betaFeedback

WorksEconomics, Econometrics and Finance

Recommendations with Costly Verification

Abstract

A recommender faces consumers who can verify recommendations, but differ in their attention costs or in the salience of their choice of adoption. We characterize the conditions under which a technology that increases transparency by lowering the cost of verifying a product's quality will increase or decrease false recommendations. Without commitment, reducing verification costs never increases false recommendations, but when the recommender can commit to a policy, transparency can worsen both the quality of recommendations and consumer welfare. The necessary and sufficient condition is that the log-elasticity of the distribution of consumer cost types is decreasing. It can hold when the distribution is sufficiently fat-tailed, or when it mixes sophisticated and naive consumers. Our results can be applied to platform-recommended listings, security ratings, and audits.

Introduction

Modern markets are often organized around recommendations. Platforms and rating agencies offer rankings and suggestions of products, services and content, which affect their perception and visibility. Most of the time, these are subject to no oversight except for individual consumers’ ability to fact-check them against other sources. Aside from fact-checking, the only disadvantage to the intermediary of providing false information is that they become less trustworthy.

This paper studies the design of the content of product recommendations by a self-interested recommender when heterogeneous consumers can verify them at a cost. The recommender, e.g., a platform, can be motivated to recommend one product. over another for reasons unrelated to quality, such as profit-sharing with the seller. However, frequent false recommendations make it more attractive to consumers to acquire additional information before acting. Our main finding is simple: increasing transparency to lower consumers’ cost of information acquisition may or may not lead to better recommendations and increased consumer welfare, with a possible negative net effect on consumers when verification costs are heavy-tailed.

In particular, we focus on the case where the intermediary may make a binary recommendation of a preferred, risky product over a safe option (or no recommendation), which, depending on its quality, may be the better or worse option for the consumer. The intermediary knows the risky product’s quality and is able to commit to their preferred recommendation policy, but cannot prevent the consumer from paying an idiosyncratic attention cost to perfectly verify the product’s quality before responding to the recommendation.

We take the canonical example of a transparency-augmenting technology to be one that lowers the cost of verification for all consumers, while fixing relative attention-cost heterogeneity between consumers. Transparency has ambiguous effects on recommendations that depend on the distribution of consumer heterogeneity even when restricting to these cases, in which a shift in the technology affects more- and less-attentive consumers in equal proportion. We also identify the marginal effect of a general, FOSD shift in the consumer distribution: when technology improves costs in a non-proportional manner, consumers benefit most from those that decrease costs slower for marginal consumers.

The reason is that false recommendations trade off the platform’s exploitation of inframarginal adopters with persuasion of those on the margin, who may be swayed to either accept or verify recommendations depending on their level of trust in the recommender. The tradeoff depends on the log-elasticity of the consumer cost distribution because a thin population of near-accepters makes marginal improvements to the policy more appealing to the platform; decreasing the cost of verification for these near-accepters makes them into inframarginal consumers who strictly prefer to verify regardless of a change in the recommendation policy.

These effects rely on commitment. Without commitment, the recommender chooses whether to make a recommendation after observing quality, and unless it recommends every product, it makes false recommendations until the point where it is indifferent between doing so and not. The consumer who is indifferent to verifying is then fixed at the same quantile of the cost distribution, no matter the verification technology, and false recommendations become less frequent as verification gets cheaper.

While we primarily apply the model to a platform recommending products, the framework applies wherever a better-informed, self-interested intermediary guides agents who can verify its recommendation at a heterogeneous cost. For example, e-commerce platforms promote listings that yield higher commissions, while consumers differ in the cost of acquiring outside reviews. Rating agencies are paid by the issuers whose securities they rate, while investors, from retail buyers to sophisticated institutions, have very different costs of conducting independent due diligence. Auditors certify the financial reports of the firms that hire them, while the scrutiny these statements receive from analysts and counterparties varies widely. In each case, the audience of recommender differs in the cost of verifying its recommendation. Any improvements in transparency, such as review aggregators, standardized disclosures, or open-data requirements, could lower these costs. Our results identify when such interventions discipline the recommender and when they instead lead it to make more false recommendations. The answer depends on the intermediary’s commitment power and on the distribution of verification costs near the marginal verifier.

