Available online at www.sciencedirect.

com

ScienceDirect
Journal of Economic Theory 213 (2023) 105722
www.elsevier.com/locate/jet

Information design for selling search goods and the

effect of competition ✩
Chen Lyu
Peking University HSBC Business School, China
Received 30 November 2021; final version received 30 July 2023; accepted 22 August 2023
Available online 1 September 2023

Abstract
I study optimal information provision by a search goods seller. While the seller controls a consumer’s
pre-search information, he cannot control post-search information because the consumer will inevitably
learn the product’s match after search. A relaxed problem approach is developed to solve the optimal de-
sign, which accommodates both continuous value distributions and ex-ante heterogeneous consumers with
privately known outside options. The optimal design is shown to crucially depend on the outside option
value distribution, and can be implemented by a simple upper-censorship signal under certain regularity
conditions. Several applications are provided, including comparing information designs for search goods
and experience goods, and studying the effect of competition with a large number of sellers.
© 2023 Elsevier Inc. All rights reserved.

JEL classification: D40; D82; D83; L10; M37

Keywords: Information design; Bayesian persuasion; Search goods; Consumer search; Competition

✩
I would like to thank Tilman Börgers, the associate editor and two anonymous referees for their highly helpful
comments and suggestions. I also thank Heski Bar-Isaac, Kyungmin Kim, Marzena Rostek, Daniel Quint, Lones Smith,
Lu Liao as well as various participants at the North American Summer Meeting of the Econometric Society (2021),
Young Economist Symposium (2021), the Latin American Meeting of the Econometric Society (2021), the 32nd Stony
Brook International Conference on Game Theory (poster session) and the theory group seminar at the University of
Wisconsin-Madison.
E-mail address: [email protected].

https://doi.org/10.1016/j.jet.2023.105722
0022-0531/© 2023 Elsevier Inc. All rights reserved.
C. Lyu Journal of Economic Theory 213 (2023) 105722

1. Introduction

Modern information technology has significantly improved firms’ ability in communicating

product information with potential customers. For examples, they can provide virtual demon-
strations online, make personalized recommendations to individual customers (e.g., via targeted
email ads or mobile pushes), or have key opinion leaders evaluate their products for specific
consumer groups. Given this ease and flexibility of information provision, what and how much
information to provide has become an important design question for firms.
Most existing studies on consumer information design have focused on experience goods,
which are products whose matches cannot be uncovered by consumers on their own before pur-
chase. For this type of products, consumers have to make purchase decisions solely based on
the information provided by the seller. However, many products in practice are better understood
as search goods, whose match values are naturally revealed to consumers after a search step
before purchase.1 Classical examples include clothing, home supplies and furniture, for which
a consumer can easily tell how she likes the product upon visiting the seller. More generally, a
product that allows returns may also be better understood as a search good because its consumers
can almost learn their match values before committing on purchase. Treating a search good as
an experience good ignores the post-search information revealing and may thus exaggerate the
seller’s control over consumer information.
The consumer’s ability to learn her true match value before purchase seems to make the sell-
er’s information design irrelevant. In many situations, however, the search step required before
one can learn the product’s match incurs a search cost. This can include the cost of visiting a
local store in off-line shopping or non-refundable shipping costs in online shopping. Although a
search goods seller’s information provision is irrelevant after search, it decides whether the con-
sumer will search for the product in the first place. By providing proper pre-search information,
the seller can attract as many as possible consumers who will purchase once having sunk their
search costs into searching.
In this paper, I study optimal pre-search information provision by search goods sellers. In the
main model, a monopoly seller designs a general pre-search signal (à la Bayesian persuasion)
for a representative consumer about her match with a product, based on which the consumer can
decide whether to search for the product.2 If the consumer searches, she will fully learn her match
value and then decide whether to make a purchase; otherwise, she will take an outside option.3
Since consumers in practice often know their best alternatives to a product, which is unknown to
the seller, I allow the consumer to have private information on her outside option value.
The post-search revelation of product match, as the identifying property of search goods,
significantly complicates the seller’s information design problem compared to that for experience
goods. In particular, the first moment of the consumer’s posterior belief on match value no longer
suffices for determining the purchase outcome.4 Together with the existence of the consumer’s

1 These concepts on goods classification are introduced by Nelson (1970).

2 Section 2.2 provides some concrete interpretations for the seller’s information provision.
3 Like many papers on consumer search (see Anderson and Renault (2018) for a survey), I assume searching is a
necessary step for purchase. This is natural in many shopping environments. For examples, visiting a store and finding
the product is necessary for off-line shopping; ordering the product for shipping and examining it upon delivery is
necessary for online shopping.
4 If the product’s match value has a chance to be really high, even if its mean is low, it can still be optimal for the
consumer to search and learn the actual match. This suggests that higher moments of the consumer’s posterior belief
must matter.

2
C. Lyu Journal of Economic Theory 213 (2023) 105722

private information on her outside option, this makes it difficult to derive the optimal design with
existing tools.5 To overcome this challenge, I propose a relaxed problem of the seller. Despite
being much simpler than the original problem, this relaxed problem is sufficient for solving the
optimal design under certain regularity conditions.
When the consumer’s outside option value is unimodal (i.e., has a quasi-concave density), I
show that it is optimal for the seller to provide an upper-censorship signal. Under this signal,
consumers with match values below a threshold will be fully informed of their matches, while
others will only learn that their matches are above the threshold. This in particular suggests that
in order to attract more consumers into purchase, coarser information should be communicated
to those with relatively high match values than to those with lower matches. The characteriza-
tion also reveals that the curvature of the consumer’s outside option value distribution plays an
important role in shaping the optimal design. In particular, when the distribution is convex over a
relevant region, the optimal signal will be fully revealing; when it is concave, the optimal signal
will be completely uninformative.
Besides solving the optimal design, my relaxed problem approach also enables a straight-
forward comparison between the information design problems of search goods and experience
goods. Specifically, the search goods seller’s relaxed problem can be treated as the experience
goods seller’s problem with just one more constraint. On one hand, the additional constraint for-
malizes the intuition that the seller’s control over consumer information is weaker with search
goods than with experience goods, which is critical in shaping the optimal information provision.
On the other hand, despite being more constrained, the relaxed problem does share an important
structure with the design problem of experience goods. This explains analogies between many
results of these two types of goods and has enabled me to extend results from one to the other.
My characterization of the optimal design enables several applications. The first one studies
the effect of policies turning experience goods into search goods.6 It is shown that while these
policies directly benefit consumers by forcing post-search information revealing, under certain
conditions they may reduce the seller’s incentive in providing pre-search information and thereby
lead to more inefficient searches. The overall effect on consumer welfare can be negative. This
suggests that for such policies to benefit consumers, additional efforts may be needed to maintain
sellers’ incentives in providing better pre-search information.
In the second application, I consider the possibility that the seller can partially observe con-
sumers’ outside option values and tailor information accordingly. I show that the seller optimally
provides better information to those who are expected to have higher outside option values, which
forms a kind of discrimination in information provision. This discrimination affects different con-
sumers heterogeneously, and its effect on total consumer welfare is generally indeterminate. This
highlights a non-price discrimination channel through which the design of consumer privacy can
influence consumer welfare.
In the third application, I examine the effect of changes in the search cost. When the product
price is exogenously fixed,7 the model predicts that the consumer welfare will unambiguously
increase when the search cost decreases, although this gain will be partially offset by coarser

5 A discussion about the existing tools is provided in Appendix A.1.

6 Examples of such policies include product labeling laws and regulations forcing sellers to accept product returns. See
my discussion in Section 4 for more details.
7 This mainly applies when the seller is a salesperson or information intermediary, who does not set price but just tries
to induce purchase via providing match information.

3
C. Lyu Journal of Economic Theory 213 (2023) 105722

pre-search information provided by the seller. This holds not only for overall consumer welfare,
but also for each individual consumer given her realized outside option value.
My last application is to study the effect of competition among multiple sellers with hori-
zontally differentiated products. Although I am not able to solve the equilibrium in general, my
approach suffices for pinning down an equilibrium when the number of sellers is sufficiently large
under certain regularity conditions. In particular, I show that as competition becomes stronger,
there is a sequence of equilibria with pre-search information converging to full information. This
extends the corresponding result in Hwang et al. (2019) from experience goods to search goods.
Related literature – Possibly due to its technical difficulty, the economic literature on information
design for search goods is relatively scarce. To the best of my knowledge, the only major study on
seller’s information provision for search goods is Anderson and Renault (2006).8 They consider a
similar setting as mine, but the consumers in their paper are ex-ante homogeneous without private
information. In Section 7.1, I will compare their optimal design with mine and explain how their
result is partially extended in my setting. I note that most of my findings in the applications above
cannot be made in their setting because the consumer’s private information plays important roles
in those findings.
My paper also relates to several other strands of literature. The first strand studies sellers’
provision of real product information without consumer search (e.g., Meurer and Stahl, 1994;
Lewis and Sappington, 1994; Johnson and Myatt, 2006; Ivanov, 2013; Boleslavsky et al., 2017;
Hwang et al., 2019). Most relatedly, Ivanov (2013) and Hwang et al. (2019) consider competitive
sellers and show that the equilibrium information converges to full information when the number
of sellers goes to infinity. In particular, the equilibrium characterization in Hwang et al. (2019)
also plays a key role in my analysis. As mentioned earlier, the corresponding result of mine
extends their results from experience goods to search goods.
The second strand of literature studies various information design questions in consumer
search environments. Unlike mine, most of these papers focus on post-search information design
instead of pre-search information design (Bar-Isaac et al., 2012; Board and Lu, 2018; Whitmeyer,
2020; Dogan and Hu, 2022; Au and Whitmeyer, 2023). Since their consumers cannot receive ad-
ditional information than that provided by the seller, these papers can still be considered as being
about experience goods. Two exceptions are Choi et al. (2019) and Hinnosaar and Kawai (2020).
Choi et al. (2019) studies the consumer-optimal pre-search information; Hinnosaar and Kawai
(2020) studies the seller-worst pre-search information in a robust mechanism design problem.
Unlike these papers, I study the seller-optimal pre-search information provision. Another differ-
ence is that I allow continuous value distributions, while the two papers above assume binary
values.
More closely related, Wang (2017) considers an environment where consumers can engage in
costly search for additional information after receiving the seller’s signal. With ex-ante homoge-
neous consumers, the paper concludes that the seller’s optimal signal should deter the consumer
from searching for more information. Matyskova and Montes (2023) also derives a similar result
in a more abstract Bayesian persuasion setting. Importantly, while search is optional in these pa-
pers, it is necessary for making purchase in my model. Thus their main topic of search deterrence
is not relevant in my study. Moreover, their analyses heavily rely on the ex-ante homogeneity of

8 Also see Anderson and Renault (2013), which extends Anderson and Renault (2006) to encompass vertical informa-
tion.

4
C. Lyu Journal of Economic Theory 213 (2023) 105722

consumers (as in Anderson and Renault (2006)). In contrast, I allow ex-ante heterogeneous con-
sumers with privately known outside option values.
My paper greatly benefits from recent developments in the consumer search literature. In
particular, I draw upon a characterization for the consumer’s purchase outcome given any be-
lief on the product match value, which is discovered in several papers (Kleinberg et al., 2016;
Armstrong, 2017; Choi et al., 2018).
Finally, the paper relates to the general Bayesian persuasion literature (Rayo and Segal, 2010;
Kamenica and Gentzkow, 2011). In particular, the relaxed problem I propose is analogous to the
optimization in Dworczak and Martini (2019) and Section 4.3 of Kolotilin (2018), but features
one more constraint. The optimality conditions I provide essentially extend Theorem 1 in Dwor-
czak and Martini (2019) to accommodate that additional constraint. Kolotilin et al. (2017) and
Guo and Shmaya (2019) also consider a receiver with private information. However, they do not
consider the design for selling search goods, whose complexity motivates my relaxed problem
approach.
The paper is organized as follows: Section 2 introduces the main model; Section 3 develops
the relaxed problem and characterizes the optimal design; Section 4 compares the designs of
search goods and experience goods; Section 5 considers comparative statics and related applica-
tions; Section 6 considers information provision by competing sellers; Section 7 provides further
discussions. All proofs are provided in the appendix.

2. The model

I first describe the model and then provide two interpretations for it in Section 2.2.

2.1. The setup

The model features a seller, a representative consumer and a single product. The consumer’s
match value with the product is denoted as U , which is initially unknown to both agents and
has a continuous distribution FU with compact support [u, ū]. At the beginning of the game, the
seller can design a pre-search signal (statistical experiment) to provide information about U to
the consumer. Following Anderson and Renault (2006) and the Bayesian persuasion literature,
I do not make any restriction on the signal structure.9 I use S to denote the signal’s realization,
and use φ(·; S) to denote the consumer’s posterior belief on U given S.
If the consumer does not consume the product, she will consume her outside option, whose
value is the consumer’s private information and is denoted as U0 . I assume U0 is drawn from a
distribution J , and is independent from U .
I will consider two kinds of sellers. A non-pricing seller treats the product’s price as exoge-
nous, and just wants to maximize the consumer’s purchase probability by providing informa-
tion10 ; a pricing seller also sets the product’s price p, and tries to maximize the expected profit.
I normalize the seller’s marginal cost to zero.
The game goes as follows:

9 Formally, a signal consists of a measurable realization space S and a transition kernel π : [u, ū] → (S) that maps
any realized match value to a probability distribution over S, according to which the signal realization S will be drawn.
10 In practice, non-pricing sellers can be salesmen, brokers or information intermediaries, who typically do not set price
but just try to induce purchase by providing information.

5
C. Lyu Journal of Economic Theory 213 (2023) 105722

1. The seller designs the pre-search signal about U . A pricing seller also chooses the product’s
price p. These are observed by the consumer.11
2. The match value U is (secretly) realized, and the pre-search signal realization S is generated
to the consumer.
3. After learning S and her outside option value U0 , the consumer decides whether to search
for the product. If not, she consumes the outside option and receives utility U0 .
4. If the consumer chooses to search, she will incur a search cost c > 0 and learn U . She then
makes her purchase decision. With purchase, her final utility will be U − p − c; without
purchase, her final utility will be U0 − c.

The assumption that the consumer can fully learn her match value after search is the identify-
ing property of search goods. Formally, I adopt the following dichotomy in Nelson (1970).

Definition 2.1. A product is a search good (abbr., SG) if its match value will be fully revealed
to the consumer after search; it is an experience good (abbr., EG) if the consumer cannot learn
additional information after search before purchase.

Notice that I assume search is a necessary step towards purchase even if the product is an experi-
ence good. This ensures the two types of goods have the same total purchase cost so that they are
comparable. I will call the seller in the model described above an SG seller, and call a seller EG
seller if he faces the same problem except that the consumer will learn no additional information
after search. In Section 4, I will compare the information design problems of these two types of
sellers, and highlight the key similarity and dissimilarity between them.

2.2. Interpretations for the model

The model admits two general interpretations. In the first interpretation, which is adopted
by Anderson and Renault (2006), the seller has a single product and faces a large population
of potential consumers, whose matches with the product are independently drawn. To provide
pre-search information, the seller advertises selected horizontal product characteristics to all con-
sumers (e.g., via mass media).12 With these characteristics, each consumer can privately update
her belief about the product’s match value based on her own taste. By selecting different charac-
teristics to advertise, the seller then imparts different information to the consumers.
In the second interpretation, which is more popular in the Bayesian persuasion literature, the
seller interacts with each individual consumer repeatedly over time. Each time, the seller offers
one issue of his product, whose match with the consumer is drawn independently from those
of the other issues. Knowing both features of the product issues and the consumer’s preference
(e.g., revealed by browsing history and demographics), the seller can predict the consumer’s
match with each issue and provide information about it (e.g., via recommendation messages).
Since the seller repeatedly interacts with the consumer, it is conceivable that he can commit
on a particular information provision rule to maximize the long-run profit, and the consumer can

11 I assume that the price set by a pricing seller is observed before search, which avoids the hold-up problem in the
Diamond’s Paradox (Diamond, 1971). This is innocuous as long as the seller can provide price information along with
the pre-search match information.
12 Only horizontally differentiating characteristics are considered here, which better informs a consumer of her individ-
ual match without vertically shifting the aggregate demand.

6
C. Lyu Journal of Economic Theory 213 (2023) 105722

correctly interpret the messages she receives.13 This justifies the seller’s commitment power over
the signal structure being used.14

3. Solving the Seller’s optimal design

I first consider a non-pricing seller with exogenous price p, and focus on the information
design problem. The pricing seller’s problem will be studied in Section 3.4.

3.1. Preliminaries

To characterize the consumer’s optimal search behavior, given any posterior belief φ on U ,
let zφ denote the corresponding Pandora’s index.15 That is, zφ solves:

[(x − zφ )+ − c]φ(dx) = 0 (3.1)

where (y)+ := max{y, 0}. Given any pre-search signal, I define the random variable Z as the
Pandora’s index of the consumer’s posterior belief conditional on the pre-search signal realiza-
tion, i.e., Z := zφ(·;S) . Then, according to the Pandora’s rule in Weitzman (1979), the consumer
will search for the product if and only if Z ≥ U0 + p. Since the consumer will purchase after
search if and only if the revealed U ≥ U0 + p, the product will be finally sold if and only if
U ∧ Z ≥ U0 + p, where x ∧ y is a shorthand for min{x, y}.16 I will call this key statistic U ∧ Z
the product’s effective-search-value.17
Let G denote the CDF of U ∧ Z induced by the pre-search signal. Let Jp denote the CDF of
U0 + p, i.e., Jp (x) := J (x − p). The consumer’s purchase probability then equals to P (U0 +
p ≤ U ∧ Z) = E[Jp (U ∧ Z)] = Jp (x)dG(x). The non-pricing seller’s problem can then be
formulated as:

max Jp (x)dG(x) (3.2)
G
s.t. G is a feasible distribution of U ∧ Z (3.3)
Unfortunately, I cannot handle this optimization directly because a full characterization of its
feasible set will be too complicated to handle in such an optimization. To get things simplified, I
will hence propose a relaxed problem of it, which is based on two lemmas below.
Given any search cost c, I will call U − c the consumer’s net-match-utility and let FU −c denote
its CDF. The following technical lemma provides a key necessary condition for the constraint
(3.3) to hold.

13 Several recent papers have studied how repeated interaction can support commitment in strategic communication.
See, e.g., Mathevet et al. (2019) and Best and Quigley (2020).
14 An implicit assumption of my model with a pricing seller is that the seller cannot price-discriminate consumers based
on their realized match values. This is natural in applications fitting the first interpretation above, since the seller there
cannot observe individual consumers’ match values. In applications fitting the second interpretation, the assumption can
also hold because, for example, consumers can often re-gain anonymity (e.g., by browsing in private mode or switching
to another account) to circumvent discriminatory prices.
15 This index is called “reservation price” in the original paper of Weitzman (1979). It is easy to show its existence and
uniqueness when c > 0.
16 I assume the consumer will search and buy the product when being indifferent.
17 This characterization of the consumer’s purchase outcome has been proposed by several papers (Kleinberg et al.,
2016; Armstrong, 2017; Choi et al., 2018).

7
C. Lyu Journal of Economic Theory 213 (2023) 105722

Lemma 3.1. Given FU and c, the distribution of U ∧ Z under any signal is a mean-preserving
contraction (MPC) of FU −c . As a special case, U ∧ Z = U − c under the fully revealing signal.

The lemma suggests that we can relax the seller’s problem by replacing constraint (3.3) with
the constraint that G is a MPC of FU −c . However, this will lead to an optimization that is often
too relaxed for its solution to be feasible for the seller. Actually, as I will discuss in Section 4, such
an optimization is equivalent to that faced by an EG seller, with which the difference between the
two types of goods will be neglected. Therefore, to derive a more useful relaxed problem, some
additional characterization for the constraint (3.3) is needed. The next lemma provides such a
result.

Lemma 3.2. Under any signal, U ∧ Z ≤ U and thus the distribution of U ∧ Z is first-order
stochastically dominated by FU .

While being obvious, the lemma highlights an important restriction faced by an SG seller.
Namely, no matter what pre-search information is provided, the consumer’s effective-search-
value U ∧ Z is always bounded by the true match value U . This limits the seller’s ability in
manipulating the consumer’s purchase behavior through inducing mean-preserving contraction
of G by withholding information.

3.2. A relaxed problem

Let MP S denote the mean-preserving spread order and let F OD denote the first-order
stochastic dominance order. Based on Lemmas 3.1 and 3.2, a Relaxed Problem of optimization
(3.2) – (3.3) can be formulated as:

max Jp (x)dG(x) (3.4)
G
s.t. G MP S FU −c (3.5)
G F OD FU (3.6)
For any distribution F , I denote its support as supp{F }. The following theorem provides the
optimality conditions for solving the linear program above. It extends Theorem 1 in Dworczak
and Martini (2019) to accommodate constraint (3.6).

Theorem 3.1. A distribution G solves problem (3.4) – (3.6) if there exists functions v(·) and ρ(·)
such that:

(C1) v(·) is convex over [u − c, ū − c] and ρ(·) is weakly increasing over [u − c, ū].
(C2) v(x)
 + ρ(x) ≥ Jp (x) for all x ∈ [u − c, ū − c], withequality holding for any x ∈ supp{G}.
(C3) v(x)dG(x) = v(x)dFU −c (x); ρ(x)dG(x) = ρ(x)dFU (x).
(C4) G satisfies constraints (3.5) and (3.6).

Moreover, if there exist G, v(·) and ρ(·) satisfying the above conditions, then another distribution
Ĝ also solves the problem if and only if it satisfies conditions (C1) – (C4) with the same v(·) and
ρ(·).

8
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. 3.1. Values of U ∧ Z given different values of U − c under an upper-censorship signal with threshold η. The axis
represents different values of U − c and the expressions above
 the brackets
 indicate the value of U ∧ Z given U − c in
each region. We can see that U ∧ Z > U − c when U − c ∈ η, z(η) and U ∧ Z < U − c when U − c > z(η).