§ 2 describes the model of recommendations, adoption/exploration decisions, and transparency. § 3 describes the recommendation threshold that the platform optimally commits to, and characterizes how it responds to the consumer’s cost of verification, while § 4 focuses on consumer welfare. § 5 drops commitment, shows that transparency then never increases false recommendations, and compares the two benchmarks. Finally, § 6 discusses non-binary recommendations and concludes. Proofs are in the Appendix.

Related literature A large literature studies information intermediaries whose incentives diverge from those of the decision makers they advise (see e.g. Lizzeri 1999, Hagiu and Jullien 2011, Armstrong and Zhou 2011). Closest to us is Armstrong and Chen [2], where an exogenous share of inattentive consumers governs a seller’s quality choice; we endogenize this share as the choice of consumers with different attention costs. Our concept of a recommendation lies under the framework of information design, as in Kamenica and Gentzkow [11]. One recent paper with a similar application is Peitz and Sobolev [19], which studies a monopolist who issues personalized recommendations to consumers who differ in whether they know their own tastes, and characterizes when recommendations are biased. Unlike theirs, our recommender is uninformed about the consumer’s type, and so faces heterogeneity when making a recommendation.

Attention costs have also been well-studied. Models with informed and uninformed consumers [22, 23] share our basic tradeoff between exploiting consumers who do not check and retaining those who do, and sellers may raise verification costs deliberately [6, 8]; our recommender deters verification through the content of its recommendations rather than through the cost. Weitzman [24] and Dovak [5] analyze the problem of costly search through sequential inspection, mandatory or not; our problem is simpler in that there is only one recommendation to verify, and allows us to consider the platform’s design problem under a variety of cost structures. Unlike Fabbri and others who consider entropy-based attention costs [see 20], we focus on the case of perfect verification, but instead allow for a complex cost distribution in the population.

A limited collection of other work focuses on persuasion when the receiver can endogenously obtain information. Matysková and Montes [17] show in a more abstract persuasion environment that a sender’s optimal signal is constructed to deter independent information acquisition, and that cheaper acquisition erodes the sender’s persuasive power. Lukyanov and Safaryan [16] study a sender’s incentive to gather information to persuade an audience that fact-checks at heterogeneous costs, and finds that cheaper fact-checking coarsens the optimal public signal.

Our characterization of the optimal threshold is related to optimal monopoly pricing: the mass of false recommendations is analogous to a price, and the share of consumers who accept a recommendation without verification is analogous to the quantity sold. Our defection elasticity ψ\psi is the generalized failure rate of the cost distribution [13, 12]. It is well-known that the response of a monopolist to a multiplicative demand shift depends on the curvature of demand [4, 25]; in the language of Mr´azov´a and Neary [18], transparency lowers standards when demand is superconvex at the optimum. Likewise, the platform’s response to cost shifts depends on cost elasticity. Two features distinguish our problem. Demand is derived from consumer and product-quality heterogeneity together, so a second elasticity, the dilution elasticity η\eta, also determines the magnitude of the platform’s response. Secondly, costs have a direct effect on consumer welfare. Roesler and Szentes [21] elegantly describe the buyer-optimal information structure under monopoly pricing; we describe the total effect of information and costs generated by costly verification on consumer welfare.

Other work on consumer heterogeneity describes shifting and bunching effects of varying preferences. Johnson and Myatt [10] show that changes in the dispersion of consumer valuations rotate demand and that the comparative statics depend on the position of the rotation point relative to the firm’s operating point; Lewis and Sappington [14] and Anderson and Renault [1] study information supplied by the seller that rotates demand. In our setting, the level and slope of the cost-growth profile at the marginal consumer play a similar role. We show that the effect of a general FOSD shift in transparency can also be broken down into these components.