In general, the Relaxed Problem does not necessarily admit a solution that is feasible for
the seller to induce. However, as I will show in the next subsection, it does under some mild
conditions on the outside option value distribution.

3.3. The optimal signal with unimodal outside option

I impose the following unimodal assumption on the distribution of U0 + p.

Assumption 3.1 (Unimodal Jp ). Jp admits a continuous quasi-concave density over [u − c, ū −

c].

Let jp denote the density of Jp . The assumption mainly requires that over the support of U − c,
jp is first increasing and then decreasing, which implies Jp to be first convex and then concave.18
Under the assumption, I will use [rpmin , rpmax ] to denote the mode interval of Jp over [u − c, ū −
c], i.e.,
[rpmin , rpmax ] := arg max jp (x) (3.7)
x∈[u−c,ū−c]

Following Kolotilin et al. (2021), I define an upper-censorship signal as follows:

Definition 3.1. A signal is an upper-censorship signal if there is a threshold η ∈ [u − c, ū − c]

such that the signal fully reveals any net-match-utility U − c below η, and pools all net-match-
utilities above η together.

Given any upper-censorship signal with threshold η, I will use Gη to denote the distribution of
U ∧ Z it induces, and use z(η) to denote the Pandora’s index of the consumer’s posterior belief
after learning U − c ≥ η.19 It is easy to see that z(η) ≥ η and is strictly increasing in η. Fig. 3.1
illustrates how U − c maps to U ∧ Z under an upper-censorship signal. We can see that compared
to the full revelation case, where U ∧ Z always equals to U − c, upper-censorship leads to a
contraction for U ∧ Z over the region above η, which by Lemma 3.1 must be mean-preserving.
In particular, consumers with U − c ∈ (η, z(η)) will have their effective-search-values increased
to either U or z(η), whichever is smaller; those with U − c > z(η) will have their effective-
search-values decreased to z(η).
To state the main result below, given any Jp satisfying the unimodal assumption, I define:
η0 := inf{η ∈ [u − c, ū − c] : z(η) ≥ rpmin } (3.8)

18 This is weaker than requiring U to admit a log-concave density, which is satisfied by many common distributions
0
(see, e.g., Table 1 in Bagnoli and Bergstrom (2005)).
19 I provide detailed properties of z(·) and the formula of G in Appendix B.3.1.
η

9
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. 3.2. Graphical illustration for the optimality condition (η∗ ) = 0. The red curve is Jp (·). The condition implies that
the secant of Jp (·) over [η∗ , (η∗ + c) ∧ z(η∗ )] is parallel to the tangent of Jp (·) at z(η∗ ). (For interpretation of the colors
in the figure(s), the reader is referred to the web version of this article.)

 
Jp (η + c) ∧ z(η) − Jp (η)  
(η) := − jp z(η) , ∀η ∈ [u − c, ū − c) (3.9)
(η + c) ∧ z(η) − η
Intuitively, η0 is the smallest η such that z(η) will fall in the region where Jp (·) is concave; (η)
measures the difference between the average slope of Jp (·) over [η, (η + c) ∧ z(η)] and its slope
at z(η). An optimal signal is characterized by the following proposition.

Proposition 3.1. Under Assumption 3.1, for a non-pricing seller:

(a) If rpmax = ū − c, the upper-censorship signal with threshold η∗ = ū − c (i.e., full disclosure)
is optimal.
(b) If rpmax < ū − c, there exists η∗ ∈ [η0 , rpmax ] such that either of the following conditions hold:
(i) η∗ > u − c and (η∗ ) = 0;
(ii) η∗ = u − c and (η∗ ) ≥ 0.
For any such η∗ , Gη∗ solves the Relaxed Problem (3.4) – (3.6) and the upper-censorship
signal with threshold η∗ is optimal.

To gain some intuition about why upper-censorship is optimal under the unimodal assumption,
consider two events with equal probabilities: in event A, U − c = a; in event B, U − c = b > a.
If they are separately revealed, then U ∧ Z = U − c in both events. Now, suppose the seller
instead pools these events. Then the benefit for him is that U ∧ Z will be raised to some a > a
J (a )−J (a)
in event A, which increases the sale probability by p 2 p ; the cost is that U ∧ Z will be
Jp (b)−Jp (b )
decreased to some b < b in event B, which decreases the sale probability by 2 . Notice
Jp (a )−Jp (a)
the mean-preserving condition implies a − a = b − b . We thus must have 2 to be less
J (b)−J (b )
(resp., greater) than p 2 p when Jp is convex (resp., concave) over [a, b]. This suggests
that the optimal signal should reveal better information over a region where Jp is convex, and
tends to withhold information over a region where Jp is concave. Under Assumption 3.1, this is
right in accordance with upper-censorship.20
In part (b) of Proposition 3.1, an interior optimal threshold can be pinned down by the con-
dition (η∗ ) = 0. Graphically, this means that the secant of Jp over [η∗ , (η∗ + c) ∧ z(η∗ )] has

20 This does not imply that the optimal threshold coincides with the mode of J , since J (a )/2 −J (a)/2 > J (b)/2 −
p p p p
Jp (b )/2 can also hold when Jp is first convex and then concave over [a, b], in which case pooling is more profitable.
Proposition 3.1 only implies η∗ ≤ rpmax .

10
C. Lyu Journal of Economic Theory 213 (2023) 105722

the same slope as Jp at z(η∗ ) (see Fig. 3.2). To understand this condition, suppose the seller
adds some additional consumers with U − c right below η∗ to the pooling region, whose to-
∗
tal mass is dm. The benefit is that these consumers will have
∗ ∗
 ∗U ∧ Z increased
 their ∗
 from∗ η to
(η +c) ∧z(η ), which increases the total sale probability by Jp (η +c) ∧z(η ) −Jp (η ) dm.
The cost is that consumers originally with U ∧ Z = z(η∗ ), I denote whose mass as M, will have
their effective-search-values marginally decreased from z(η∗ ) by some  |dz| to obey the mean-
preserving condition. This decreases the total sale probability by Jp z(η∗ ))|dz|M. The seller’s
    
net benefit thus equals to Jp (η∗ + c) ∧ z(η∗ ) − Jp (η∗ ) dm − Jp z(η∗ ))|dz|M, which is pro-
portional to:
  ∗  
Jp (η + c) ∧ z(η∗ ) − Jp (η∗ ) 
 − Jp z(η∗ ))
|dz|M dm
By the mean-preserving condition, we must have [(η∗ + c) ∧ z(η∗ ) − η∗ ]dm = |dz|M. The above
expression is thus equal to (η∗ ). Therefore, the graphical condition guarantees that any marginal
deviation from η∗ will not be profitable for the seller.21
Proposition 3.1 suggests that the seller’s optimal design crucially depends on the curvature
of the consumer’s outside option value distribution. The following corollary of it illustrates this
with the extreme cases.

Corollary 3.1. For a non-pricing seller, we have:

(a) If Jp (·) is convex over [u − c, ū − c], then the fully revealing signal is optimal.
(b) If Jp (·) is concave over [u − c, ū − c], then the fully pooling signal is optimal.

In general, the seller’s optimal design is not unique. For example, suppose an optimal upper-
censorship signal is fully revealing for U − c ∈ [a, b] and Jp is affine over this region. Then
another signal pooling over this region (otherwise identical to the original signal) will also be
optimal. However, under a strict version of the unimodal assumption on Jp , we do have a unique-
ness result.

Assumption 3.2 (Strictly Unimodal Jp ). Assumption 3.1 holds with jp being strictly quasi-
concave over [u − c, ū − c].

The assumption requires jp to be first strictly increasing and then strictly decreasing over [u −
c, ū − c]. When it holds, I will use rp to denote the unique mode of Jp over [u − c, ū − c], i.e.,
rp = arg maxx∈[u−c,ū−c] jp (x).

Proposition 3.2. Under Assumption 3.2, all optimal signals induce the same joint distribution of
(U, Z), and the optimal upper-censorship signal is unique.

Notice that the pair of (U, Z) determines a consumer’s search and purchase decisions as well
as her ex-post utility given any U0 . Thus the proposition implies that under Assumption 3.2, the
search-purchase outcome and the consumer surplus in equilibrium are both unique. This makes
it convenient to study any comparative statics.

21 For corner solution η∗ = u − c, the threshold cannot be further lowered, so we only need to rule out profitable upward
deviations from it. Therefore, we only need to require (η∗ ) ≥ 0.

11
C. Lyu Journal of Economic Theory 213 (2023) 105722

3.4. The optimal design for a pricing seller

Given the results for a non-pricing seller, characterizing the optimal design for a pricing seller
is straightforward. By the same argument as in Section 3.1, the pricing seller’s problem can be
written as:

max{p J (x − p)dG(x)} (3.10)
G, p

s.t. G is a feasible distribution of U ∧ Z (3.11)

The following assumption allows me to use the simple characterization of the non-pricing
seller’s optimal design in Section 3.3.

Assumption 3.3 (Unimodal J ). J admits a continuous quasi-concave density.

The assumption guarantees that for any price p, the distribution of U0 + p (i.e., Jp ) satisfies
Assumption 3.1 and hence there exists an optimal upper-censorship signal. Since an upper-
censorship signal is solely pinned down by its threshold η, we can transform the pricing seller’s
problem into an optimization over (p, η), which leads to the following result.

Proposition 3.3. Under Assumption 3.3, a price p ∗ and an upper-censorship signal with thresh-
old η∗ are optimal for the pricing seller if and only if:

∗ ∗
(p , η ) ∈ arg max {p J (x − p)dGη }
p≥0; η∈[u−c,ū−c]

where the formula of Gη is given by equation (B.3) in Appendix B.3.1. Moreover, if J (·) is log-
concave over (−∞, x0 ) for some x0 > 0, then an optimal pair of (p ∗ , η∗ ) exists.

Proposition 3.3 has simplified the original infinite-dimensional optimization (3.10) – (3.11)
into a two-dimensional problem. It will allow me to study several applications of the model in
later sections.

4. Search goods vs. experience goods

In practice, whether a product is a search good (SG) or an experience good (EG) not only de-
pends on its own property, but also depends on the shopping environment and related consumer
protection policies. For examples, product labeling laws require sellers to provide detailed prod-
uct information on packages, which enables consumers to learn product characteristics (e.g.,
food nutrition) that they would otherwise not know before (or even after) purchase; consumer
protection laws in many countries require online sellers to accept returns without any reason
within certain time period.22 These policies allow consumers to better learn their matches with a

22 For instances, China requires online sellers to fully refund no-reason returns (except for special products) within 7
days of the sale, excluding any shipping cost; European Union has a similar policy with the cooling period being 14 days
and the refund there includes initial (standard) shipping charges. See http://lawinfochina.com/display.aspx?id=23187%
26lib=law and https://europa.eu/youreurope/citizens/consumers/shopping/guarantees-returns.

12
C. Lyu Journal of Economic Theory 213 (2023) 105722

product after some search step but before committing on purchase, and can thus approximately
transform experience goods into search goods.23
In this section, I investigate how such policies can change the equilibrium outcomes and com-
pare the two types of goods from an information design perspective. For simplicity, I will focus
on non-pricing sellers with exogenous product price unless otherwise stated. Some numerical
analyses will be provided in Appendix A.2 for the case of pricing sellers.

4.1. Comparison between the design problems

When the product is an experience goods, the consumer will receive no additional information
after the search step. She will hence buy the product if and only if her posterior mean net-match-
utility E[U − c|S] exceeds U0 + p. The statistic E[U − c|S] thus replaces the role of U ∧ Z in
determining the purchase outcome. Let H denote the distribution of E[U − c|S] to be induced. It
is well known that H is feasible if and only if it is a MPC of FU −c .24 A non-pricing EG seller’s
problem can thus be written as:

max Jp (x)dH (x) (4.1)
H
s.t. H MP S FU −c (4.2)
This kind of optimization has been well studied in the Bayesian persuasion literature (e.g.,
Kolotilin (2018) and Dworczak and Kolotilin (2019)).
Comparing optimization (4.1) – (4.2) with optimization (3.4) – (3.6), one can see some key
similarity and dissimilarity between the problems of the two types of goods. In terms of similar-
ity, both problems involve the same MPC constraint. This roots from the fact that both U ∧ Z
and E[U − c|S] must be a MPC of the true net-match-utility U − c. This commonality allows
many results to be carried over between the two types of goods. In particular, under the unimodal
assumption on Jp , the optimal design of an EG seller also features upper-censorship signals.
Moreover, the results in Corollary 3.1 equally hold for both types of goods.
The key difference between the two problems is that an SG seller faces the additional con-
straint (3.6) requiring G F OD FU , which is absent in the EG seller’s problem. This is because
U ∧ Z must be bounded by U , but E[U − c|S] needs not be. In this sense, it is harder to induce
MPC for U ∧ Z by withholding information than for E[U − c|S], which makes the SG seller’s
problem more “constrained”. An obvious implication of this is the following, which holds for
both pricing and non-pricing sellers.

Proposition 4.1. The seller’s profit is (weakly) lower when the product is a search good than
when it is an experience good.

In the following subsections, I will further compare the two types of goods in terms of the
equilibrium information provision and consumer welfare. For convenience, I first present a useful
lemma here.

23 To interpret the search goods model in the situation of online shopping with returnable products, one should consider
c to also include any return cost of the consumer, and consider U to be the product’s consumption utility plus the return
cost. Then the consumer’s final utility with search without purchase is indeed U0 − c, and her utility with purchase is
indeed U − c − p.
24 See, e.g., Proposition 2 in Kolotilin (2018).

13
C. Lyu Journal of Economic Theory 213 (2023) 105722

Lemma 4.1. For any belief φ on U , we have zφ ≥ EU ∼φ [U − c], which holds as equality if and
only if inf(supp{φ}) ≥ EU ∼φ [U − c].

The lemma implies that given any posterior belief φ of the consumer, the corresponding
Pandora’s index zφ is no less than the posterior mean of U − c. This is intuitive since even if
the consumer always ignores the post-search information, it is optimal for her to search when
U0 + p ≤ EU ∼φ [U − c]. Thus zφ must be no less than EU ∼φ [U − c] for the Pandora’s rule to
be optimal. The lemma also implies that these two values are equal when c is sufficiently large
such that even the smallest value in the support of φ is greater than EU ∼φ [U − c]. Intuitively,
when the search cost is very large, the post-search information will not make a difference to the
consumer’s decisions because it comes only after the consumer has sunk a significant cost. The
range of U0 that makes it optimal for the consumer to search should hence be the same regardless
of the product’s type, which implies zφ = EU ∼φ [U − c].

4.2. Comparison for the equilibrium information provision

I first define some notations. Let μ(η) := E[U − c|U − c ≥ η] and

 
Jp μ(η) − Jp (η)  
 (η) :=
E
− jp μ(η) , ∀η ∈ [u − c, ū − c) (4.3)
μ(η) − η
Intuitively, μ(η) is the posterior mean of U − c after observing the pooling signal realization of
an upper-censorship signal with threshold η; E (η) measures the difference between the average
slope of Jp over [η, μ(η)] and its slope at μ(η). I note that these functions are fully determined
by Jp and FU −c .
I will focus my analysis on the case where the following assumption holds.

Assumption 4.1. Jp satisfies Assumption 3.2 with rp < ū − c and E (u − c) < 0.

Here, the conditions rp < ū − c and E (u − c) < 0 guarantee that the EG seller’s optimal signal
is neither fully revealing nor fully pooling. These are not essential for the analysis, but avoid
separate discussions of “corner” solutions. Under Assumption 4.1, the EG seller’s optimal signal
is characterized by Kolotilin et al. (2021). For convenience, I summarize its properties in the
following proposition.

Proposition 4.2. Assume Assumption 4.1 holds. For a non-pricing EG seller, an upper-
censorship signal with threshold ηE∗ is optimal, where η∗ is the unique solution to E (η) = 0.
E
Moreover, all optimal signals are outcome-equivalent to it.

The optimality condition E (ηE ∗ ) = 0 implies that the secant of J over [η∗ , μ(η∗ )] is tangent
p E E
∗
to Jp at μ(ηE ) (see Fig. 4.1). The intuition behind it is analogous to that for the SG seller’s opti-
mality condition (η∗ ) = 0 derived in Section 3.3. By marginally decreasing the threshold from
ηE∗ , the EG seller’s benefit is that consumers with U − c right below η∗ will have their purchase
E
probability increased from Jp (ηE ∗ ) to J (μ(η∗ )), while his cost is that consumers originally in
p E
the pooling region will have their purchase probability marginally decreased from Jp (μ(ηE ∗ )).
∗
The condition  (ηE ) = 0 guarantees that these cost and benefit are just balanced out. Compar-
E

ing this analysis to its search goods counterpart in Section 3.3, one can see that both the benefit

14
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. 4.1. An illustration for the cases in Proposition 4.3(b). The red curves represent three versions of Jp . Regardless
of the version, the solid black line is tangent to Jp , which implies E (ηE ∗ ) = 0 and thus η∗ is the EG seller’s optimal
E
threshold. The two dashed black segments are parallel, which implies (ηE ∗ ) = 0 when J = J 2 . As J shifts from J 1
p p p p
∗ ) decreases, and hence (η∗ ) shifts from negative to positive.
to Jp3 , its slope at z(ηE E

and the cost associated to a marginal change in the signal threshold are generally different. This
explains why the two types of sellers have different optimality conditions.
For non-pricing SG and EG sellers respectively, let ηS∗ and ηE∗ denote the thresholds of their

optimal upper-censorship signals. Under Assumption 4.1, comparing the equilibrium informa-
tion provisions boils down to comparing these two thresholds. Notice that by Proposition 4.2,
ηE∗ can be equivalently defined as the unique solution to E (η) = 0. Given the value of it, the

following proposition provides conditions for comparing ηS∗ with ηE ∗:

Proposition 4.3. Assume Assumption 4.1 holds. We have:

(a) If c ≥ μ(ηE ∗ ) − η∗ , then η∗ = η∗ and the equilibrium outcomes of the two types of goods
E S E
are the same.
(b) If c < μ(ηE ∗ ) − η∗ , then we have: (i) (η∗ ) < 0 ⇒ η∗ > η∗ ; (ii) (η∗ ) = 0 ⇒ η∗ = η∗ ;
E E S E E S E
(iii) (ηE ) > 0 ⇒ ηS∗ < ηE
∗ ∗.

Since μ(·) and E (·) depend on c only through FU −c , so does μ(ηE ∗ ) − η∗ . Thus given any
E
fixed FU −c , part (a) of the proposition applies when c is sufficiently large, while part (b) applies
otherwise.
Part (a) of Proposition 4.3 implies that if c is sufficiently large, then the equilibrium outcome
will be the same for the two types of goods. This is intuitive since when c is large, the post-
search information revealing will make no difference to the consumer’s optimal decisions as I
have discussed below Lemma 4.1. The discrepancy between the two types of goods should hence
disappear.
When c < μ(ηE ∗ ) − η∗ , the post-search information revealing does make a difference. In this
E
case, part (b) of Proposition 4.3 shows that the order between ηS∗ and ηE ∗ is decided by the sign of
∗
(ηE ). To understand this, notice that according to my discussion in Section 3.3, (ηE ∗ ) reflects
∗ ∗
the SG seller’s profit from marginally changing the threshold from ηE . If (ηE ) > 0, a decrease
will be profitable; if (ηE ∗ ) < 0, an increase will be profitable. The proof of Proposition 4.3(b)

shows that this local analysis can be extended globally, which implies the result.

15
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. 4.2. When Assumption 4.1 holds and ηS∗ < ηE ∗ , the red curve represents the surplus change of consumers with
different U0 + p when the product is turned from an EG to an SG. See Appendix C.4 for the proof of its main qualitative
properties.

Fig. 4.1 illustrates each case of Proposition 4.3(b). Notice when c < μ(ηE ∗ ) − η∗ , Lemma 4.1
E
∗ ∗ ∗ ∗
implies z(ηE ) > μ(ηE ) > ηE + c. Thus (ηE ) equals to the difference between the slope of
25

the secant of Jp over [ηE ∗ , η∗ + c] and the slope of J at z(η∗ ), which generally differs from
E p E
∗
 (ηE ) (= 0). In the figure, we can see that if Jp shifts from Jp1 to Jp2 and then to Jp3 , (ηE
E ∗)

will change from negative to zero and then to positive, which correspond to the three cases in
Proposition 4.3(b). This in particular suggests that given the portion of Jp over [ηE ∗ , μ(η∗ )]
E
fixed, if its slope (density) turns low sufficiently fast after μ(ηE ), then we will have ηS∗ < ηE
∗ ∗.

In this case, the post-search information revealing of a search good will “crowd out” the seller’s
pre-search information provision.

4.3. Comparison for the consumer welfare

Above analyses have illustrated that the pre-search information in equilibrium may get either
better or worse after an experience good is turned into a search good. If it gets better (i.e., ηS∗ >
ηE∗ ), all consumers will certainly be better-off due to better information both before and after

search. If it gets worse (i.e., ηS∗ < ηE ∗ ), however, the welfare impact is unclear.

When Assumption 4.1 holds and ηS∗ < ηE ∗ , Fig. 4.2 illustrates that consumers with different

outside option values will be heterogeneously affected when we change an experience good into
a search good.It shows that consumers with relatively large U0 , including those with U0 + p ∈
μ(ηE ∗ ), z(η∗ ) , will be better-off.26 However, the opposite is true for those with relatively small
S
U0 , including those with U0 + p ∈ (ηS∗ , ηE ∗ + c]. For these consumers, the poorer pre-search

information will lead to too many additional inefficient searches, which makes them worse-off
despite the better post-search information available.
Now, a natural question to ask is whether it is possible for the total consumer welfare to de-
crease when the product is turned into a search good. Proposition 4.4 below provides a sufficient
condition for this to happen. For tractability, I still focus on strictly unimodal Jp and use rp to
denote its mode over [u − c, ū − c]. I also assume the model primitives (FU , c, rp ) satisfy the
following assumption.

Assumption 4.2. (FU , c, rp ) satisfy:

25 Under the EG seller’s optimal signal, the lowest value of U leading to the pooling signal realization is η∗ + c, which
E
∗ ) here. Thus z(η∗ ) > μ(η∗ ) by Lemma 4.1.
is strictly less than μ(ηE E E
26 I note that in the current situation, although η∗ < η∗ , we must have z(η∗ ) > μ(η∗ ) for the seller’s design to be
S E S E
optimal. This means that search given the pooling signal realization must be more attractive when the product is a search
good. Its formal proof is provided in Appendix C.4.

16
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. 4.3. Shape of jp tending to induce cs S < cs E .

(1) FU has a log-concave density fU over [u, ū].

(2) c < μ(rp ) − rp .
(3) z(u − c) < rp < ū − c

Condition (1) is a mild condition on FU that is satisfied by many common distributions.

Condition (2) rules out the case of Proposition 4.3(a) so that the discrepancy between the two
types of goods truly matters. Condition (3) simplifies the discussion by ruling out corner solutions
∗ and η∗ .
for ηE S
Let cs E and cs S denote the equilibrium consumer surplus with experience goods and search
goods respectively. Given any (Fu , c, rp ) satisfying Assumption 4.2, the following proposition
provides a condition regarding Jp ’s behavior on the two sides of its mode rp that guarantees
cs S < cs E . As before, I define η0 as the smallest η such that z(η) surpasses rp , i.e., η0 := inf{η ∈
[u − c, ū − c] : z(η) ≥ rp }.

Proposition 4.4. Given any (Fu , c, rp ) satisfying Assumption 4.2, there exist κ > 0 and ν > 0
such that the following holds: if Jp satisfies Assumption 3.2 with its mode being rp , and it satisfies
J (ū−c)−Jp (u−c)
(i) jp (rp + ν) < jp (η0 ) and (ii) jp (η0 ) > κ p ū−u , then cs S < cs E .27

The proposition implies that given any (Fu , c, rp ) satisfying Assumption 4.2, one can find a
pair of (κ, ν) such that as long as Jp is strictly unimodal and satisfies conditions (i) and (ii), then
cs S < cs E . To understand these conditions, notice η0 < rp < rp + ν. Thus condition (i) holds
when jp drops sufficiently fast to the right of rp versus to the left of rp . This guarantees that the
shape of Jp is more aligned with Jp3 in Fig. 4.1 than with Jp1 , which leads to ηS∗ < ηE ∗ . Actually,
∗
when ν is chosen to be small enough, the condition ensures that ηS will be considerably less than
ηE∗ , which makes the deterioration of pre-search information significant. Condition (ii) holds

when jp is sufficiently “fat” to the left of its mode. It guarantees that there will be a significant
proportion of consumers with U0 + p lying in the region of [ηS∗ , ηE ∗ + c]. As I have shown in

Fig. 4.2, such consumers have a net loss when the product changes into a search good. We thus
have cs S < cs E when their proportion is large. Overall, Proposition 4.4 suggests that the total
consumer surplus tends to be lower in the search goods case when jp has a moderately sloped
“left-hillside” and a steep “right-hillside” relative to its mode, as is illustrated in Fig. 4.3.
My analyses have shown that policies turning experience goods into search goods may crowd
out the seller’s pre-search information provision and unintentionally reduce the total consumer
welfare. For such policies to truly benefit consumers, we may thus need to provide extra incen-
tives to sellers for offering pre-search information. In the context of online shopping, for example,

27 The proof shows how the pair of (κ, ν) can be constructed given any (F , c, r ). The parameter κ can actually be
u p
chosen arbitrarily if ν is chosen accordingly.

17
C. Lyu Journal of Economic Theory 213 (2023) 105722

one possibility is to require the sellers to afford some shipping cost for returned products.28 This
makes wasteful searches also costly for the sellers and can thus incentivize them to provide better
pre-search information.

5. Comparative statics

5.1. Discriminatory information provision

In this subsection, I consider comparative statics about the seller’s optimal information pro-
vision with respect to the consumer’s outside option value distribution. The analysis reveals
how the seller may want to tailor different pre-search signals to different consumer groups,
which forms a kind of discriminatory information provision. Because I will maintain the strictly
unimodal assumption on the distribution of U0 , Proposition 3.2 implies that I can focus on com-
parative statics regarding the threshold of the seller’s optimal upper-censorship signal, which I
denote as η∗ . The following proposition provides such a result for a non-pricing seller.

Proposition 5.1. Consider a non-pricing seller and assume Jp remains in the family of dis-
tributions satisfying Assumption 3.2. Then η∗ (weakly) increases when Jp increases in the
likelihood-ratio order.

The proposition suggests that it is optimal for a non-pricing seller to provide better pre-search
information when the distribution of U0 + p is higher (in the likelihood-ratio order). If the seller
can distinguish consumers or consumer groups with different outside option value distributions,
this implies that he tends to impart different information to them accordingly.29
The analysis can also be extended to pricing sellers in a more concrete setting. Specifically,
assume that the consumer’s outside option value now consists of two parts: U0 = W + , where
W and are independent random variables. Let fW and f denote the densities of W and
respectively. If the seller can observe W and tailor pre-search information and price accordingly,
then we have the following result on the seller’s discriminatory information provision:

Proposition 5.2. Assume f is strictly log-concave. For both pricing and non-pricing sellers, if
he can observe W , then η∗ (weakly) increases in W .30

The proposition suggests that it is optimal for the seller to provide better information to con-
sumers with higher average outside option value. For an intuition on this, notice that conditional
on higher W + p, the distribution of U0 + p is located more to the right and is thus convex over
a larger portion of [u − c, ū − c], which makes the trade-off mentioned below Proposition 3.1
favor a higher η∗ . This implies the stated result for a non-pricing seller. For a pricing seller, the
proof shows that although the seller may want to charge a lower p when W is higher, optimality
requires W + p to be increasing in W . We thus also have η∗ to be increasing in W .

28 European Union indeed has such a policy (see footnote 22).

29 Another implication of Proposition 5.1 is that a non-pricing seller tends to provide better information when p is
higher. Formally, if U0 has a log-concave density, then U0 + p will increase in p in the likelihood-ratio order. The
proposition then implies η∗ to be increasing in p.
30 This is in the strong set order if we have a pricing seller and the optimal design is not unique.

18
C. Lyu Journal of Economic Theory 213 (2023) 105722

The corollary below shows how discriminatory information provision affects consumer wel-
fare. For simplicity, it only considers the case of a non-pricing seller. This shuts down the
traditional effect of price discrimination and allows us to solely focus on informational discrimi-
nation.

Corollary 5.1. Consider a non-pricing seller and assume:

(A1) fW is log-concave
 and f is strictly log-concave;
(A2) arg maxx { fW (x − y)f (y)dy} is a singleton.31

Then there exists w ∗ such that when the seller can discriminate based on W , all consumers
with W < w ∗ becomes (weakly) worse-off and all consumers with W > w ∗ becomes (weakly)
better-off, compared to the case without discrimination.

The corollary suggests that the interests of consumers with different W are generally not
aligned. Those with higher W are more likely to benefit from discriminatory information provi-
sion, while the opposite is true for those with lower W . A particular implication of this is that a
consumer with high W may have incentive to voluntarily share it with the seller (e.g., by choos-
ing a low privacy setting). Since this will help the seller to also identify those with lower W , it
can lead to unraveling of W and harm consumers with lower average outside option values.32
As the discriminatory information provision has different welfare implications for different
consumers, a natural question is then what its effect on total consumer welfare is. The answer to
this is ambiguous and generally depends on the curvatures of the value distributions. If the CDF
of W + is convex over the support of U − c, the seller will provide full information without
discrimination. Hence allowing discrimination will merely harm consumers with low W . If the
CDF of W + is concave over [u − c, ū − c], the seller will provide no information without
discrimination. Then discrimination will only help by allowing some consumers with high W to
receive better information.
I conclude this subsection with a note on the related literature. Since higher outside option
value is equivalent to lower inside option value, the result in Proposition 5.2 can also be inter-
preted as that a seller with lower product quality tends to provide better pre-search information.
Anderson and Renault (2013) draw a similar lesson in their Appendix B.33 My result extends and
refines theirs in two ways. First, I allow for ex-ante heterogeneous consumers; second, I provide
a detailed comparative statics on the optimal signal threshold, while they only concern whether
any threshold information will be provided.

5.2. The effect of changing search cost

Technological developments can significantly reduce a consumer’s cost in accessing products.

For examples, modern means of transportation make it easier to visit physical stores; advanced
logistics systems make product shipping cheaper in online shopping; faster internet connection

31 These are satisfied, for example, when W and are normally distributed.
32 Such an effect is well-recognized for price discrimination (e.g., Acquisti et al. (2016) page 453).
33 Also related are earlier papers including Lewis and Sappington (1994), Sun (2011) and Bar-Isaac et al. (2010). These
papers only consider experience goods and restricted information structures.

19
C. Lyu Journal of Economic Theory 213 (2023) 105722

reduces the time needed for loading digital contents. These all help to reduce a consumer’s search
cost in various contexts.
The following proposition describes the effect of changing search cost with a non-pricing
seller.

Proposition 5.3. Consider a non-pricing seller and assume the following:

(A1) Jp admits a continuous density that is strictly quasi-concave over (−∞, ū].34
(A2) 1 − Fu is log-concave.

Then, when the search cost c decreases, we have:

(a) The seller’s optimal upper-censorship signal will become less informative.
(b) A consumer with any outside option value will become better-off.

When the product price is fixed and the regularity conditions hold, the proposition shows that
the consumer welfare will unambiguously increase when the search cost drops, although this gain
from lower search cost will be partially offset by coarser pre-search information. This holds not
only for consumers as a whole, but also for each individual consumer given her realized outside
option value.
A natural question is how the results will change if we have a pricing seller who can adjust
his price accordingly when the search cost changes. The general answer to this is ambiguous.
The reason is that a pricing seller can take advantage of reduced search cost not only by altering
information provision, but also by raising the product price. Since a higher price tends to moti-
vate better pre-search information provision (see footnote 29) but directly harms consumers, the
results in Proposition 5.3 can be reversed when the value distributions are such that the seller’s
price increment will be sufficiently large relative to the search cost’s drop.35

6. Information provision by competing sellers

In this section, I turn to consider multiple sellers and study the effect of competition on equi-
librium information provision.

6.1. The model

Consider a discrete choice model with horizontally differentiated products. There are N sell-
ers, each of whom sells a product with zero marginal cost. I call the product sold by seller i as
product i. There is a (representative) consumer, whose match value with product i is denoted as
Ui . Assume that {Ui }Ni=1 are independently drawn from a common distribution FU with compact
support [u, ū]. The consumer can consume one product and is assumed to have no outside option
for simplicity.

34 If one considers search costs less than some value c̄, this only needs to hold over [u − c̄, ū].
35 As has been noticed by Anderson and Renault (2006), when the search cost decreases, the seller’s optimal price can
increase faster than how the search cost changes. With ex-ante heterogeneous consumers in my setting, this especially
hurts the “infra-marginal” consumers with low outside option values and can thus lead to lower consumer surplus.

20
C. Lyu Journal of Economic Theory 213 (2023) 105722

Because it is not harder to consider pricing sellers than non-pricing sellers, I will assume all
sellers are pricing sellers. The analysis for non-pricing sellers is analogous and will lead to the
same qualitative result. The game proceeds as follows:

1. Each seller i decides his price pi and designs a pre-search signal about Ui . These are done
by all sellers simultaneously and observed by the consumer.
2. {Ui }N
i=1 are (secretly) realized and the pre-search signal realizations {Si }i=1 are generated to
N

the consumer.
3. After observing {Si }N i=1 , the consumer starts a sequential search process as in Weitzman
(1979). Specifically, at any time she decides whether to keep searching. If so, she chooses
one of the products to search, pay the search cost c, and then learns the product’s match
value. If not, she chooses one of the products that have been searched before for purchase,
and the game ends.36

I will call the setting above the SG-environment. As in Definition 2.1, I will say a product is
an experience good (EG) if the consumer will receive no additional information after searching
it. The environment is an EG-environment if all products are experience goods and everything
else is the same as above. In particular, the consumer still needs to pay the search cost c in order
to buy any product.

6.2. Equilibrium definitions

For any i, I use Zi to denote the Pandora’s index of the consumer’s posterior belief on Ui after
observing Si . The following lemma characterizes the consumer’s consumption outcome, where
part (a) has been discovered in several papers (Kleinberg et al., 2016; Armstrong, 2017; Choi et
al., 2018) and part (b) is trivial to see.

Lemma 6.1. We have the following results:

(a) In the SG-environment, the consumer will buy from a seller with the highest Ui ∧ Zi − pi .
Her expected surplus is E[maxi {Ui ∧ Zi − pi }].
(b) In the EG-environment, the consumer will buy from a seller with the highest E[Ui − c|Si ] −
pi . Her expected surplus is E[maxi {E[Ui − c|Si ] − pi }].

Let Gi and Hi denote the distributions of Ui ∧ Zi and E[Ui − c|Si ] respectively. The lemma
implies that in the SG-environment (resp., EG-environment), {(Gi , pi )}N i=1 (resp., {(Hi , pi )}i=1 )
N

is sufficient in determining the equilibrium outcomes. We can thus equivalently consider each
seller i’s decision as choosing the pair of (Gi , pi ) in the SG case and choosing (Hi , pi ) in the
EG case.
Given {(Gk , pk )}k=i with {Gk : k = i} being continuous, seller i’s best response problem in
the SG-environment can be written as:

max {pi Gk (x − pi + pk )dGi (x)} (6.1)
Gi , pi
k=i

36 As is standard in the literature, I assume that a product must be searched before purchase and it is free for the
consumer to recall an earlier searched product for purchase.

21
C. Lyu Journal of Economic Theory 213 (2023) 105722

s.t. Gi is a feasible distribution of Ui ∧ Zi (6.2)

Notice the integration in the objective function measures P (Ui ∧ Zi − pi ≥ Uk ∧ Zk − pk , ∀k),
which equals to the sale probability of seller i by Lemma 6.1.
Similarly, given {(Hk , pk )}k=i with {Hk : k = i} being continuous, seller i’s best response
problem in the EG-environment can be written as:

max {pi Hk (x − pi + pk )dHi (x)} (6.3)
Hi , pi
k=i
s.t. Hi MP S FU −c (6.4)
I define the equilibria with the two types of goods as follows.

Definition 6.1 (SG-equilibrium). A pair of (G∗ , p ∗ ) is a symmetric SG-equilibrium if G∗ is con-

tinuous and (G∗ , p ∗ ) solves optimization (6.1) – (6.2) given (Gk , pk ) = (G∗ , p ∗ ) for all k = i.

Definition 6.2 (EG-equilibrium). A pair of (H ∗ , p ∗ ) is a symmetric EG-equilibrium if H ∗ is

continuous and (H ∗ , p ∗ ) solves optimization (6.3) – (6.4) given (Hk , pk ) = (H ∗ , p ∗ ) for all
k = i.

I note that in the above definitions, I have focused on equilibria where G∗ or H ∗ is continuous.
This avoids the discussion of ties in the consumer’s choices, and also suffices for presenting my
results below.37

6.3. Equilibrium with a large number of sellers

I assume the following regularity condition holds:

Assumption 6.1. FU admits a log-concave density fU , which is continuously differentiable with

bounded derivative over [u, ū]. Moreover, fU is strictly positive over [u, ū].38

Under Assumption 6.1, Theorem 3 in Hwang et al. (2019) shows that there is a unique sym-
metric EG-equilibrium. Given any number of competing sellers N , I denote this equilibrium as
(HNex , pN
ex ). My main result in this section is the following:

Proposition 6.1. Under Assumption 6.1, there exists N ∗ < ∞ such that for all N ≥ N ∗ ,
(G∗ , p ∗ ) = (HNex , pN
ex ) is a symmetric SG-equilibrium when there are N competing sellers.

The proposition suggests that when the number of competing sellers is sufficiently large, the
distribution of Ui ∧ Zi in the SG-environment will be the same as the distribution of E[Ui − c|Si ]
in the EG-environment, and the product prices in the two environments will also be the same. By
Lemma 6.1, this implies that the purchase outcome, sellers’ profits and consumer welfare will

37 Requiring H ∗ to be continuous in the EG-equilibrium is actually without loss of generality because Hwang et al.
(2019) shows that it holds in any symmetric equilibrium with EG.
38 The assumption that f is strictly positive over [u, ū] is not crucial for my results. However, it allows me to directly
U
refer to certain results in Hwang et al. (2019), who make such kind of assumption. Details on this are available upon
request.

22
C. Lyu Journal of Economic Theory 213 (2023) 105722

all be the same in the two environments. Actually, the proposition’s proof shows that when N
is large, any equilibrium profile of signals in the EG-environment also serves as an equilibrium
profile of signals in the SG-environment. Under these signals, Zi ≤ Ui almost surely for all i,
and thus the consumer will always purchase from the first seller being searched by the Pandora’s
rule. The discrepancy between the two types of environments therefore vanishes when there is
sufficient competition among sellers.
In the EG-environment, Hwang et al. (2019) shows that the equilibrium information will con-
verge to full information when the number of competing sellers goes to infinity. The following
corollary of Proposition 6.1 extends this result to the SG-environment.

Corollary 6.1. Assume Assumption 6.1 holds. For any > 0, there exists N < ∞ such that
for all N ≥ N : there exists a symmetric SG-equilibrium in which G∗ (x) = FU −c (x) for all
x ≤ ū − c − .

The corollary implies that when the number of competing sellers goes to infinity, the equi-
librium distribution of Ui ∧ Zi will converge to FU −c and hence the equilibrium pre-search
information will converge to full information. This suggests that strong competition leads to
informational efficiency also for search goods.
I note that Proposition 6.1 and Corollary 6.1 only consider one symmetric equilibrium for
search goods. Although I conjecture that the symmetric equilibrium will be unique when N
is large, showing this formally is beyond the scope of this paper.39 The main difficulty is that
without a complete and simple characterization for the feasible set of Gi , it is hard to derive
meaningful necessary optimality conditions for the best response problem (6.1) – (6.2), which
are needed to establish the equilibrium uniqueness.

6.4. Discussion: equilibrium with a small number of sellers

I have studied the case with a monopoly seller and the case with a large number of competing
sellers. The remaining question is what will happen with a small number of sellers. One may
guess that the equilibrium signal informativeness will be in the middle of the two extreme cases.
Unfortunately, however, a formal analysis for this is difficult.
Similar to what has been done in the proof of Proposition 6.1, one tentative approach is to first
consider a “relaxed game”, where each seller chooses Gi only subject to the two constraints in
the Relaxed Problem (3.4) – (3.6). After deriving the equilibrium for this relaxed game, one can
check whether the equilibrium Gi can be induced by some signal. If so, this will indeed be an
equilibrium of the original game. The main difficulty is that there is no easy sufficient condition
to show a distribution is inducible for Ui ∧ Zi . Throughout the paper, I have been showing this by
explicitly constructing the underlying signal. However, such a construction is typically difficult
when Gi has a degree of complexity that one must face in the competing environment. I thus
leave this analysis for future studies.

39 I have the conjecture because of the similarity between the two types of goods and that we do have equilibrium
uniqueness in the EG-environment.

23
C. Lyu Journal of Economic Theory 213 (2023) 105722

7. Additional discussions

7.1. Comparison with Anderson and Renault (2006)

As is mentioned in the introduction, Anderson and Renault (2006) studies a similar setting as
mine while assuming that the consumers are ex-ante homogeneous without private information.
Under this assumption, they solve the optimal signal and show that it suffices for the seller to
provide threshold information, which just informs the consumer whether her match value is
above certain threshold.
My analyses have partially extended their result that the optimal design involves threshold
information to the situation where the consumers’ outside option values are private and admit a
unimodal distribution. Unlike in Anderson and Renault (2006), however, the optimal signal in my
setting typically also requires fully revealing matches below the threshold, especially when the
distribution of U0 is strictly unimodal. To understand this discrepancy, let η∗ denote the threshold
net-match-utility under the optimal design. In Anderson and Renault (2006), U0 is common for
all consumers and we have U0 > η∗ − p for sure. Thus once a consumer has learned her U − c is
below η∗ , she will not search for the product anyway. Providing additional below-threshold in-
formation is therefore useless. In my setting, however, U0 has a non-degenerate distribution and
typically some consumers will have U0 < η∗ − p. Providing additional below-threshold infor-
mation allows these consumers to tell how far their net-match-utilities are below the threshold,
and can thereby alter their decisions. If the CDF of U0 is strictly convex below η∗ − p, which
is always the case under the strict unimodal assumption on U0 , this will indeed increase these
consumers’ total purchase probability in light of my discussion right below Proposition 3.1. Thus
finer below-threshold information can attract more consumers with relatively low outside option
values into purchase.

7.2. Incentive for multi-stage information provision

I have been studying situations where the seller controls the consumer’s pre-search informa-
tion, while many earlier studies reviewed in the introduction have focused on cases where the
seller controls the consumer’s post-search information. A natural question is then what if the
seller can control information at both stages.
There is a simple answer to this question: the seller never benefits from multi-stage informa-
tion provision. As is mentioned in Section 4, an SG seller’s pre-search design problem is more
constrained than the EG seller’s. This implies that the seller is weakly better-off when the con-
sumer cannot receive additional information after search. Thus even if the seller can control both
pre-search and post-search information, it is optimal for him to only provide information at the
pre-search stage. The equilibrium outcome will hence be identical to that with an EG seller.

7.3. Incentive to subsidize search

I have been assuming that the seller cannot incentivize consumer search by directly subsi-
dizing it. An interesting question is how the equilibrium outcome may change if we allow such
subsidization. When the distribution of outside option value is unimodal and the seller is a pric-
ing seller, a short answer to this is available: the seller has no incentive to subsidize search, and
thus the equilibrium outcome will remain unchanged. A detailed analysis for this is provided in
Appendix A.3.

24
C. Lyu Journal of Economic Theory 213 (2023) 105722

7.4. Multimodal outside option

Most of my analyses have relied on the unimodal assumption on the consumer’s outside op-
tion value distribution. What if we go beyond this assumption? Although my relaxed problem
approach is not guaranteed to find the seller’s optimal design in that case, one can still try the
following procedure: (1) solving the Relaxed Problem (3.4) – (3.6); (2) constructing a signal that
implements the Relaxed Problem’s solution. If the second step is feasible, then the signal con-
structed will be optimal. In Appendix A.4, I discuss when this procedure works and provide an
analytical example. It is shown that the optimal signal can involve multiple pooling and revealing
intervals, which is more complex than under the unimodal assumption.

8. Concluding remarks

I have examined the optimal pre-search information provision by search goods sellers. A re-
laxed problem approach is developed to solve the information design problem, which is more
challenging than its experience goods counterpart. When the consumer’s outside option value
distribution is unimodal, I show the optimal design features a simple upper-censorship structure,
based on which several applications can be considered. These include studies on (1) compari-
son between search goods and experience goods, (2) discriminatory information provision, (3)
the effect of reduction in search cost and (4) competition by multiple sellers. In some of these
applications, my approach has allowed existing results of experience goods to be qualitatively
extended to search goods.
Many research questions still remain open. Examples include how to solve the optimal design
with multimodal outside option in general, and how to characterize the equilibrium with a small
number of competing sellers. These probably require sharper characterizations for the feasible
set of the effective-search-value distribution. My lemmas in Section 3.1 and the relaxed problem
approach can be considered as a starting point for this more general research agenda.
The paper’s setting can also be generally understood as having a sender persuading a receiver
to take certain action that involves two costly steps, after the first of which the receiver will in-
evitably learn additional information and can choose to quit. This kind of environment is very
common in practice. For example, one can consider an entrepreneur persuading an investor to fi-
nance a project involving two phases, before the second of which the investor can choose whether
to continue funding based on the information revealed in the first phase. My approach may also
be useful in such applications.

Declaration of competing interest

I have no interest to declare for my paper “Information design for selling search goods and
the effect of competition”.

Data availability

No data was used for the research described in the article.

25
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. A.1. j (·) in the example of Appendix A.2.

Appendix A. Additional results and discussions

A.1. Tentative existing approaches to the seller’s problem

As is mentioned in the introduction, the first moment of the consumer’s posterior belief does
not suffice for deciding the purchase outcome. In fact, with continuous U0 , the sender’s objec-
tive function must depend on infinitely many moments of the posterior belief. This makes the
approaches in Dworczak and Martini (2019) and Dworczak and Kolotilin (2019) hard to apply.
Another tentative approach is the concavification method in Kamenica and Gentzkow (2011). As
is well known, this method is hardly tractable unless U takes no more than three values, which
is very restrictive in the current context. In particular, the clean signal structures derived in this
paper and Anderson and Renault (2006), like the threshold structure, would not be available if
U only takes a few discrete values. A third approach is the linear programming approach in
Kolotilin (2018), which handles receiver private information. With this approach, one will have
to consider an optimization over the joint distribution of (U, Z), which turns out to be very chal-
lenging. A fourth approach is to draw upon a result in Guo and Shmaya (2019) and transform
the problem into a nonlinear optimization over some thresholds as in their Section 4. Unless
one assumes U0 takes only finite values, the optimization will be over infinite-dimensional func-
tions. Even if U0 takes finite values, it is still hard to derive semi-analytical solution for such a
non-linear optimization.

A.2. SG vs. EG with a pricing seller

In Section 4, I have shown that with a non-pricing seller, turning an experience good into
a search good may unintentionally crowd-out the seller’s pre-search information provision and
decrease the total consumer welfare. With a numerical example, I show that these can also happen
with a pricing seller below.
Assume c = 0.1, U ∼ Uniform[0, 1] and U0 has an approximate skew-normal distribution
(Ashour and Abdel-hameed, 2010). Specifically, U0 has density:
x − ξ   x −ξ
j (x) = 2φ F −8×
0.3 0.3
where φ is the standard normal density; F is defined in equation (2.1) in Ashour and Abdel-
hameed (2010); ξ is chosen such that the mean of U0 is normalized to zero. The density of U0 is
plotted in Fig. A.1.
The comparison for equilibrium outcomes of the two types of goods is summarized in Ta-
ble A.1. As one can see, when the product turns into a search good, the threshold of the seller’s

26
C. Lyu Journal of Economic Theory 213 (2023) 105722

Table A.1
Equilibrium outcomes in the example of Appendix A.2.
Product type η∗ Equilibrium Price Consumer Surplus
Experience Goods 0.284 0.387 0.0400
Search Goods 0.273 0.414 0.0366
% Change -3.89 6.93 -8.33

optimal upper-censorship signal (η∗ ) and the total consumer surplus both drop. Moreover, the
consumer welfare decreases due to not only poorer pre-search information, but also higher prod-
uct price.

A.3. Incentive for subsidizing search

In this appendix, I show a pricing seller has no incentive to subsidize the consumer’s search.
Given any original product price, consider two strategies of the seller:

A. Offer a subsidy y > 0 per search.

B. Give no search subsidy, but offer a price discount y for the product.

Without loss of generality, assume y is less than the original product price. Then, we have the
following observation:

Proposition A.1. Assume J is unimodal (i.e., Assumption 3.3 holds). A pricing seller’s maximal
profit under strategy A is lower than that under strategy B.40

The proposition implies that subsidizing for search is always an inferior option for the seller
than directly lowering the product price. Two intuitions are underlying this result. First, given
any purchase probability, strategy A is more costly to implement than strategy B since it offers
subsidy not only to those who purchase, but also to those who search without purchase. Second,
by making search cheaper, strategy A weakens a consumer’s reliance on pre-search informa-
tion. This makes it less effective to manipulate consumer behavior by controlling the pre-search
information and thus decreases the maximal purchase probability.

A.4. Multimodal outside option

When Jp is multimodal over [u − c, ū − c], my results about the seller’s optimal design in
Section 3.3 does not apply. In this case, one can try the following approach for solving the non-
pricing seller’s optimal signal: (1) solving the Relaxed Problem (3.4) – (3.6); (2) constructing a
signal that implements the Relaxed Problem’s solution. If the second step is possible, then the
signal constructed will be optimal.
Let G∗RP denote a solution to the Relaxed Problem. Let ẑ(a, b) be the Pandora’s index of the
posterior belief on U after learning U − c ∈ [a, b]. An important case where the above procedure
works is when G∗RP features a monotone-partitionable structure.

40 This holds strictly as long as there will be some search without purchase under strategy A.

27
C. Lyu Journal of Economic Theory 213 (2023) 105722

Definition A.1. Distribution G is called monotone-partitionable if there exists ({si }ni=0 , {ηi }ni=0 )
such that u − c = s0 < ... < sn = ū − c; si ≤ ηi ≤ si+1 ∀i; and
⎧
⎪
⎪ FU −c (x) si ≤ x < ηi
⎪
⎨F
U −c (ηi ) ηi ≤ x < (ηi + c) ∧ ẑ(ηi , si+1 )
G(x) = (A.1)
⎪
⎪ FU (x) (ηi + c) ≤ x < ẑ(ηi , si+1 )
⎪
⎩
FU −c (si+1 ) ẑ(ηi , si+1 ) ≤ x ≤ si+1

With a similar argument as in the proof of Lemma B.2 (in Appendix B.3.1), it is easy to see
that if G∗RP is monotone-partitionable, it can be induced by a monotone-partitioning signal that
fully reveals U − c ∈ [si , ηi ] ∀i, and pools U − c ∈ [ηi , si+1 ] separately for each i. Intuitively,
such a signal first partitions net-match-utilities into subintervals {[si , si+1 ]}ni=1 , and then imposes
an upper-censorship signal on each of these subintervals. Thus, if the Relaxed Problem turns out
to have a monotone-partitionable solution, one can conclude that it is indeed inducible and the
optimal signal can be the monotone-partitioning signal described above.
In the rest of this appendix, I consider a particular situation where the Relaxed Problem in-
deed features a monotone-partitionable solution. This both illustrates how the procedure above
works and shows that when Jp is not unimodal, the optimal signal can involve more complicated
structure than upper-censorship.41
Assume the following assumption holds:

Assumption A.1. Jp admits a continuous density jp over [u − c, ū − c]. There exist r1 and r2
(r1 < r2 ) in (u − c, ū − c) such that jp (·) is strictly increasing on [u − c, r1 ] and [r2 , ū − c], and
is strictly decreasing on [r1 , r2 ].

Further assume for now that there exists η∗ ∈ [u − c, r1 ] and ξ ∗ ∈ [r2 , ū − c] such that
ẑ(η∗ , ξ ∗ ) ∈ [r1 , r2 ] and:
 
Jp (η∗ + c) ∧ ẑ(η∗ , ξ ∗ ) − Jp (η∗ )  ∗ ∗  Jp (ξ ∗ ) − Jp (ẑ(η∗ , ξ ∗ ))
= j p ẑ(η , ξ ) = (A.2)
(η∗ + c) ∧ ẑ(η∗ , ξ ∗ ) − η∗ ξ ∗ − ẑ(η∗ , ξ ∗ )
Fig. A.2 plots Jp satisfying Assumption A.1 with (η∗ , ξ ∗ ) satisfying condition (A.2). Intuitively,
the condition requires that the two black secants of Jp plotted in the graph have the same slope,
which is equal to the slope of Jp at ẑ(η∗ , ξ ∗ ).
Then a solution to the Relaxed Problem (G∗RP ) turns out to be monotone-partitionable as in
Definition A.1, with n = 2, s0 = u − c, η1 = η∗ , s1 = ξ ∗ and η2 = s2 = ū − c. As the discussion
earlier suggests, this G∗RP can be induced by a monotone-partitioning signal that fully reveals
net-match-utilities in [u − c, η∗ ] ∪ [ξ ∗ , ū − c] and pools the rest in the middle. Thus G∗RP is
indeed feasible for the seller, and the signal inducing it is optimal.
To check that G∗RP suggested above indeed solves the Relaxed Problem, one can apply The-
orem 3.1. Namely, one can find a convex function v(·) and an increasing function ρ(·) such that
(G∗RP , v, ρ) satisfies conditions (C1) – (C4) in Theorem 3.1. Fig. A.3 gives a graphical illustra-
tion for this. The fact that conditions (C1) and (C2) hold is evident from the graph (notice ρ(·)
is represented by the difference between the solid blue curve and the dotted blue curve). (C4)
holds since G∗RP is indeed inducible and thus satisfies the constraints. To check (C3), notice that:

41 More examples are available upon request.

28
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. A.2. Jp (·) satisfying Assumption A.1 with (η∗ , ξ ∗ ) satisfying (A.2). The red curve is Jp ; the black segments are
secants of Jp with common slope jp (ẑ(η∗ , ξ ∗ )).

Fig. A.3. Graphical Check of Optimality Conditions for (G∗RP , v, ρ). The red curve is Jp ; the solid blue curve is v + ρ;
the dashed blue curve is v; the orange area on x-axis is the support for G∗RP .

(1) over regions where v(·) is strictly convex, G∗RP equals to FU −c ; (2) over the region where
ρ(·) is strictly increasing, G∗RP equals to FU . Given that G∗RP satisfies the constraints, these
imply that (C3) indeed holds. (A detailed proof for this is similar to that for Proposition 3.1. See
Observation (4) in Appendix B.3.3.)
Up to now, I have been assuming the existence of (η∗ , ξ ∗ ) such that ẑ(η∗ , ξ ∗ ) ∈ [r1 , r2 ] and
(A.2) holds. Actually, when such η∗ and ξ ∗ do not exist, one can still derive the optimal signal
by solving the Relaxed Problem, but just ends up with a simpler signal structure, which can
be thought of as a corner solution. In particular, the optimal signal can feature either upper-
censorship or lower-censorship (i.e., it pools values below a threshold and fully reveals values
above the threshold). In general, as long as Assumption A.1 holds, the Relaxed Problem always
has a monotone-partitionable solution, and the seller’s optimal signal can be found with the above
procedure.

Appendix B. Proofs for Section 3

B.1. Proof for Lemma 3.1

Proof. First notice by the definition of Pandora’s index (3.1), we have E[(U − Z)+ − c|S] = 0
for any signal realization S. Since (U − Z)+ = U − U ∧ Z, we have E[U − c|S] = E[U ∧ Z|S]
(i.e., the posterior mean of net-match-utility and the effective-search-value are equal).42

42 Throughout my proofs, the (in)equalities involving conditional expectation or Z are understood as holding almost
surely when appropriate.

29
C. Lyu Journal of Economic Theory 213 (2023) 105722

Next, I show that conditioning on any signal realization S, the conditional distribution of
U ∧ Z single crosses that of U − c from below. To see this, given any S, let zS be the Pandora’s
index of posterior belief φ(·; S). Then, for any u < zS , we have:

P (U ∧ Z ≤ u|S) = P (U ∧ zS ≤ u|S) = P (U ≤ u|S) ≤ P (U − c ≤ u|S)

For any u ≥ zS , we have:

P (U ∧ Z ≤ u|S) = P (U ∧ zS ≤ u|S) = 1 ≥ P (U − c ≤ u|S)

Together with the observation E[U − c|S] = E[U ∧ Z|S] above, this single crossing property
implies that U ∧ Z is MPC of U − c conditional on S. Then the result that U ∧ Z is MPC of
U − c follows because MPC is preserved under mixture. (To see this, just notice for any convex
function v(·), E[v(U ∧ Z)|S] ≤ E[v(U − c)|S] ⇒ E[v(U ∧ Z)] ≤ E[v(U − c)].)
To see U ∧ Z = U − c under fully revealing signal, take φ = δU into (3.1), where δU is the
Dirac measure at U . It is easy to derive zδU = U − c then. 

B.2. Proof for Theorem 3.1

Proof. For the first claim (the part before “Moreover”), let distribution G, convex function v
and increasing function ρ satisfy all the conditions. Let Ĝ be another distribution that satisfies
constraints (3.5) and (3.6). We have:

Jp (x)dG(x)
  
= [Jp (x) − v(x) − ρ(x)]dG(x) + v(x)dG(x) + ρ(x)dG(x)
  
= [Jp (x) − v(x) − ρ(x)]dG(x) + v(x)dFU −c (x) + ρ(x)dFU (x)
  
≥ [Jp (x) − v(x) − ρ(x)]d Ĝ(x) + v(x)d Ĝ(x) + ρ(x)d Ĝ(x)

= Jp (x)d Ĝ(x)

The second equality is because G satisfies the complementary-slackness condition (C3). For the
inequality in fourth line, notice it holds for each term on the two sides. Specifically,
(1) Since
 Ĝ is MPC of FU −c , its support is a subset
 of [u − c, ū − c]. Then by condition (C2),
we know [Jp (x) − v(x) − ρ(x)]d Ĝ(x) ≤ 0 = [Jp (x) − v(x) − ρ(x)]dG(x).

(2) v being convex over [u − c, ū − c] and Ĝ MP S FU −c imply that v(x)dFU −c (x) ≥

v(x)d Ĝ(x).

(3) ρ being increasing on [u − c, ū] and Ĝ F OD FU imply that ρ(x)dFU (x) ≥

ρ(x)d Ĝ(x).
Thus, the inequality above holds. Since Ĝ is an arbitrary feasible distribution, I conclude that
G is optimal.
For the second claim (the part after “Moreover”), notice for Ĝ to be another optimal solution,
we need the inequality above to hold as equality. Based on observations (1) – (3) above, this
requires equality to hold for each term. Specifically, we need:

30
C. Lyu Journal of Economic Theory 213 (2023) 105722

 
First, [Jp (x) − v(x) − ρ(x)]d Ĝ(x) = [Jp (x) − v(x) − ρ(x)]dG(x) = 0. Since Jp (x) −
v(x) − ρ(x)≤ 0 on the support  of Ĝ, this implies condition
 (C2) holds
 for Ĝ.
Second, v(x)d Ĝ(x) = v(x)dFU −c (x) and ρ(x)d Ĝ(x) = ρ(x)dFU (x). This means
that condition (C3) holds for Ĝ.
Thus, the triple (Ĝ, v, ρ) satisfies (C1) – (C4). 

B.3. Proof for Proposition 3.1

Since rpmax = ū − c implies that Jp (·) is convex over [u − c, ū − c], part (a) is directly implied
by Lemma 3.1. I prove part (b) with the following three steps.

B.3.1. Step 1: preliminaries

First, we need the following lemma regarding z(·):

Lemma B.1.

(a) ū − c > z(η) > η for η ∈ [u − c, ū − c); z(η) = η = ū − c for η = ū − c.

(b) z(·) is continuous and strictly increasing on [u − c, ū − c].

Proof. (a) The result for η = ū − c holds obviously by the definition of the Pandora’s index.
When η < ū − c, the definition of the Pandora’s index implies E[(U − z)+ − c|U − c ≥
η]|z=z(η) = 0. In contrast, E[(U − z)+ − c|U − c ≥ η]|z=η = E[(U − η)+ |U − η ≥ c] − c =
E[U − η|U − η ≥ c] − c > 0, where the inequality is strict because P [U − η > c] > 0;
E[(U − z)+ − c|U − c ≥ η]|z=ū−c = E[(U − (ū − c))+ − c|U − c ≥ η] < E[(ū − (ū −
c))+ − c|U − c ≥ η] = 0, where the inequality is strict because P (ū > U ≥ η + c) > 0.
Since E[(U − z)+ − c|U − c ≥ η] is decreasing in z, these imply ū − c > z(η) > η.
(b) Suppose η1 , η2 ∈ [u − c, ū − c] are such that η1 < η2 but z(η1 ) ≥ z(η2 ). Then by the result
of (a), we must have η1 < η2 < z(η2 ) ≤ z(η1 ) < ū − c. By the definition of the Pandora’s
index, we have:

E[(U − z(η1 ))+ − c|U − c ≥ η1 ] = 0 (B.1)

E[(U − z(η2 ))+ − c|U − c ≥ η2 ] = 0 (B.2)
 
The later implies that E[ (U − z(η2 ))+ − c 1{U −c≥η2 } ] =
 0, where 1 denotes the indicator
function. This further implies that E[ (U − z(η2 ))+ − c 1{U −c≥η1 } ] < 0 because the event
{η2 > U − c > η1 } has strictlypositive probability and (U − z(η2 ))+ − c < 0 on it. Thus,
wehave E[ (U − z(η2 ))+ − c |U − c ≥ η1 ] < 0. Since z(η2 ) ≤ z(η1 ), this further implies
E[ (U − z(η1 ))+ − c |U − c ≥ η1 ] < 0, which contradicts with (B.1). Therefore, we must
have η1 < η2 ⇒ z(η1 ) < z(η2 ).
 ū
To show the continuity of z(·), define ψ(z; η) := η+c [(x − z)+ − c]dFU (x). Then by the
definition of z(·), we must have ψ(z(η); η) = 0. Moreover, it is easy to see that the value of
z satisfying ψ(z; η) = 0 is unique when η < ū − c. Thus,

z(η) = arg max −(ψ(z; η))2

z∈[u−c,ū−c]

for any η < ū − c. Because ψ(z, η) is continuous in (z, η) due to the assumption that FU is
continuous, this implies that z(·) is continuous at any η ∈ [u − c, ū − c) by the Maximum

31
C. Lyu Journal of Economic Theory 213 (2023) 105722

Theorem. To show it is also continuous at η = ū − c, just notice the results in (a) implies
limηū−c z(η) = ū − c = z(ū − c). 

Next, I provide the detailed formula of Gη , the distribution of U ∧ Z under an upper-

censorship signal with threshold η.

Lemma B.2.
⎧
⎪
⎪ FU −c (x) x<η
⎪
⎨F
U −c (η) η ≤ x < (η + c) ∧ z(η)
Gη (x) = (B.3)
⎪
⎪ F U (x) (η + c) ≤ x < z(η)
⎪
⎩
1 z(η) ≤ x

(The third piece vanishes if η + c ≥ z(η).)

Proof. Under the upper-censorship signal with threshold η, all match values with U − c < η
are fully revealed while the others are pooled together. Therefore, when U − c < η, we have
Z = U − c and thus U ∧ Z = U − c 43 ; when U − c ≥ η, we have Z = z(η) and thus U ∧ Z =
U ∧ z(η) ≥ (η + c) ∧ z(η). Therefore, we have:

• For x < η, U ∧ Z ≤ x if and

 only if U − c ≤ x. Thus Gη (x) = FU −c (x).
• For x ∈ η, (η + c) ∧z(η) , U ∧ Z ≤ x if and only if U − c ≤ η. Thus Gη (x) = FU −c (η).
• For x ∈ η + c, z(η) , U ∧ Z ≤ x if and only if either U − c < η or U ∧ z(η) ≤ x. Since
z(η) > x ≥ η + c here, this is equivalent to having U ≤ x. Thus Gη (x) = FU (x).
• For x ≥ z(η), obviously U ∧ Z ≤ x for sure. Thus Gη (x) = 1. 

B.3.2. Step 2: existence of such an η∗ in Proposition 3.1(b)

Lemma B.3. Under the assumptions in Proposition 3.1(b), there exists η∗ ∈ [η0 , rpmax ] such that
one of conditions (i) and (ii) is satisfied. Moreover, z(η∗ ) ≥ rpmin for any such η∗ .

Proof. I first show some basic properties of the function  and the set [η0 , rpmax ].
 
J (η+c)∧z(η) −J (η)  
By definition, (η) = p (η+c)∧z(η)−ηp − jp z(η) . Since Jp (·), jp (·) and z(·) are all con-
tinuous and z(η) > η (thus the denominator is non-zero) for any η < ū − c (by Lemma B.1(a)),
(·) is continuous on [u − c, rpmax ].
By definition, η0 = inf{η ∈ [u − c, ū − c] : z(η) ≥ rpmin }. Because Lemma B.1(a) implies
z(rpmin ) ≥ rpmin , we have η0 ≤ rpmin and thus the set [η0 , rpmax ] is non-empty. Moreover, since
z(·) is continuous and increasing (Lemma B.1 (b)), we have z(η) ≥ rpmin for any η ≥ η0 . This
verifies that the desired η∗ must satisfy z(η∗ ) ≥ rpmin .
Now, I assume that no η∗ ∈ [η0 , rpmax ] satisfies condition (ii) and show that there then must
exist η∗ ∈ [η0 , rpmax ] satisfying condition (i). For this purpose, I show (η0 ) ≤ 0 and (rpmax ) >
0.

43 To check this, notice z = u − c solves equation (3.1) when φ places probability 1 at u.

32
C. Lyu Journal of Economic Theory 213 (2023) 105722

• Check (η0 ) ≤ 0:
If η0 = u − c, then the assumption that no η∗ ∈ [η0 , rpmax ] satisfies condition (ii) directly
implies that (η0 ) < 0.
If η0 > u − c, then by the continuity of z(·) we must have z(η0 ) = rpmin and thus Jp is convex
over [η0 , z(η0 )]. This convexity implies (η0 ) ≤ 0.44
• Check (rpmax ) > 0:
Since rpmax < (rpmax + c) ∧ z(rpmax ) ≤ z(rpmax ) by Lemma B.1(a), (rpmax ) > 0 is implied by
the concavity of Jp (·) over [rpmax , ū − c] and the strict concavity of it over [rpmax , rpmax + ]
for some > 0.

These observations together with the continuity of (·) imply that there exists η∗ ∈ [η0 , rpmax ]
s.t. (η) = 0 by the Intermediate Value Theorem. Since condition (ii) does not hold for this η∗
by my assumption earlier, it must satisfy condition (i). 

B.3.3. Step 3: optimality of upper-censorship signal with threshold η∗

Proof. Pick any η∗ ∈ [η0 , rpmax ] satisfying one of the conditions (i) and (ii) in the proposition
(existence has been shown in Step 2). Then (η∗ ) ≥ 0, where the inequality is strict only if
η∗ = η0 = u − c. It suffices to show Gη∗ solves the Relaxed Problem (3.4) – (3.6).
To ease notation, fixing an η∗ , I denote z(η∗ ) as z∗ for short. Moreover, I use ( · ; b, (x0 , y0 ))
to denote an affine function with slope b passing point (x0 , y0 ).
Define
 ∗
Jp (x)  x<η
v(x) =   ∗  ∗ 
 x; jp z , (η + c) ∧ z∗ , Jp ((η∗ + c) ∧ z∗ ) x ≥ η∗

By construction, v follows Jp when x < η∗ , is affine with slope jp (z∗ ) when x ≥ η∗ , and inter-
sects with Jp at x = (η∗ + c) ∧ z∗ .
Define
⎧
⎪
⎨0 x < (η∗ + c) ∧ z∗
ρ(x) = Jp (x) − v(x) (η∗ + c) ∧ z∗ ≤ x ≤ z∗
⎪
⎩
Jp (z∗ ) − v(z∗ ) x > z∗
Notice if η∗ + c ≥ z∗ , then the second piece vanishes and ρ ≡ 0 (Jp (z∗ ) − v(z∗ ) = 0 in this case
by definition of v). If η∗ + c < z∗ , ρ equals to the difference between Jp and v over interval
[η∗ + c, z∗ ], and is constantly extended to regions above z∗ or below η∗ + c.
Now, it suffices to check that Gη∗ (whose formula is given in Lemma B.2), v and ρ satisfy the
optimality conditions in Theorem 3.1. This is done by showing four observations below. While
the detailed proofs are abstract and tedious, I plot the key functions in Fig. B.1, which serves as a
graphical check for the optimality conditions. I highlight that by construction and the conclusion
in Lemma B.3, we have η∗ ≤ rpmax and z∗ ≥ rpmin (i.e., η∗ is always on the convex portion of Jp
and z∗ is always on the concave portion of Jp ). This fact is used repeatedly below.

    
44 To see this, notice (η0 ) = η(η0 +c)∧z(η0 ) j (x)dx (η0 + c) ∧ z(η0 ) − η0 − j z(η0 ) ≤
 
0
 (η +c)∧z(η )    
η0
0 0 j z(η ) dx
0 (η0 + c) ∧ z(η0 ) − η0 − j z(η0 ) = 0, where the inequality is because j (x) is
increasing on [η0 , z(η0 )].

33
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. B.1. Graphical Check of Optimality Conditions for (Gη∗ , v, ρ). The red curve is Jp ; the solid blue curve is v + ρ;
the dashed blue curve is v; the orange area on x-axis is the support for Gη∗ .

Observation (1). v is convex over [u − c, ū − c]. Moreover, v(η∗ ) ≥ Jp (η∗ ), which holds as
equality if η∗ > u − c.

Subproof. To show v(η∗ ) ≥ Jp (η∗ ), just notice that

    
v(η∗ ) = η∗ ; jp z∗ , (η∗ + c) ∧ z∗ , Jp ((η∗ + c) ∧ z∗ )
  
=Jp ((η∗ + c) ∧ z∗ ) − jp z∗ (η∗ + c) ∧ z∗ − η∗
 
≥Jp ((η∗ + c) ∧ z∗ ) − Jp ((η∗ + c) ∧ z∗ ) − Jp (η∗ ) = Jp (η∗ )
 
J (η∗ +c)∧z∗ −J (η∗ )  
where the inequality holds due to the condition p (η∗ +c)∧z∗ −ηp∗ ≥ jp z∗ , which is implied
by either of condition (i) and condition (ii). If η∗ > u − c, it must hold as equality (i.e., condition
(i) holds) and thus v(η∗ ) = Jp (η∗ ). Now I show convexity of v. If η∗ = u − c, convexity trivially
holds since v is affine then. Thus we only need to consider the case where η∗ > u − c and
condition (i) holds (i.e., (η∗ ) = 0). In this case, notice that both pieces of v are separately
convex, and that v(η∗ ) = Jp (η∗ ) shown above implies continuity of v at η∗ . Thus it suffices to
show convexity around η∗ . Specifically, we need to show v+ (η∗ ) ≥ v− (η∗ ), where v+ and v−
denote v’s right and left derivatives respectively. Consider two cases:
  J (η∗ +c)∧z∗ −J (η∗ )
• (η∗ +c) ∧z∗ ≤ rpmax . In this case, notice v+ (η∗ ) = jp z∗ = p (η∗ +c)∧z∗ −ηp∗ ≥ jp (η∗ ) =
v− (η∗ ), where the inequality is by convexity of Jp over [u − c, rpmax ].
• (η∗ + c) ∧ z∗ > rpmax . In this case, v (x) = jp (z∗ ) ≤ jp (x) ∀x ∈ [rpmax , z∗ ) due to the con-
cavity of Jp over [rpmax , z∗ ]. Since the affine part of v intersects Jp at (η∗ + c) ∧ z∗ by
construction, this implies that v(x) ≥ Jp (x) for all x ∈ [rpmax , (η∗ + c) ∧ z∗ ) and, in partic-
ular, v(rpmax ) ≥ Jp (rpmax ).45 Since it has been shown that v(η∗ ) = Jp (η∗ ), this implies that

45 If J is strictly concave over [r max , ū − c], then the inequalities hold strictly.
p p

34
C. Lyu Journal of Economic Theory 213 (2023) 105722

Jp (r max )−Jp (η∗ )

v+ (η∗ ) ≥ p
rpmax −η∗ ≥ jp (η∗ ) = v− (η∗ ), where the second inequality is due to convex-
ity of Jp over [u − c, rp ]. 
max

Observation (2). ρ is (weakly) increasing.

Subproof. It suffices to consider the case η∗ + c < z∗ and show that ρ is increasing over [η∗ +
c, z∗ ]. Notice for all x ∈ max{rpmin , (η∗ + c)}, z∗ , we have
ρ (x) = jp (x) − v (x) = jp (x) − jp (z∗ ) ≥ 0
where the inequality holds because Jp is concave over this region. Moreover, for all x ∈ [η∗ +
c, rpmin ] (supposing rpmin > η∗ + c), we have
 
Jp η∗ + c − Jp (η∗ )
jp (x) ≥ ≥ jp (z∗ ) = v (x)
c
where the first inequality is due to convexity of Jp over [u − c, rpmin ] and the second inequality
holds because (η∗ ) ≥ 0 (recall that in the current case z∗ > η∗ + c). In sum, ρ = jp − v is
non-negative over [η∗ + c, z∗ ], and thus ρ is increasing.46 

Observation (3). v(x) + ρ(x) ≥ Jp (x) over [u − c, ū − c], where equality holds for x ∈ [u −
c, η∗ ) ∪ [(η∗ + c) ∧ z∗ , z∗ ]. If η∗ > u − c, equality also holds for x = η∗ .

Subproof. The fact that equality holds when x ∈ [u − c, η∗ ) ∪ [(η∗ + c) ∧ z∗ , z∗ ] is obvious by

construction of v and ρ. If η∗ > u − c, we also have v(η∗ ) + ρ(η∗ ) = Jp (η∗ ) because ρ(η∗ ) = 0
and v(η∗ ) = Jp (η∗ ) by Observation (1).
For x > z∗ , by construction we have v(x) + ρ(x) = v(x) + Jp (z∗ ) − v(z∗ ) = jp (z∗ )(x − z∗ ) +
Jp (z∗ ) ≥ Jp (x), where the last inequality is due to concavity of Jp over [rpmin , ū − c].47
 
For x ∈ η∗ , (η∗ + c) ∧ z∗ , ρ(x) = 0 and thus it suffices to show v(x) ≥ Jp (x) over this
interval. v(η∗ ) ≥ J ∗
 p (η ) has been shown in Observation (1). Thus it suffices to consider x ∈
∗ ∗ ∗
η , (η + c) ∧ z . Consider two possibilities48 :

• (η∗ + c) ∧ z∗ ≤ rpmax . In this case, Jp is convex and v is affine over the interval. Thus the
   
desired result is implied by v(η∗ ) ≥ Jp (η∗ ) and v (η∗ + c) ∧ z∗ = Jp (η∗ + c) ∧ z∗ .
• (η∗ + c) ∧ z∗ > rpmax . In this case, for x ∈ [rpmax , (η∗ + c) ∧ z∗ ), it has been shown that v(x) ≥
Jp (x) (see the last paragraph in the proof of Observation (1)). For x ∈ (η∗ , rpmax ), the result
holds because Jp is convex over [η∗ , rpmax ], v(η∗ ) ≥ J (η∗ ) and v(rpmax ) ≥ J (rpmax ). 

Observation (4). Gη∗ , v and ρ satisfy condition (C3) in Theorem 3.1.

 
Subproof. • v(x)dGη∗ (x) = v(x)dFU −c (x) holds because Gη∗ agree with FU −c over
[u − c, η∗ ) (see Lemma B.2) and v is affine over [η∗ , ū − c]. (Recall that Gη∗ MP S FU −c .)

46 If J is strictly convex over [u−c, r min ] and is strictly concave over [r min , ū −c], then the corresponding inequalities
p p p
hold strictly and thus ρ is strictly increasing over [η∗ + c, z∗ ].
47 The last inequality holds strictly if J is strictly concave over [r min , ū − c].
p p
48 In both cases, if J is strictly convex over [u − c, r max ] and is strictly concave over [r max , ū − c], then the same
p
 p p
argument shows v(x) > Jp (x) over η∗ , (η∗ + c) ∧ z∗ . When referring to the proof of Observation (1), also refer to
footnote 45.

35
C. Lyu Journal of Economic Theory 213 (2023) 105722

 
• When η∗ + c ≥ z∗ , ρ(x)dGη∗ (x) = ρ(x)dFU (x) holds trivially since ρ ≡ 0. When η∗ +
c < z∗ , we have:
 
ρ(x)dGη∗ (x) − ρ(x)dFU (x)
 
= ρ(x)dGη∗ (x) − ρ(x)dFU (x)
[z∗ ,ū] [z∗ ,ū]
    
= Jp (z∗ ) − v(z∗ ) 1 − lim∗ Gη∗ (x) − 1 − lim∗ FU (x) = 0
xz xz

where the first equality holds because ρ(x) ≡ 0 for x ≤ η∗ + c and Gη∗ agrees with FU over
(η∗ + c, z∗ ); the second equality is because ρ(x) ≡ Jp (z∗ ) − v(z∗ ) when x ≥ z∗ ; the last
line equals to zero because Gη∗ = FU on a left neighborhood of z∗ when η∗ + c < z∗ (see
Lemma B.2). 

To sum up, we have:

• (v, ρ) satisfies (C1) by Observations (1) and (2);

• Gη∗ supports within [u − c, η∗ ] ∪ [(η∗ + c) ∧ z∗ , z∗ ] when η∗ > u − c and supports within
[(η∗ + c) ∧ z∗ , z∗ ] when η∗ = u − c (see Lemma B.2). Thus (Gη∗ , v, ρ) satisfies (C2) by
Observation (3).
• (Gη∗ , v, ρ) satisfies (C3) by Observation (4).
• G∗η satisfies (C4) since it is induced by a particular feasible signal (upper-censorship signal
with threshold η∗ ).

Thus, Gη∗ is indeed optimal for the Relaxed Problem, and the corresponding upper-censorship
signal is optimal. 

B.4. Proof for Corollary 3.1

Proof. Part (a) is equivalent to Proposition 3.1(a). For part (b), if rpmax = ū − c, then Jp is also
convex and thus affine on [u − c, ū − c]. Then the result is directly implied by Lemma 3.1. If
rpmax < ū − c, it suffices to check that η∗ = u − c satisfies the conditions in Proposition 3.1(b).
Indeed, concavity of Jp implies rpmin = u − c and thus η0 = u − c, so u − c is within [η0 , rpmax ].
Moreover, concavity of Jp directly implies (u − c) ≥ 0 and thus condition (ii) holds. 

B.5. Proof for Proposition 3.2

Part 1: uniqueness of the optimal distribution of U ∧ Z and the optimal upper-censorship signal

Proof. Case 1: rp = ū − c
In this case, jp is strictly increasing over [u − c, ū − c] and G = FU −c is optimal for the
Relaxed Problem. Then, for any optimal distribution G∗ of the effective-search-value, we must
have:
 
∗
0= Jp (x)dG (x) − Jp (x)dFU −c (x)
(u−c,ū−c] (u−c,ū−c]

36
C. Lyu Journal of Economic Theory 213 (2023) 105722

 ū−c
 ū−c
= Jp (x) G∗ (x) − FU −c (x)  − [G∗ (x) − FU −c (x)]jp (x)dx
u−c
u−c

ū−c
=− [G∗ (x) − FU −c (x)]jp (x)dx
u−c

 x  ū−c

=− [G∗ (t) − FU −c (t)]dt jp (x) 
u−c
u−c
  x 
+ [G∗ (t) − FU −c (t)]dt djp (x)
(u−c,ū−c] u−c
  x 
= [G∗ (t) − FU −c (t)]dt djp (x)
(u−c,ū−c] u−c

(For any function h, h(x)|ba := h(b) − h(a).) In the first equality, point u − c is omitted from the
integrals because FU −c is atom-less and because of this, G∗ also puts zero probability at u − c
to be a MPC of FU −c . The 2nd and 4th equalities are by integration by parts.49 The 3rd equality
holds because G∗ MP S FU −c ⇒ supp{G∗ } ⊂ supp{FU −c } and thus G∗ (x) =FU −c (x) for
x = u − c or ū − c. The 5th equality holds because G∗ MP S FU −c implies that u−c [G∗ (t) −
x

FU −c (t)]dt = 0 for x = ū − c.
Now, notice term u−c [G∗ (t) − FU −c (t)]dt ≤ 0 for all x since G∗ MP S FU −c . As jp is
x
x
increasing, djp is a positive measure over (u−c, ū −c] and the above result implies u−c [G∗ (t) −
FU −c (t)]dt = 0 for djp − a.e. x in (u − c, ū − c].
Suppose G∗ (t0 ) > FU −c (t0 ) for some t0 ∈ [u − c, ū − c). Then by right-continuity we
have G∗ (t) > FU −c (t) over some right-neighborhood of t0 , denoted as I . Then  x over inter-
val I , u−c [G∗ (t) − FU −c (t)]dt is strictly increasing in x. This implies that u−c [G∗ (t) −
x

FU −c (t)]dt = 0 over some subinterval I ⊂ I . Then we must have jp constant over I , which
violates the assumption that jp is strictly increasing. Similar argument also rules out the possi-
bility that G∗ (t0 ) < FU −c (t0 ). Thus G∗ (t) = FU −c (t) for all t ∈ [u − c, ū − c). We also have
G∗ (ū − c) = 1 = FU −c (ū − c) since supp{G∗ } ⊂ supp{FU −c }. In conclusion, G∗ = FU −c is the
unique optimal solution.
Case 2: rp < ū − c
Let η∗ be an optimal threshold characterized in Proposition 3.1(b), and let v and ρ be the
corresponding functions defined in Section B.3.3. I have shown that (Gη∗ , v, ρ) satisfies all con-
ditions in Theorem 3.1. Thus, by Theorem 3.1, if another distribution G∗ also solves the Relaxed
Problem (which is now necessary for it to be optimal), then (G∗ , v, ρ) must also satisfy all opti-
mality conditions in Theorem 3.1. Then we have following observations:
 
Observation (1). G∗ has zero probability over η∗ , (η∗ + c) ∧ z∗ and (z∗ , ū − c].


49 Integration by parts holds in the two equalities respectively because J (x) and x [G∗ (t) − F
p u−c U −c (t)]dt are con-
tinuous in x. See Folland (1999) Theorem 3.36.

37
C. Lyu Journal of Economic Theory 213 (2023) 105722

Subproof. By condition (C2) in Theorem 3.1, it suffices to show that v(x) + ρ(x) > Jp (x)
for x ∈ η∗ , (η∗ + c) ∧ z∗ ∪ (z∗ , ū − c]. This is proved by slightly modifying the subproof of
Observation (3) in the proof for Proposition 3.1 in Section B.3.3. See footnotes 47 and 48 for
details. (Notice that rpmin = rpmax = rp now under Assumption 3.2.) 

Observation (2). G∗ (x) = FU −c (x) for x ∈ [u − c, η∗ ].

Subproof. Since G∗ MP S FU −c , we must have G∗ (u − c) = FU −c (u − c) = 0 (by assumption,

FU −c is atomless). Thus it suffices to show the result assuming η∗ > u− c.
By condition (C3) in Theorem 3.1, we must have v(x)dG∗ (x) = v(x)dFU −c (x). By two
steps of integration by parts (similar to the procedure in Case 1), this implies that
  x 
[G∗ (t) − FU −c (t)]dt dv (x) = 0
(u−c,ū−c] u−c
x
and we thus have u−c [G∗ (t) − FU −c (t)]dt = 0 for dv − a.e. x, where v denotes the right-
continuous (almost everywhere) derivative of convex function v, and dv is the positive measure
induced by it. As v(x) = Jp (x) for x ∈ [u − c, η∗ ], v (x) = jp (x) is strictly increasing over [u −
c, η∗ ). Thus by similar argument as in Case 1, we have G∗ (t) = FU −c (t) for all t ∈ [u − c, η∗ ).
Finally, consider the point η∗ . Notice G∗ (t) = FU −c (t) ∀t ∈ [u − c, η∗ ) and FU −c being con-
tinuous at η∗ imply that G∗ (η∗ ) ≥ FU −c (η∗ ) (otherwise G∗ must jump down at η∗ , violating its
monotonicity). Suppose G∗ (η∗ ) > FU −c (η∗ ). Then ∗
 x by∗ right-continuity G (x) > FU −c (x) over
a right-neighborhood of η . This implies that u−c [G (t) − FU −c (t)]dt > 0 for some x > η∗ ,
∗

violating G∗ MP S FU −c . Thus G∗ (η∗ ) = FU −c (η∗ ). 

Observation (3). If z∗ > η∗ + c, then G∗ (x) = FU (x) for x ∈ [η∗ + c, z∗ ).

 
Subproof. Condition (C3) in Theorem 3.1 requires that ρ(x)dG∗ (x) = ρ(x)dFU (x). This
implies that
 
0 = ρ(x)dG∗ (x) − ρ(x)dFU (x)

ū ū
∗ 
= ρ(x)(G (x) − FU (x)) − [G∗ (x) − FU (x)]dρ(x)
u−c
u−c

ū
=− [G∗ (x) − FU (x)]dρ(x)
u−c

where the integration by parts formula for the 2nd equality is valid because ρ is continuous by
construction. Since G∗ (x) ≥ FU (x) by FOSD and ρ is increasing, we must have G∗ (x) = FU (x)
for dρ − a.e. x in [u − c, ū]. By footnote 46, we know ρ is strictly increasing over [η∗ + c, z∗ ].
Thus G∗ (x) = FU (x) for all x ∈ [η∗ + c, z∗ ) (using right-continuity of CDF). 

Together with the requirement that supp{G∗ } ⊂ [u − c, ū − c] (since G∗ MP S FU −c ), the

above observations pin down a unique formula for G∗ , which is the same as Gη∗ . Finally, notice

38
C. Lyu Journal of Economic Theory 213 (2023) 105722

that two upper-censorship signals with different thresholds necessarily lead to different G. Thus
the optimal upper-censorship signal is unique under Assumption 3.2. 

Part 2: uniqueness of the optimal joint distribution of (U, Z)

Proof. I first need a simple observation:

Observation (4). Under any signal, we have: (i) E[U − c − U ∧ Z|Z] = 0 a.s.; (ii) P (U ∧ Z =
Z|Z) > 0 a.s.

Subproof. By definition of Z, we have E[U − c − U ∧ Z|S] = E[(U − Z)+ − c|S] = 0. This im-
plies (i) since Z is measurable w.r.t. S. Suppose (ii) does not hold. Then with positive probability
P (U ∧ Z = U |Z) = 1 and thus E[U − c − U ∧ Z|Z] = −c, which contradicts with (i). 

Given any signal inducing Gη∗ , I show two claims below.

Claim 1: On the event {U − c ≤ η∗ }, Z = U − c almost surely (i.e., P (U − c ≤ η∗ , Z = U − c) =
0).
Notice under any signal inducing Gη∗ , for any t ≤ η∗ we have:
E[(U − c − t)1{U − c ≤ t}]
=E[(U ∧ Z − t)1{U ∧ Z ≤ t}]
 
=E[(U ∧ Z − t) 1{Z ≤ t} + 1{Z > t ≥ U } ]
≥E[(U ∧ Z − t)1{Z ≤ t}] + E[(U − c − t)1{Z > t ≥ U }]
=E[(U − c − t)1{Z ≤ t}] + E[(U − c − t)1{Z > t ≥ U }]
=E[(U − c − t)1{U ∧ Z ≤ t}]
where the first equality holds because Gη∗ (x) = FU −c (x) for x ≤ η∗ ; the inequality in the fourth
row holds because U ∧ Z = U > U − c when Z > U ; the equality in the 5th row holds because
of result (i) in Observation (4) above. This result further implies:
 
0 =E[(U − c − t) 1{U − c ≤ t} − 1{U ∧ Z ≤ t} ]
 
=E[(U − c − t) 1{U − c ≤ t; U ∧ Z > t} ]
 
− E[(U − c − t) 1{U − c > t; U ∧ Z ≤ t} ]
Notice that in the second line, the first term is non-positive and the second term is non-negative.
Thus they must both be zero. This implies that P (U − c < t; U ∧ Z > t) = P (U − c > t; U ∧
Z ≤ t) = 0. Since this holds for arbitrary t ≤ η∗ , we have for any t1 < t2 ≤ η∗ , P (U − c ∈
(t1 , t2 ), U ∧ Z ∈/ [t1 , t2 ]) = 0. Since U is continuous, we have a stronger result that P (U − c ∈
[t1 , t2 ], U ∧ Z ∈
/ [t1 , t2 ]) = 0.
Now, notice that whenever U − c ≤ η∗ and U − c = U ∧ Z, we can always find t1 , t2 ∈ Q
such that t1 < t2 ≤ η∗ , U − c ∈ [t1 , t2 ] and U ∧ Z ∈ / [t1 , t2 ]. Thus

P (U − c ≤ η∗ ; U − c = U ∧ Z) ≤ P (U − c ∈ [t1 , t2 ]; U ∧ Z ∈
/ [t1 , t2 ])
t1 ,t2 ∈Q: t1 <t2 ≤η∗

The conclusion above implies the RHS equals to 0. Thus P (U − c ≤ η∗ ; Z = U − c) = P (U −

c ≤ η∗ ; U − c = U ∧ Z) = 0.

39
C. Lyu Journal of Economic Theory 213 (2023) 105722

Claim 2: On the event {U − c > η∗ }, Z = z(η∗ ) a.s. (i.e., P (U − c > η∗ ; Z = z(η∗ )) = 0.)
First notice that Observation (4)(ii) implies that the support of Z is a subset of the support of
U ∧ Z.50 Thus we have supp{Z} ⊂ [u − c, η∗ ] ∪ [(η∗ + c) ∧ z(η∗ ), z(η∗ )]. Also recall that in
the proof of Claim 1, I have shown P (U − c > t; U ∧ Z ≤ t) = 0, ∀t ≤ η∗ . With t = η∗ , this
implies P (U − c > η∗ ; Z ≤ η∗ ) = 0. So, on the event {U − c > η∗ } we have Z ∈ [(η∗ + c) ∧
z(η∗ ), z(η∗ )] a.s. Therefore, when ∗
 ∗η + c ≥∗z(η
∗
 ), Claim 2 has been proved; when η + c <
∗
∗
z(η ), it suffices to show P Z ∈ η + c, z(η ) = 0.
Now, assume η∗ + c < z(η∗ ). We have:

E[(U ∧ Z)1{U ≥ η∗ + c}] = E[(U ∧ Z)1{U ∧ Z ≥ η∗ + c}]

     
= E U ∧ z(η∗ ) 1{U ∧ z(η∗ ) ≥ η∗ + c} = E U ∧ z(η∗ ) 1{U ≥ η∗ + c}
where the first equality holds because almost surely U ≥ η∗ + c ⇒ Z ≥ η∗ + c as shown above
(the event {U = η∗ + c} has zero probability and thus can be ignored); the second equality holds
because Gη∗ agrees with the distribution of U ∧ z(η∗ ) over [η∗ + c, +∞). Since Z ≤ z(η∗ ) a.s.,
U ∧ Z ≤ U ∧ z(η∗ ) a.s. and thus the above result implies U ∧ Z = U ∧ z(η∗ ) a.s. on the event
{U − c > η∗ }. 
Notice
 that Claim 1 implies almost surely Z > η∗ ⇒ U − c > η∗ . Thus we have Z ∈ η∗ +
c, z(η∗ ) ⇒ U ∧ Z = U ∧ z(η∗ ) a.s. Also notice when
 ∗ Z < z(η
∗ ), U ∧ Z = U ∧ z(η∗ ) only if

∗
U < Z. This further implies that almost surely Z ∈ η  + c,z(η ) ⇒ U <Z. Since Observation
(4)(ii) implies P (Z ≤ U |Z) > 0 a.s., we must have P Z ∈ η∗ + c, z(η∗ ) = 0. This completes
the proof for Claim 2.
The two claims above pin down a unique distribution of Z conditional on U . Thus the optimal
joint distribution of (U, Z) is unique. 

B.6. Proof for Proposition 3.3

Proof. I show the existence of optimal (p ∗ , η∗ ) when J (·) has a log-concave left-tail. The other
part of the proposition is evident from the discussion in the main text. W.l.g., I assume J (ū − c) >
0. If this does not hold, profit is always 0 for any p > 0 and thus every policy is optimal. Notice
that by continuity of J , J (ū − c) > 0 implies that the probability of sale is strictly positive for
some p > 0 under fully revealing signal. Thus the seller’s maximal expected profit is strictly
positive.
Taking the formula of Gη (equation (B.3)) into the optimization in the proposition, we get:

  z(η)
η

max p J (x − p)dFU −c (x) + J (x − p)dFU (x)

p≥0
η∈[u−c,ū−c] u−c (η+c)∧z(η)

    
+ J z(η) − p 1 − FU max{η + c, z(η)}

Because FU is continuous, the integrals are continuous in their limits. Due to the continuity
of z(·) (Lemma B.1(b)) and J (·), the objective function is continuous in (p, η). Thus, by the

50 Otherwise, there exists set B s.t. P (Z ∈ B) > 0 but P (U ∧ Z ∈ B) = 0. This implies E[P (U ∧ Z = Z|Z)|Z ∈ B] =
P (U ∧ Z = Z|Z ∈ B) = P (U ∧ Z = Z; Z ∈ B)/P (Z ∈ B) = 0, which contradicts with Observation (4)(ii).

40
C. Lyu Journal of Economic Theory 213 (2023) 105722

Maximum Theorem, the solution exists if we can show that any sufficiently large p is suboptimal
(so that the feasible set can be shrunk to be compact). If the support of J (·) is bounded from
below, this is obviously true since a sufficiently large p will lead to zero quantity of sales then.
Thus it suffices to consider the case where the support of J (·) is unbounded from below.
Notice that the seller’s profit is always bounded by (p) := pJ (ū − c − p) given any p.
Since J has a log-concave left-tail, there exists p0 s.t. for p > p0 we have log(J (ū − c − p))
being concave in p and thus ∂ log(J (∂p
ū−c−p))
being decreasing in p. Moreover, because J (·) is a
∂ log(J (ū−c−p))
CDF with its support unbounded from below, ∂p must be strictly negative for some
p > p0 . These together imply that there exists > 0 such that ∂ log(J (∂p
ū−c−p))
< − when p is
large enough.
Now, notice that ∂ log((p))
∂p = 1/p + ∂ log(J (∂p
ū−c−p))
. The above result then implies that
∂ log((p))
∂p < − /2 for p large enough. This implies that limp→+∞ log((p)) = −∞ and thus
limp→+∞ (p) = 0. Therefore, any p that is sufficiently large is indeed suboptimal. 

Appendix C. Proofs for Section 4

C.1. Proof for Lemma 4.1

Proof. Notice:
 
    
x − EU ∼φ [U − c] + − c φ(dx) ≥ x − EU ∼φ [U − c] − c φ(dx)
 
= (x − c)φ(dx) − EU ∼φ [U − c] = 0 = [(x − zφ )+ − c]φ(dx)

where the last equality holds by the definition of zφ . This implies that EU ∼φ [U − c] ≤ zφ . More-
over, notice that the inequality above holds as equality if and only if x − EU ∼φ [U − c] ≥ 0 for
φ-a.e. x. Thus EU ∼φ [U − c] = zφ if and only if inf(supp{φ}) ≥ EU ∼φ [U − c]. 

C.2. Proof for Proposition 4.2

The proposition can be proved by applying Theorem 1 in Dworczak and Martini (2019) to the
optimization (4.1) – (4.2). Because it is analogous to (but simpler than) the proofs for the search
goods case, the details are omitted. An alternative proof, which treats the proposition as a special
case of the results for search goods, is also available upon request.

C.3. Proof for Proposition 4.3

The proof requires a sequence of lemmas that are also useful in some other proofs later. I first
show a technical one:

Lemma C.1. A continuously differentiable function ϒ(·) is strictly convex over [a, t] and strictly
concave over [t, b]. Pick any x, y, w, x , y ∈ [a, b] such that x < y ≤ w, x < y , x ≤ x, y ≤ y
and (x , y ) = (x, y). Then we have ϒ(y)−ϒ(x)
y−x ≤ ϒ (w) ⇒ ϒ(yy)−ϒ(x
−x
)
< ϒ(y)−ϒ(x)
y−x .

ϒ(y)−ϒ(x)
Proof. Assume x, y, w, x and y are picked as in the lemma and y−x ≤ ϒ (w). Notice
ϒ(y)−ϒ(x)
by the strict concavity of ϒ over [t, b], y−x ≤ ϒ (w) ⇒ x < t. Moreover, if y ≤ t, then the

41
C. Lyu Journal of Economic Theory 213 (2023) 105722

result is directly implied by the strict convexity of ϒ over [a, t]. Thus we only need to consider
the case where x < t < y.
For any s1 < s2 , let ϒ (s1 , s2 ) denote the average slope of ϒ over [s1 , s2 ], i.e., ϒ (s1 , s2 ) :=
ϒ(s2 )−ϒ(s1 )
s2 −s1 . Then we have ϒ (x, y) ≤ ϒ (w) and want to show ϒ (x , y ) < ϒ (x, y).
First, we can show x < x ⇒ ϒ (x , x) < ϒ (x, y). To see this, notice that by the strict con-
cavity of ϒ over [t, b], we have ϒ (t, y) > ϒ (w). Thus ϒ (x, y) ≤ ϒ (w) ⇒ ϒ (x, t) < ϒ (x, y)
since ϒ (x, y) is a weighted average of ϒ (x, t) and ϒ (t, y). By the convexity of ϒ over [a, t],
we have x < x ⇒ ϒ (x , x) ≤ ϒ (x, t). These together imply x < x ⇒ ϒ (x , x) < ϒ (x, y).
If y ≤ x, then the above result directly imply ϒ (x , y ) < ϒ (x, y) since ϒ (x , y ) ≤
ϒ (x , x) in this case by the convexity of ϒ over [a, t]. Thus it remains to consider the case
where y > x.
When y > x, we can prove y < y ⇒ ϒ (x, y ) < ϒ (x, y). Since ϒ (x, ·) is increasing over
(x, t] by the convexity of ϒ over that region, it suffices to prove this when y > y ≥ t. In this
case, by the strict concavity of ϒ over [t, b], we have ϒ (y , y) > ϒ (w). Thus our assumption
ϒ (x, y) ≤ ϒ (w) implies ϒ (x, y ) < ϒ (x, y) since ϒ (x, y) is a weighted average of ϒ (x, y )
and ϒ (y , y). This concludes the proof for y < y ⇒ ϒ (x, y ) < ϒ (x, y). Together with the
earlier conclusion x < x ⇒ ϒ (x , x) < ϒ (x, y), this further implies ϒ (x , y ) < ϒ (x, y). 

Next, I show an important property of the function .

Lemma C.2. Under Assumption 3.2, (·) single-crosses zero from below over [u − c, ū − c), i.e.,
(η1 ) ≥ (>)0 ⇒ (η2 ) ≥ (>)0 for any η2 > η1 in [u − c, ū − c).

Proof. I prove the contrapositive for this: for any η1 , η2 ∈ [u − c, ū − c) s.t. η2 > η1 , I show
(η2 ) ≤ (<)0 ⇒ (η1 ) ≤ (<)0. Notice that if z(η1 ) < rp , the strict convexity of Jp over [u −
c, rp ] would directly imply (η1 ) < 0. Thus it suffices to assume z(η1 ) ≥ rp .
Notice the definition of  implies
   
Jp (η1 + c) ∧ z(η1 ) − Jp (η1 ) Jp (η2 + c) ∧ z(η2 ) − Jp (η2 )
(η1 ) − (η2 ) = −
(η1 + c) ∧ z(η1 ) (η2 + c) ∧ z(η2 ) − η2
     
=A =B
 
+ jp (z(η2 )) − jp (z(η1 ))

Since z(η1 ) ≥ rp , we have η2 > η1 ⇒ z(η2 ) ≥ z(η1 ) ⇒ jp (z(η2 )) ≤ jp (z(η1 )). Thus, it now
suffices to show that (η2 ) ≤ 0 ⇒ A ≤ B, which then implies (η1 ) ≤ (η2 ). This is directly
implied by Lemma C.1 with ϒ = Jp , a = u − c, b = ū − c, t = rp , x = η2 , y = (η2 + c) ∧ z(η2 ),
w = z(η2 ), x = η1 and y = (η1 + c) ∧ z(η1 ). 

With Lemma C.2, the following lemma is almost immediate, which is key to the proof of
Proposition 4.3(b).

Lemma C.3. Assume Assumption 3.2 holds. Let ηS∗ be the (unique) optimal upper-censorship
signal threshold of the non-pricing seller for search goods. For any x ∈ [u − c, ū − c):

(a) If (x) < 0, then ηS∗ > x.

(b) If (x) > 0, then ηS∗ ≤ x, where the inequality is strict unless x = u − c.

42
C. Lyu Journal of Economic Theory 213 (2023) 105722

(c) If (x) = 0, then ηS∗ = x.

Proof. If rp = ū − c, then Jp is strictly convex over [u − c, ū − c]. Thus η∗ = ū − c and (x) < 0
for all x ∈ [u − c, ū − c). So the results trivially hold. If rp < ū − c, η∗ (as the unique optimal
threshold) must satisfy the condition in Proposition 3.1(b). For result (a), notice (x) < 0 implies
(x ) < 0 for all x ≤ x by Lemma C.2, so η∗ > x. For result (b), notice (x) > 0 implies (x ) >
0 for all x ≥ x by Lemma C.2, so either η∗ < x or η∗ = u − c ≤ x. For result (c), notice that
under Assumption 3.2, (x) = 0 also guarantees x ∈ [η0 , rp ]. Indeed, if x ∈ (rp , ū − c), then the
strict concavity of  over [rp , ū − c] would imply (x) > 0; if x ∈ [u − c, η0 ) and thus z(x) < rp ,
then the strict convexity of  over [u − c, rp ] would imply (x) < 0. Given this observation, the
desired result is directly implied by Proposition 3.1(b). (Notice rpmin = rpmax = rp here under
Assumption 3.2.) 

Now, we are ready to prove Proposition 4.3.

Proof for Proposition 4.3. When c ≥ μ(ηE ∗ ) − η∗ , Lemma 4.1 implies z(η∗ ) = μ(η∗ ) ≤ η∗ +
E E E E
c. Thus under the EG seller’s optimal signal, U ∧ Z ≡ E[U − c|S] and their common distribution
solves the EG seller’s problem (4.1) – (4.2). Since the SG seller’s problem (3.2) – (3.3) is more
constrained, the same distribution also solves it. Thus the same signal is also optimal for the SG
seller and we have ηS∗ = ηE∗ . Moreover, because z(η∗ ) = μ(η∗ ), the consumer’s search decision
E E
will be the same in equilibrium regardless of the product’s type; because z(ηE ∗ ) ≤ η∗ + c, we
E
∗ ∗
have U0 < z(ηE ) ⇒ U0 < ηE + c and thus no one will search without purchase even when the
product is a search good. Thus the equilibrium outcomes of the two types of goods are the same.
This proves part (a).
For part (b), notice that under Assumption 4.1 we have E (ηE ∗ ) = 0 ⇒ η∗ ∈ (u − c, ū − c).
E
The results are then directly implied by Lemma C.3. 

C.4. Proof for Properties of Fig. 4.2

When Assumption 4.1 holds and ηS∗ < ηE ∗ , I prove the properties of Fig. 4.2. I first show a

lemma that is also useful somewhere else:

∗ ) > η∗ + c. Then we have z(η∗ ) > μ(η∗ ).

Lemma C.4. Assume Assumption 4.1 holds and μ(ηE E S E

Proof. Suppose z(ηS∗ ) ≤ μ(ηE ∗ ). Notice by Lemma 4.1, we have μ(η∗ ) > η∗ + c ⇒ z(η∗ ) >
E E E
∗ ). Thus z(η∗ ) ≤ μ(η∗ ) implies η∗ < η∗ .
μ(ηE S E S E
∗ ))−J (η∗ )
Jp (μ(ηE
∗ ) = 0, i.e.,
Notice Proposition 4.2 implies E (ηE p E
= jp (μ(ηE∗ )). Invoking
∗ )−η∗
μ(ηE
 ∗
E 
∗ ) , w = μ(η∗ ),
Lemma C.1 with ϒ = Jp , [a, b] = [u − c, ū − c], t = rp , (x, y) = ηE , μ(ηE E
 ∗ ∗  ∗ ∗
J (η +c)∧z(η ) −J (η ) ∗ J (μ(η∗ ))−J (η∗ )
(x , y ) = ηS , (ηS + c) ∧ z(ηS∗ ) , we then have p (ηS∗ +c)∧z(ηS∗ )−ηp∗ S < p μ(ηE∗ )−ηp∗ E =
S S S  E  E
∗ )). Notice by Proposition 3.1(b), we must have z(η∗ ) ≥ r and p S J (η∗ +c)∧z(η∗ ) −J (η∗ )
jp (μ(ηE S p
S p S
(ηS∗ +c)∧z(ηS∗ )−ηS∗ ≥
∗ ∗ ∗ ∗ ∗
jp (z(ηS )) (i.e., (ηS ) ≥ 0). These together imply jp (μ(ηE )) > jp (z(ηS )) but μ(ηE ) ≥ z(ηS ) ≥∗

rp , which contradicts with the strict concavity of Jp over [rp , ū − c]. Thus we must have
z(ηS∗ ) > μ(ηE ∗ ). 

43
C. Lyu Journal of Economic Theory 213 (2023) 105722

Notice when Assumption 4.1 holds and ηS∗ < ηE ∗ , Proposition 4.3 implies that we must have

μ(ηE∗ ) > η∗ + c. The above lemma then implies z(η∗ ) > μ(η∗ ). Thus we have z(η∗ ) > μ(η∗ ) >
E S E S E
∗ + c as is indicated in the figure.
ηE
Now, I turn to show the monotonicity properties of red curve, which represents the consumer’s
surplus change given U0 + p when the product changes from an EG to an SG. For this, we need
a way to compute the consumer’s surplus. The following lemma serves this role:

Lemma C.5. Under any pre-search signal, let G and H respectively denote the distributions
of U ∧ Z and E[U − c|S]. Then given any outside option value u0 and product price p,
the consumer’s expected utility is EX∼G [max{X, u0 + p}] − p if the product is an SG and is
EX∼H [max{X, u0 + p}] − p if the product is an EG.

Proof. When the product is an EG, the consumer will search and buy the product if and only  if
E[U
 − c|S] − p ≥ u 0 . Thus her
 expected utility is obviously E max{E[U − c|S] − p, u0 =
}
E max{E[U − c|S], u0 + p} − p. When the product is an SG, several recent papers (Kleinberg
et al., 2016; Armstrong, 2017; Choi et al., 2018) have shown that the consumer surplus can be
computed in a similar way with E[U − c|S] replaced by U ∧ Z. In particular, Corollary 1 in Choi
et al. (2018) implies the desired result. 

Recall that Gη is defined as the distribution of U ∧ Z under the upper-censorship signal with
threshold η; similarly, I define Hη as the distribution of E[U − c|S] under such a signal. Then
Lemma C.5 implies that given U0 = u0 , the consumer’s surplus change when the product changes
from an EG to an SG will be:

max{u0 + p, x}[GηS∗ (dx) − HηE∗ (dx)]

= max{u0 + p − x, 0}[GηS∗ (dx) − HηE∗ (dx)]
u0 +p u0 +p

= (u0 + p − x)[GηS∗ (dx) − HηE∗ (dx)] = (GηS∗ (x) − HηE∗ (x))dx (C.1)
u−c u−c

where the first equality holds because x[GηS∗ (dx) − HηE∗ (dx)] = 0 by the mean-preserving
property; the second equality holds even if u0 + p < u − c because both GηS∗ and HηE∗ support
within [u − c, ū − c]; the third equality holds by integration-by-parts. The derivative of the last
expression above w.r.t. u0 + p is GηS∗ (u0 + p) − HηE∗ (u0 + p).
Recall that the formula of Gη is provided in Lemma B.2. The formula of Hη is easily seen to
be:
⎧
⎪
⎨FU −c (x) x<η
Hη (x) = FU −c (η) η ≤ x < μ(η) (C.2)
⎪
⎩
1 μ(η) ≤ x
When ηS∗ < ηE ∗ < η∗ + c < μ(η∗ ) < z(η∗ ), the relative patterns of G ∗ and H ∗ are illustrated
E E S ηS ηS
in Fig. C.1. In particular, we can see GηS∗ (x) − HηE∗ (x) is strictly negative over (ηS∗ , ηE
∗ + c) ∪
∗ ∗ ∗ ∗
(μ(ηE ), z(ηS )), is strictly positive over (ηE + c, μ(ηE )), and is zero elsewhere. Thus we have
the monotonicity of the surplus change as a function of u0 + p in each of these intervals as is
depicted in Fig. 4.2.

44
C. Lyu Journal of Economic Theory 213 (2023) 105722

Fig. C.1. Comparison between Gη∗ and Hη∗ . The black curve represents Hη∗ and the red curve represents Gη∗ . For
S E E S
clarity, I have picked FU to be a uniform distribution.

C.5. Proof for Proposition 4.4

Given any (FU , c, rp ) satisfying Assumption 4.2 and an arbitrary κ > 0, I define δ := μ(rp ) −
 η+c+δ
rp − c and A := minη∈[u−c,rp ] η+c (η + c + δ − x)fU (x)dx, where the minimum is achievable
by the Weierstrass Theorem. By condition (2) in Assumption 4.2, we have δ > 0. Since fU −c has
full support on [u − c, ū − c], this further implies A > 0. Also define Mf := supx∈[u,ū] fU (x). It
is finite since fU is log-concave by Assumption 4.2.
Now, I choose ν to be the largest strictly positive number satisfying:
κc  1 
A − [1 − FU (rp )]ν − Mf ν 2 − [1 − FU (rp )]ν ≥ 0 (C.3)
2(ū − u) 2
Notice that the LHS is continuous and decreasing in ν, and is strictly positive when ν = 0. Thus
such a ν exists.
Now, to prove Proposition 4.4, it suffices to show the following result:

Proposition C.1. Given (Fu , c, rp ) satisfying Assumption 4.2 and κ > 0, let ν be chosen as
above. Assume Jp satisfies Assumption 3.2 with its mode being rp and (i) j (rp + ν) < j (η0 ); (ii)
J (ū−c)−Jp (u−c)
j (η0 ) > κ p ū−u . Then we have cs S < cs E .

Proof. I first note that under the current assumptions, Assumption 4.1 holds.51 By Propo-
sition 4.2, this implies that the equilibrium outcome with experience goods is unique and
E (ηE ∗ ) = 0. Due to the strict convexity-concavity of J , E (η∗ ) = 0 further implies η∗ < r
p E E p
and μ(ηE ∗)>r .
p
Moreover, under the strictly unimodality of Jp and condition (3) in Assumption 4.2, Propo-
sitions 3.1 and 3.2 imply that the equilibrium outcome with search goods is also unique with
ηS∗ ∈ [η0 , rp ] and (ηS∗ ) = 0. To see this, notice condition (3) in Assumption 4.1 states that
rp < ū − c and z(u − c) < rp . The former of these implies that ηS∗ must satisfy the conditions
in Proposition 3.1(b); the latter of these, together with the continuity of z(·) (see Lemma B.1),
implies η0 > u − c and thus rules out the possibility of ηS∗ = u − c.
I now show the following properties regarding ηE ∗ and η∗ :
S

51 The only non-trivial part for checking this is to show E (u − c) < 0. For this, notice that z(u − c) < r by condition
p
(3) in Assumption 4.1. Since Lemma 4.1 implies μ(u − c) ≤ z(u − c), we then have μ(u − c) < rp . The strict convexity
of Jp over [u − c, rp ] then implies E (u − c) < 0.

45
C. Lyu Journal of Economic Theory 213 (2023) 105722

Observation (1). We have η0 ≤ ηS∗ < ηE

∗ < η∗ + c < μ(η∗ ) < z(η∗ ) ≤ μ(η∗ ) + ν and F (η∗ +
∗
 E  E S E U E
c) − FU (ηS + c) ≥ A − [1 − FU (rp )]ν − 12 Mf ν 2 c.

Subproof. The proof consists of several parts.

• Part 1: μ(ηE∗ ) − (η∗ + c) > δ.
E
By definition, δ = μ(rp ) − (rp + c), which is > 0 by Assumption 4.2. Since ηE ∗ < r as
p
mentioned earlier, it suffices to show μ(η) − η = E[U − c − η|U − c − η ≥ 0] is decreasing
 in
η. Notice the log-concavity of fU implies that fU (x + (c + η)) is log-supermodular in x, −(c +

η) .52 Thus fU −c−η (x) = fU (x + (c + η)) decreases in the likelihood-ratio order when c + η
increases. This implies that we indeed have E[U − c − η|U − c − η ≥ 0] decreasing in η.53
• Part 2: z(ηS∗ ) > ηS∗ + c.
By Proposition 3.1, we have ηS∗ ≤ rp . The similar proof as in Part 1 then implies μ(ηS∗ ) −
(ηS∗ + c) > δ. Since z(ηS∗ ) ≥ μ(ηS∗ ) by Lemma 4.1, we have z(ηS∗ ) > ηS∗ + c.
• Part 3: 0 < z(ηS∗ ) − μ(ηE ∗ ) < ν.

Given the conclusion in Part 1 and that Assumption 4.1 holds here as I mentioned earlier, the
result z(ηS∗ ) − μ(ηE∗ ) > 0 is directly implied by Lemma C.4.

Now, suppose z(ηS∗ ) − μ(ηE ∗ ) ≥ ν. Since E (η∗ ) = 0 ⇒ μ(η∗ ) > r due to the strict con-
E E p
vexity of Jp over [u − c, rp ], we will then have z(ηS∗ ) > rp + ν. The condition ∗ j (rp + ν) <
j (η0 ) together
 with
 the strict quasi-concavity of j (·) then implies that j z(ηS ) < j (x) for
all x ∈ η0 , z(ηS∗ ) . Since ηS∗ ≥ η0 as is mentioned earlier, we then must have (ηS∗ ) =
Jp ((ηS∗ +c)∧z(ηS∗ ))−Jp (ηS∗ )
− j (z(ηS∗ )) > 0, which contradicts with my earlier conclusion of (ηS∗ ) =
(ηS∗ +c)∧z(ηS∗ )−ηS∗
0. Therefore, we must have z(ηS∗ ) − μ(ηE ∗ ) < ν.

∗ + c) − F (η∗ + c) ≥ A − [1 − F (r )]ν − 1 M ν 2 c.

• Part 4: FU (ηE U S U p 2 f
By the definitions of z(·) and μ(·), we have:
ū ū
  
x − z(ηS∗ ) + − c Fu (dx) = 0; ∗
[x − μ(ηE ) − c]Fu (dx) = 0
ηS∗ +c ∗ +c
ηE

Taking difference of these two equations, we get:

∗ +c
ηE ū ū
 
cFU (dx) = x − z(ηS∗ ) + Fu (dx) − ∗
[x − μ(ηE )]Fu (dx) (C.4)
ηS∗ +c ηS∗ +c ∗ +c
ηE

∗ + c) − F (η∗ + c)]. Notice ū

  ∗

The LHS above equals to c[FU (ηE U S ηS∗ +c x − z(ηS ) + Fu (dx) =
 ū ∗ ∗ ∗
z(η∗ ) [x − z(ηS )]Fu (dx) because z(ηS ) > ηS + c as shown in Part 2. Thus the RHS equals to
S

z(ηS∗ ) ∗
ū   E)
μ(η
∗
[μ(ηE ) − z(ηS∗ )]FU (dx) + ∗
[μ(ηE ) − x]FU (dx) + ∗
[μ(ηE ) − x]FU (dx)
z(ηS∗ ) ∗)
μ(ηE ∗ +c
ηE

52 See, for example, Lemma 2.6.2(b) in Topkis (1998). Apply it to log f (·) .
 
U
53 See, e.g., Theorem 1.4.6 in Müller and Stoyan (2002).

46
C. Lyu Journal of Economic Theory 213 (2023) 105722

Because z(ηS∗ ) ≥ rp and z(ηS∗ ) − μ(ηE∗ ) < ν, the first term above is greater than −[1 − F (r )]ν.
U p
 z(ηS∗ )  ∗
∗ ) − x]F (dx) ≥ z(ηS ) [μ(η∗ ) − x]M dx ≥
For the second term, we have μ(η∗ ) [μ(ηE U ∗)
μ(ηE E f
 μ(ηE∗ )+v E

μ(ηE∗) [μ(ηE ) − x]Mf dx = −Mf ν /2, where the second inequality is again due to z(ηS∗ ) −
∗ 2
 ∗  ∗
∗ ) < ν. For the third term, we have μ(ηE ) [μ(η∗ ) − x]F (dx) ≥ ηE +c+δ [η∗ + c + δ −
μ(ηE ∗ +c
ηE E U ∗ +c
ηE E
∗ ∗
x]FU (dx) ≥ A, where the first inequality is because μ(ηE ) − (ηE + c) > δ as proved in Part 1,
and the second inequality is due to the definition of A and that ηE ∗ ∈ [u − c, r ]. These together
  p
imply that the RHS of equation (C.4) is greater than A − 1 − FU (rp ) ν − 12 Mf ν 2 . The desired
inequality is thus implied by equation (C.4).
∗ > η∗ ≥ η .
• Part 5: ηE S 0
 
Since ν > 0 satisfies condition (C.3), we must have A − 1 − FU (rp ) ν − 12 Mf ν 2 > 0. Thus
∗ > η∗ is directly implied by the result of Part 4. Moreover, η∗ ≥ η since η∗ ∈ [η , r ], as I
ηE S S 0 S 0 p
have mentioned in the main proof. 

Recall that I have defined Gη and Hη as the distributions of U ∧ Z and E[U − c|S] under
the upper-censorship signal with threshold η respectively (see expressions (B.3) and (C.2)). In
Appendix C.4, I have shown that given U0 = u0 , a consumer’s surplus change when the product
u0 +p
changes from an EG to an SG will be u−c (GηS∗ (x) − HηE∗ (x))dx, where the patterns of GηS∗
and HηE∗ are illustrated in Fig. C.1 when ηS∗ < ηE ∗ < η∗ + c < μ(η∗ ) < z(η∗ ). This implies
  y
E E S
cs S − cs E = [ u−c (GηS∗ (x) − HηE∗ (x))dx]jp (y)dy. To show this is negative, I prove two more
observations below.
 ηE∗ +c  y 
Observation (2). η∗ [ u−c (HηE∗ (x) − GηS∗ (x))dx]jp (y)dy ≥ 12 jp (η0 )c A − [1 − FU (rp )]ν −
1 2
 E

2 Mf ν .

 ηE∗ +c ∗ + c) − F (η∗ + c)]c.

Subproof. First, one can show u−c (HηE∗ (x) − GηS∗ (x))dx = [FU (ηE U S
To see this, notice that the integral represents the space of the gray area in Fig. C.1. The area’s
height is FU (ηE ∗ + c) − F (η∗ + c) at any x, and its width is c at any y. Thus the integral equals
U S
to [FU (ηE + c) − FU (ηS∗ + c)]c.54
∗

Next, notice that over the region [ηE ∗ , η∗ + c], we have H ∗ (x) − G ∗ (x) ≤ F (η∗ + c) −
E ηE ηS U E
 ∗
∗ , η∗ + c], we have ηE +c (H ∗ (x) − G ∗ (x))dx ≤
FU (ηS∗ + c). This implies that for any y ∈ [ηE E y ηE ηS
∗ ∗ ∗
 ηE∗ +c
[FU (ηE + c) − FU (ηS + c)](ηE + c − y). Together with the conclusion u−c (HηE∗ (x) −
GηS∗ (x))dx = [FU (ηE ∗ + c) − F (η∗ + c)]c above, this implies that for any y ∈ [η∗ , η∗ +
U S
y ∗ + c) − F (η∗ + c)](y − η∗ ). This further im-
E E
c], u−c (HηE∗ (x) − GηS∗ (x))dx ≥ [FU (ηE U S E
 η∗ +c  y  η∗ +c ∗ + c) − F (η∗ + c)](y −
plies η∗E [ u−c (HηE∗ (x) − GηS∗ (x))dx]jp (y)dy ≥ η∗E [FU (ηE U S
E E
∗
ηE )jp (y)dy.

Now, it suffices to show the RHS of the last inequality is greater than 12 jp (η0 )c A −
 ∗ < r , we have j (η∗ ) > j (η ). Moreover, be-
[1 − FU (rp )]ν − 12 Mf ν 2 . Since η0 < ηE p p E p 0

54 Notice this argument does not require the left and right sides of the area to be linear. We only need them to be parallel,
which is true because Hη∗ follows FU −c over [ηS∗ , ηE ∗ ] and G ∗ follows F over [η∗ + c, η∗ + c].
η U S E
E S

47
C. Lyu Journal of Economic Theory 213 (2023) 105722

cause E (ηE ∗ ) = 0, we must have j (μ(η∗ )) ≥ j (η∗ ). By the strict quasi-concavity of j ,

p E p E p
 ∗
∗ , μ∗ (η∗ )]. Thus we have ηE +c [F (η∗ +
these together imply jp (y) > jp (η0 ) for all y ∈ [ηE E η∗ U E
∗ ∗ ∗ ∗
 ηE∗ +cE ∗ )dy =
c) − FU (ηS + c)](y − ηE )jp (y)dy ≥ [FU (ηE + c) − FU (ηS + c)]jp (η0 ) η∗ (y − ηE
∗ ∗
 E 
2 jp (η0 )[FU (ηE + c) − FU (ηS + c)]c ≥ 2 jp (η0 )c A − [1 − FU (rp )]ν − 2 Mf ν , where the
1 2 1 1 2

last inequality is implied by Observation (1). 

y
Observation (3). u−c (GηS (x) − HηE (x))dx ≤ [1 − FU (rp )]ν for any y ∈ [u − c, ū − c].
∗ ∗

y
Subproof. As is illustrated in Fig. 4.2, u−c (GηS∗ (x) − HηE∗ (x))dx achieves its maximum at
 ∗
∗ ). It thus suffices to show μ(ηE ) (G ∗ (x) − H ∗ (x))dx ≤ [1 − F (r )]ν. By the mean-
y = μ(ηE u−c ηS ηE U p
 z(ηS∗ )
preserving condition, we have u−c (GηS (x) − HηE (x))dx = 0 (notice GηS∗ (x) = HηE∗ (x) = 1
∗ ∗

for x > z(ηS∗ )). Since μ(ηE ∗ ) > r , we have G ∗ (x) − H ∗ (x) > F (r ) − 1 for all x ≥ μ(η∗ ).
p ηS ηE U p E
∗ ∗
 z(ηS∗ )
Since z(ηS ) − μ(ηE ) ≤ ν by Observation (1), this implies μ(η∗ ) (GηS∗ (x) − HηE∗ (x))dx ≥
 z(η∗ ) E

(FU (rp ) − 1)ν. Together with the condition u−cS (GηS∗ (x) − HηE∗ (x))dx = 0, this further im-
y
plies u−c (GηS∗ (x) − HηE∗ (x))dx ≤ [1 − FU (rp )]ν. 

Jp (ū−c)−Jp (u−c)
Observation (2), together with the condition jp (η0 ) > κ ū−u , imply

∗ +c
ηE y
[ (HηE∗ (x) − GηS∗ (x))dx]jp (y)dy
∗
ηE u−c

1 Jp (ū − c) − Jp (u − c)  1 
> κ c A − [1 − FU (rp )]ν − Mf ν 2 .
2 ū − u 2

Observation (3) implies

 y
[ (GηS∗ (x) − HηE∗ (x))dx]jp (y)dy
∗ ,η∗ +c)
[u−c,ū−c]\(ηE u−c
E

≤ [1 − FU (rp )]ν[Jp (ū − c) − Jp (u − c)].

 y
Combining these inequalities, we know [ u−c (GηS∗ (x) − HηE∗ (x))dx]jp (y)dy is less than

 κc  1 
[1 − FU (rp )]ν − A − [1 − FU (rp )]ν − Mf ν 2 [Jp (ū − c) − Jp (u − c)]
2(ū − u) 2

which is strictly negative as ν satisfies condition (C.3) by construction. Thus we indeed have
cs S − cs E < 0. 

48
C. Lyu Journal of Economic Theory 213 (2023) 105722

Appendix D. Proofs for Section 5

D.1. Proof for Proposition 5.1

Proof. Assume Jp22 dominates Jp11 in likelihood ratio order (abbr. Jp11 LRD Jp22 ). This means
that for any x < y, we have jp11 (x)jp22 (y) ≥ jp11 (y)jp22 (x). Let rp11 and rp22 denote the modes of
jp11 and jp22 over [u − c, ū − c] respectively. Let (η01 , 1 ) and (η02 , 2 ) be the pair of (η0 , )
defined in Section 3.3 given distributions Jp11 and Jp22 respectively. Notice Jp11 LRD Jp22 implies
rp11 ≤ rp22 . By the definition of η0 , this further implies η01 ≤ η02 because z(·) is increasing.
Let η1∗ and η2∗ denote the thresholds of the optimal upper-censorship signals given Jp11 and Jp22
respectively. Under Assumption 3.2, they are fully characterized by conditions in Proposition 3.1
(given Jp11 and Jp22 respectively). Consider following cases:
Case 1: rp11 = ū − c.
In this case, rp22 also equals to ū − c since rp22 ≥ rp11 . Thus η2∗ = ū − c ≥ η1∗ .
Case 2: rp11 < ū − c.
According to Proposition 3.1(b), in this case we must have either (i) 1 (η1∗ ) = 0 or (ii)
1 (η1∗ ) ≥ 0 and η1∗ = u − c. If (ii) holds, then η1∗ ≤ η2∗ trivially. Thus I assume w.l.g. that (i)
holds. I then have the following observation:

Observation. 1 (η1∗ ) = 0 ⇒ 2 (η1∗ ) ≤ 0.

Subproof. Notice if jp11 (z(η1∗ )) and jp22 (z(η1∗ )) are non-zero, we have:
(η1∗ +c)∧z(η

∗
1)
 1 
1 (η1∗ ) = 0 ⇔ jp1 (t) − jp11 (z(η1∗ )) dt = 0
η1∗
(η1∗ +c)∧z(η1∗ )

 jp11 (t) 
⇔ − 1 dt = 0
jp11 (z(η1∗ ))
η1∗
(η1∗ +c)∧z(η1∗ )

 jp22 (t) 
⇒ − 1 dt ≤ 0
jp22 (z(η1∗ ))
η1∗
(η1∗ +c)∧z(η1∗ )

  2  
⇔ jp2 (t) − jp22 (z(η1∗ )) dt ≤ 0 ⇔ 2 (η1∗ ) ≤ 0
η1∗

jp1 (t) jp2 (t)

where the third row uses Jp11 LRD Jp22 ⇒ 1
jp (z(η1∗ ))
1 ≥ 2
jp (z(η1∗ ))
2 for t ≤ z(η1∗ ).
1 2
Now, it suffices to consider either jp11 (z(η1∗ )) or jp22 (z(η1∗ )) being zero. Let I1 and I2 denote
the supports of j1 and j2 within [u − c, ū − c] respectively. Notice the unimodality assumption
implies that these supports are intervals.

49
C. Lyu Journal of Economic Theory 213 (2023) 105722

• Suppose jp11 (z(η1∗ )) = 0. Then 1 (η1∗ ) = 0 implies that jp11 ≡ 0 on [η1∗ , (η1∗ + c) ∧ z(η1∗ )].
Notice that since η1∗ ≤ rp11 , this interval must be to the left of I1 . Since Jp11 LRD Jp22 implies
inf I1 ≤ inf I2 , we must have jp22 ≡ 0 on [η1∗ , (η1∗ + c) ∧ z(η1∗ )] too. This further implies
2 (η1∗ ) ≤ 0.
• Suppose jp22 (z(η1∗ )) = 0. If z(η1∗ ) is to the left of I2 , then jp22 (x) = 0 for all x ≤ z(η1∗ )
and thus 2 (η1∗ ) = 0. If z(η1∗ ) is to the right of I2 , then because Jp11 LRD Jp22 implies
sup I1 ≤ sup I2 , we must also have jp11 (z(η1∗ )) = 0. This implies 2 (η1∗ ) ≤ 0 as has been
shown above. 

By the observation, we have 2 (η1∗ ) ≤ 0. Lemma C.3 in Appendix C.3 then implies η2∗ ≥
η1∗ . 

D.2. Proof for Proposition 5.2

Proof. Let F and f denote the CDF and PDF of respectively. For any p, let JpW be the
conditional distribution of U0 + p conditioning on W . Then, JpW (x) = F (x − (p + W )). Notice
the log-concavity of f implies that f (x − (p + W )) is log-supermodular in (x, p + W ),55 and
thus JpW (x) = F (x − (p + W )) increases in the likelihood-ratio order when p + W increases.
Also notice f being strictly log-concave implies that JpW satisfies Assumption 3.2. The optimal
η∗ is thus unique and all optimal signals are outcome-equivalent. The desired result for a non-
pricing seller is directly implied by Proposition 5.1.
For a pricing seller, I prove the result with two steps:

• Step 1. Given any p and realized W , the seller’s problem is just the non-pricing seller’s
optimization (3.2) – (3.3) with Jp (x) replaced by JpW (x) = F (x − (p + W )). Notice this
optimization depends on (p, W ) only through  := p + W . Given any , let Q∗ () denote
the optimal value and let η∗ () denote the threshold of the optimal upper-censorship signal.
Then, the argument above for a non-pricing seller implies η∗ () is a singleton and increases
in . Moreover, Q∗ () decreases in  since the objective function decreases in .
• Step 2. Given Q∗ (·), the optimization over p can be written as:

max{pQ∗ (p + W )} ⇐⇒ max{( − W )Q∗ ()}

p 

Since Q∗ () is decreasing in , the objective function has increasing differences in (, W ).
Thus the optimal  increases (in the strong set order) in W . The desired result is then implied
by η∗ () being increasing in .56 

D.3. Proof for Corollary 5.1

Proof. Without discrimination based on W , the setting is the same as in Section 3. Since log-
concavity is preserved under convolution, condition (A1) implies that U0 (and thus U0 + p)
admits a log-concave density. The strict log-concavity of f and condition (A2) further imply


55 See, for example, Topkis (1998) Lemma 2.6.2(b). Apply it to log f (·) .

56 Notice that the log-concavity of f implies log-concavity of F (· − W ) given any realized W . Thus the existence of
optimal solution is guaranteed by Proposition 3.3.

50
C. Lyu Journal of Economic Theory 213 (2023) 105722

that U0 has full support over R and has a single mode. Together with the log-concavity, these
imply that the density of U0 + p satisfies Assumption 3.2. Thus, without discrimination, there
is a unique optimal upper-censorship signal for the seller and all optimal signals are outcome
∗ denote the threshold of this optimal upper-censorship signal.
equivalent. Let ηnd
With discrimination, as is shown in the proof of Proposition 5.2, for any W there is a unique
optimal upper-censorship signal and all optimal signals are outcome-equivalent. Let η∗(W ) de-
note the threshold of the optimal upper-censorship signal.
By Proposition 5.2, η∗ (W ) is (weakly) increasing in W . Let w ∗ = inf{W : η∗ (W ) ≥ ηnd ∗ }.
∗ ∗ ∗ ∗ ∗ ∗
Then we have W > w ⇒ η (W ) ≥ ηnd and W < w ⇒ η (W ) < ηnd . This implies that
for W > w ∗ , the pre-search signal is more informative with discrimination; for W < w ∗ , the
pre-search signal is less informative with discrimination. Thus we have the conclusion in the
corollary. 

D.4. Proof for Proposition 5.3

Proof. Given any search cost c, let z(η; c) equal to the z(η) defined in Section 3.3; let (η; c)
equal to the (η) defined in equation (3.9); let rp (c) denote the maximum point of jp over
[u − c, ū − c]. Let t ∗ ∈ [−∞, ū] be the maximum point of jp over (−∞, ū]. Notice that rp (c) =
t ∗ if and only if t ∗ ∈ [u − c, ū − c].
Pick any search costs c1 and c2 . Let η1 and η2 denote the thresholds of optimal upper-
censorship signals given these two search costs respectively, which are unique under condition
(A1) (by Proposition 3.2) and are characterized by Proposition 3.1. Let z1 := z(η1 ; c1 ) and
z2 := z(η2 ; c2 ). We have the following observation:

Observation (1). If η1 < η2 and η1 + c1 < η2 + c2 , then z1 ≥ z2 .

Subproof. Since u ≤ ηi + ci ≤ ū (i = 1, 2), condition η1 + c1 < η2 + c2 further implies u <

η2 + c2 and η1 + c1 < ū. Consider two cases:

• Case 1: η2 + c2 = ū.
In this case, full disclosure is optimal under search cost c2 (since η2 = ū − c2 ) but is not
optimal under search cost c1 (since η1 < ū − c1 ). Thus Jp is convex over [u − c2 , ū − c2 ]
but is not convex over [u − c1 , ū − c1 ]. This implies that ū − c2 ≤ t ∗ < ū − c1 .
Now, suppose z1 < z2 . Since z2 ≤ ū − c2 (see Lemma B.1(a)), we have:

u − c1 ≤ z1 < z2 ≤ ū − c2 ≤ t ∗ < ū − c1

Notice u − c1 < t ∗ < ū − c1 implies rp (c1 ) = t ∗ < ū − c1 and thus η1 is characterized by the
condition in Proposition 3.1(b). Then, by Lemma B.3, we must have z1 ≥ rp (c1 ) = t ∗ . This
contradicts with the fact z1 < t ∗ derived above. Thus z1 ≥ z2 .
• Case 2: η2 + c2 < ū.
In this case, full disclosure is suboptimal given both search costs c1 and c2 , and thus Jp
is not convex over either [u − c1 , ū − c1 ] or [u − c2 , ū − c2 ]. This implies that the optimal
signals are characterized by the condition in Proposition 3.1(b) under both search costs. Thus
(ηi ; ci ) ≥ 0 for i = 1, 2. Moreover, since η2 + c2 > u as is shown earlier, we further have
(η2 ; c2 ) = 0.

51
C. Lyu Journal of Economic Theory 213 (2023) 105722

Then, it can be shown that:

   
Jp (η1 + c1 ) ∧ z1 − Jp (η1 ) Jp (η2 + c2 ) ∧ z2 − Jp (η2 )
jp (z1 ) ≤ ≤ = jp (z2 )
(η1 + c1 ) ∧ z1 − η1 (η2 + c2 ) ∧ z2 − η2
The first inequality is equivalent to (η1 ; c1 ) ≥ 0. The last equality is equivalent to
(η2 ; c2 ) = 0. Given the last equality, the second inequality is implied by Lemma C.1 in Ap-
pendix C.3 with ϒ = Jp , a = min{u − c1 , u − c2 }, b = ū, t = t ∗ , x = η2 , y = (η2 + c2 ) ∧ z2 ,
w = z2 , x = η1 and y = (η1 + c1 ) ∧ z1 .
By Lemma B.3, we know that z1 ≥ rp (c1 ) and z2 ≥ rp (c2 ). As is mentioned above, Jp is
not convex over either [u − c1 , ū − c1 ] or [u − c2 , ū − c2 ]. These together imply that z1 and
z2 lie in the region where Jp is strictly concave and jp is strictly decreasing. Thus the fact
jp (z1 ) ≤ jp (z2 ) derived above implies z1 ≥ z2 . 

Now, I show part (a) and part (b) of the proposition in sequence.
Part (a): c1 < c2 ⇒ η1 + c1 ≤ η2 + c2 .57
Suppose c1 < c2 but η1 + c1 > η2 + c2 . Then we must have η1 > η2 . According to Observation
(1) above, conditions η1 > η2 and η1 + c1 > η2 + c2 imply z1 ≤ z2 . (When using the observation,
interchange the positions of (c1 , η1 , z1 ) and (c2 , η2 , z2 ).)
However, because η1 + c1 > η2 + c2 , the posterior belief on U after learning U − c1 ≥ η1 is
superior to that after learning U −c2 ≥ η2 (in terms of first-order stochastic dominance). Together
with the assumption that c1 < c2 , it is easy to see that z1 > z2 . This contradicts with the result
z1 ≤ z2 above. Thus c1 < c2 must imply η1 + c1 ≤ η2 + c2 .
Part (b): Given any realized U0 , the consumer surplus is decreasing in c.
Assume c1 < c2 . I first show z1 ≥ z2 by considering the following two cases.

• Case 1: η1 < η2 .
In this case, c1 < c2 further implies η1 + c1 < η2 + c2 . Then by Observation (1) above, we
have z1 ≥ z2 .
• Case 2: η1 ≥ η2 .
Suppose z1 < z2 . We have:
0 =E[(U − z1 )+ − c1 |U − c1 ≥ η1 ]
 
≥E[ (U − c1 ) + c1 − z1 + − c1 |U − c1 ≥ η2 ]
 
≥E[ (U − c1 ) + c2 − z1 + − c2 |U − c1 ≥ η2 ]
 
≥E[ (U − c2 ) + c2 − z1 + − c2 |U − c2 ≥ η2 ]
=E[(U − z1 )+ − c2 |U − c2 ≥ η2 ]
≥E[(U − z2 )+ − c2 |U − c2 ≥ η2 ] = 0
The first equality and the last equality hold by the definition of z1 and z2 . The first inequality

holds because η1 ≥ η2 . The second inequality holds because c1 < c2 and the term (U −
c1 ) + x − z1 + − x is (weakly) decreasing in x. To show the third inequality holds, notice
under assumption (A2) of the proposition, c1 < c2 implies that U − c1 dominates U − c2

57 Recall that η is a threshold on U − c according to Definition 3.1. Thus η + c is the corresponding threshold on U
i i i i
and the signal is more informative on U if ηi + ci is larger.

52
C. Lyu Journal of Economic Theory 213 (2023) 105722

in the hazard rate order.58 This further implies U − c1 | {U − c1 ≥ η2 } first-order stochastic

dominates U − c2 | {U − c2 ≥ η2 } and thus the inequality holds.59 The last inequality holds
since z1 < z2 as is supposed.
The above result implies E[(U − z1 )+ − c2 |U − c2 ≥ η2 ] = 0 and thus z1 = z2 . This contra-
dicts with what I supposed (i.e., z1 < z2 ). Thus we must have z1 ≥ z2 .

Now, for i = 1, 2, let Zi denote the Pandora’s index for the realized posterior belief given
search cost ci and the corresponding equilibrium signal. Then we have:

U − ci if U ≤ ηi + ci (full disclosure region)
Zi =
zi if U > ηi + ci (pooling region)
Recall that I have shown η1 + c1 ≤ η2 + c2 (part (a) of the proposition) and z1 ≥ z2 above. These
imply that Z1 ≥ Z2 . (When U ≤ η1 + c1 , Z1 = U − c1 > U − c2 = Z2 ; when η1 + c1 < U ≤
η2 + c2 , Z1 = z1 ≥ z2 ≥ η2 ≥ U − c2 = Z2 ; when U > η2 + c2 , Z1 = z1 ≥ z2 = Z2 . ) Thus we
have U ∧ Z1 ≥ U ∧ Z2 .
When c = ci , as I discussed in Section 5.1, the consumer’s expected surplus would be U0 +
E[max{U ∧ Zi − p − U0 , 0}] given any realized U0 .60 Thus U ∧ Z1 ≥ U ∧ Z2 implies that the
consumer surplus is higher when c = c1 compared to that when c = c2 . This concludes the proof
for part (b). 

Appendix E. Proofs for Section 6

E.1. Proof for Proposition 6.1

Proof. By the definition of SG-equilibrium, it suffices to show:


 ex 
ex N −1
(HN , pN ) ∈ arg max{pi
ex ex
HN (x − pi + pN ) dGi (x)} (E.1)
(Gi , pi )
s.t. Gi is a feasible distribution for Ui ∧ Zi
Since (HNex , pN
ex ) is an EG-equilibrium, we know:

 ex 
ex N −1
(HNex , pN
ex
) ∈ arg max{pi HN (x − pi + pN ) dHi (x)} (E.2)
(Hi , pi )
s.t. Hi MP S FU −c
Crucially, Lemma 3.1 implies that the optimization in (E.1) is more constrained than that in (E.2).
Thus, to show condition (E.1) holds, it suffices to check that HNex is indeed a feasible distribution
for Ui ∧ Zi when N is large enough. For this purpose, I need the following property of HNex :

58 See Section 1.3 in Müller and Stoyan (2002) for the definition and properties of hazard rate order. Formally,
assumption (A2) implies that FU has increasing hazard rate and thus for any x s.t. FU (x + c2 ) < 1, we have
fU (x+c1 ) fU (x+c2 ) fU −c1 (x) fU −c2 (x)
1−F (x+c ) ≤ 1−F (x+c ) , which is equivalent to 1−F
U 1 U 2 (x) ≤ 1−F
U −c1 (x) . This implies that U − c1 dominates
U −c2
U − c2 in the hazard rate order by Theorem 1.3.3 in Müller and Stoyan (2002).
59 This follows from the discussion in Müller and Stoyan (2002) right above their Definition 1.3.2.
60 See, for example, Corollary 1 in Choi et al. (2018).

53
C. Lyu Journal of Economic Theory 213 (2023) 105722

Observation (1). For any > 0, there exists N < ∞ such that whenever N ≥ N , we have
HNex (x) = FU −c (x) for all x ≤ ū − c − .

Subproof. This observation is directly implied by Proposition 2 and footnote 17 in Hwang et

al. (2019) (June 5th version). When referring to Hwang et al. (2019), one need to notice that the
consumer’s value for a product in their paper is corresponding to the net-match-utility Ui − c
in my paper, so their range for product value [v, v̄] is corresponding to [u − c, ū − c] in my
model. 

According to the observation, we can find N ∗ such that N ≥ N ∗ ⇒ HNex (x) = FU −c (x), ∀x ≤
ū −2c. Now, fix any N ≥ N ∗ . Let (Si , πi ) be a particular signal under which E[Ui −c|Si ] ∼ HNex .
Since HNex (x) = FU −c (x) for all x ≤ ū − 2c, this signal fully reveals Ui − c less than ū − 2c, and
thus Ui − c ≤ ū − 2c ⇒ Zi = Ui − c < Ui .61 Since Zi ≤ ū − c under any signal, we also have
Ui − c > ū − 2c ⇒ Zi ≤ Ui . Combining these facts, we always have Zi ≤ Ui under (Si , πi ).
This further implies:

0 = E[(Ui − Zi )+ − c|Si ] = E[Ui − Zi − c|Si ] = E[Ui − c|Si ] − Zi

where the first equality holds by the definition of Zi and the second equality holds because
Zi ≤ Ui . Thus E[Ui − c|Si ] = Zi = Ui ∧ Zi under (Si , πi ). Therefore, under (Si , πi ) we also
have Ui ∧ Zi ∼ HNex . This shows that HNex is indeed a feasible distribution for Ui ∧ Zi . 

E.2. Proof for Corollary 6.1

Proof. This is directly implied by Proposition 6.1 and Observation (1) in the proof of Proposi-
tion 6.1 above. 

Appendix F. Other proofs

Proof for Proposition A.1. I first enlarge the model to accommodate search subsidy and price
discount. Let c† denote the objective search cost (without deducting any search subsidy) and let
U † denote the product’s (uncertain) consumption utility. Let b and d denote the search subsidy
and price discount respectively. Let p denote the original product price without any discount.
Given any search subsidy b and price discount d, let c := c† − b and U := U † + d. Then
the consumer’s behavior is characterized in the same way as in the baseline model. In particular,
the consumer would search if Z ≥ U0 + p and would purchase if U ∧ Z ≥ U0 + p, where Z is
defined identically as in Section 3 (with respect to the U and c define here).62 I still use G to
denote the distribution of U ∧ Z under any signal and use J to denote the distribution of U0 .
Under strategy A, we have (b, d) = (y, 0) and the seller’s maximal expected profit given any
original price p is:

max{pP (purchase) − yP (search)} (F.1)

≤ max{(p − y)P (purchase)} (F.2)

61 A formal proof for this is similar to that for Claim 1 in part 2 of the proof for Proposition 3.2.
62 Notice all U , c and Z here depend on the underlying (b, d). I suppress this dependence to ease the notations.

54
C. Lyu Journal of Economic Theory 213 (2023) 105722


max{(p − y) J (x − p)dG(x)}
= G (F.3)
s.t. G MP S FU −c (= FU † −c† +y ); G F OD FU (= FU † )
where the maximizations in the first two lines are over all pre-search signals. The inequality
holds because P (purchase) ≤ P (search); the equality holds because under the unimodality as-
sumption of J , the seller’s optimization over purchase probability (given any (p, b, d)) is fully
characterized by the Relaxed Problem.
Under strategy B, we have (b, d) = (0, y) and the seller’s maximal expected profit given p is:
max{(p − y)P (purchase)} (F.4)

max{(p − y) J (x − p)dG(x)}
= G (F.5)
s.t. G MP S FU −c (= FU † −c† +y ); G F OD FU (= FU † +y )
Comparing optimization (F.5) with optimization (F.3), one can see that they are the same except
that the second constraint in (F.5) is less restrictive. Thus the maximal profit under strategy B is
larger than that under strategy A. Also notice that the inequality in the line of (F.2) holds strictly
when P (purchase) > 0. Thus strategy A is strictly dominated as long as there would be some
search without purchase under it. 

References

Acquisti, A., Taylor, C., Wagman, L., 2016. The economics of privacy. J. Econ. Lit. 54 (2), 442–492.
Anderson, S.P., Renault, R., 2006. Advertising content. Am. Econ. Rev. 96 (1), 93–113.
Anderson, S.P., Renault, R., 2013. The advertising mix for a search good. Manag. Sci. 59 (1), 69–83.
Anderson, S.P., Renault, R., 2018. Firm pricing with consumer search. In: Handbook of Game Theory and Industrial
Organization, vol. 2, pp. 177–224.
Armstrong, M., 2017. Ordered consumer search. J. Eur. Econ. Assoc. 15 (5), 989–1024.
Ashour, S.K., Abdel-hameed, M.A., 2010. Approximate skew normal distribution. J. Adv. Res. 1 (4), 341–350.
Au, P.H., Whitmeyer, M., 2023. Attraction versus persuasion: information provision in search markets. J. Polit. Econ. 131
(1), 202–245.
Bagnoli, M., Bergstrom, T., 2005. Log-concave probability and its applications. Econ. Theory 26 (2), 445–469.
Bar-Isaac, H., Caruana, G., Cuñat, V., 2010. Information gathering and marketing. J. Econ. Manag. Strategy 19 (2),
375–401.
Bar-Isaac, H., Caruana, G., Cuñat, V., 2012. Search, design, and market structure. Am. Econ. Rev. 102 (2), 1140–1160.
Best, J., Quigley, D., 2020. Persuasion for the long run. Available at SSRN 2908115.
Board, S., Lu, J., 2018. Competitive information disclosure in search markets. J. Polit. Econ. 126 (5), 1965–2010.
Boleslavsky, R., Cotton, C.S., Gurnani, H., 2017. Demonstrations and price competition in new product release. Manag.
Sci. 63 (6), 2016–2026.
Choi, M., Dai, A.Y., Kim, K., 2018. Consumer search and price competition. Econometrica 86 (4), 1257–1281.
Choi, M., Kim, K., Pease, M., 2019. Optimal information design for search goods. AEA Pap. Proc. 109, 550–556.
Diamond, P.A., 1971. A model of price adjustment. J. Econ. Theory 3 (2), 156–168.
Dogan, M., Hu, J., 2022. Consumer search and optimal information. Rand J. Econ. 53 (2), 386–403.
Dworczak, P., Kolotilin, A., 2019. The persuasion duality. preprint. arXiv:1910.11392.
Dworczak, P., Martini, G., 2019. The simple economics of optimal persuasion. J. Polit. Econ. 127 (5), 1993–2048.
Folland, G.B., 1999. Real Analysis: Modern Techniques and Their Applications (vol. 40). John Wiley & Sons.
Guo, Y., Shmaya, E., 2019. The interval structure of optimal disclosure. Econometrica 87 (2), 653–675.
Hinnosaar, T., Kawai, K., 2020. Robust pricing with refunds. Rand J. Econ.
Hwang, I., Kim, K., Boleslavsky, R., 2019. Competitive advertising and pricing. Working paper.
Ivanov, M., 2013. Information revelation in competitive markets. Econ. Theory 52 (1), 337–365.
Johnson, J.P., Myatt, D.P., 2006. On the simple economics of advertising, marketing, and product design. Am. Econ.
Rev. 96 (3), 756–784.

55
C. Lyu Journal of Economic Theory 213 (2023) 105722

Kamenica, E., Gentzkow, M., 2011. Bayesian persuasion. Am. Econ. Rev. 101 (6), 2590–2615.
Kleinberg, R., Waggoner, B., Weyl, E.G., 2016. Descending price optimally coordinates search. preprint. arXiv:1603.
07682.
Kolotilin, A., 2018. Optimal information disclosure: a linear programming approach. Theor. Econ. 13 (2), 607–635.
Kolotilin, A., Mylovanov, T., Zapechelnyuk, A., 2021. Censorship as optimal persuasion. Available at SSRN 3783291.
Kolotilin, A., Mylovanov, T., Zapechelnyuk, A., Li, M., 2017. Persuasion of a privately informed receiver. Economet-
rica 85 (6), 1949–1964.
Lewis, T.R., Sappington, D.E., 1994. Supplying information to facilitate price discrimination. Int. Econ. Rev., 309–327.
Mathevet, L., Pearce, D., Stacchetti, E., 2019. Reputation and information design. New York University. Unpublished
paper.
Matyskova, L., Montes, A., 2023. Bayesian persuasion with costly information acquisition. J. Econ. Theory 105678.
Meurer, M., Stahl II, D.O., 1994. Informative advertising and product match. Int. J. Ind. Organ. 12 (1), 1–19.
Müller, A., Stoyan, D., 2002. Comparison Methods for Stochastic Models and Risks, vol. 389. Wiley.
Nelson, P., 1970. Information and consumer behavior. J. Polit. Econ. 78 (2), 311–329.
Rayo, L., Segal, I., 2010. Optimal information disclosure. J. Polit. Econ. 118 (5), 949–987.
Sun, M., 2011. Disclosing multiple product attributes. J. Econ. Manag. Strategy 20 (1), 195–224.
Topkis, D.M., 1998. Supermodularity and Complementarity. Princeton University Press.
Wang, C., 2017. Advertising as a search deterrent. Rand J. Econ. 48 (4), 949–971.
Weitzman, M.L., 1979. Optimal search for the best alternative. Econometrica, 641–654.
Whitmeyer, M., 2020. Persuasion produces the (Diamond) paradox. preprint. arXiv:2011.13900.

56