Model

A platform chooses a product to recommend to a consumer. There are two possible products:

  • The safe/standard product (SS) has known consumption value vS>0v_S > 0 to the consumer, and the platform obtains benefit pSp_S if the consumer chooses to adopt it.

  • The risky product (RR) has a consumption value vv known only to the platform. The platform obtains a benefit pRp_R from recommending the risky product to the consumer.

We assume that pR>pSp_R > p_S, that is, the platform prefers that the consumer adopt the risky product to having them adopt the safe product. There can be several possible reasons for this: the platform may want to generate reviews on a new product in order to feed its algorithm, or its share of the revenue from the risky product could be higher. It will be convenient to work with the normalized margin

γ≡pSpR∈[0,1).\gamma\equiv\frac{p_S}{p_R} \in[0,1).

The consumption values vSv_S and vv give a reduced-form description of the consumer’s utility. We assume that vS≥0v_S \ge0, so that the consumer has a straightforward preference to accept the recommendation of the safe product.

Our focal point is the problem of the consumer when recommended the risky product: they do not know vv, though they know the ex-ante distribution of vv is given by GG; we assume it is supported on an interval of the form [v‾,v‾][\underline{v},\overline{v}]. Note that v‾\underline{v} can be taken as the expected consumption value given the platform’s information – the platform need not be fully informed, only better-informed than the consumer.

Assumption 1. Quality distribution GG is continuous, with a density gg that is continuous and strictly positive on [v‾,0)[\underline{v},0) and on (0,v‾](0,\overline{v}], where v‾<0<v‾\underline{v} < 0 < \overline{v}; gg may jump at $v = 0.

For products listed by a seller with an exogenous price, we could think of the consumption value as the total of the product’s (possibly unknown) quality minus its (known) price, v=q−pv = q - p. Alternatively, if the product is not priced – e.g., it is content – then the consumption value directly describes the consumer’s welfare.

Heterogeneous verification costs The consumer is able to check the quality of a product before choosing to adopt it or not. The consumer’s cost of verification depends on a privately known attention cost parameter, xx, which is positive and distributed ∼Fc\sim F_c in the population. Verification is perfect: the consumer learns vv perfectly after paying the attention cost xx, and is then informed when choosing whether to accept or reject the risky recommendation.

We also allow the cost of verification to depend on a shared, commonly known parameter cc that describes the verification technology that determines the opacity of information about the (risky) product. We assume that the single-parameter family of distributions FcF_c describes costs that are smoothly increasing in cc in the sense of first-order stochastic dominance.

Assumption 2 (Differentiability and FOSD). For every cc, FcF_c is an atomless distribution on [0,∞)[0,\infty) whose support is an interval [x‾c,x‾c][\underline{x}_c,\overline{x}_c] with 0≤x‾c<x‾c≤∞0 \le\underline{x}_c < \overline{x}_c \le\infty; the endpoints may depend on cc, and x‾c\underline{x}_c is nondecreasing in cc. On the interior of the

support, {(x,c):x‾c<x<x‾c}\{(x,c): \underline{x}_c < x < \overline{x}_c\}, Fc(x)F_c(x) is twice continuously differentiable in (x,c)(x,c) with fc(x)>0f_c(x)>0 and ∂Fc(x)∂c<0\frac{\partial F_c(x)}{\partial c}<0.2

It will be helpful to index the consumer’s type by its quantile, q∈[0,1]q \in[0,1], so that the cost borne by the consumer at quantile qq is

κ(c,q)=Fc−1(q).\kappa(c,q)=F_c^{-1}(q).

Before proceeding further, we highlight the case where interventions in transparency (or opacity) have a uniformly proportional impact on all consumers.

Definition 2.1. The family FcF_c with quantile cost function κ(c,q)\kappa(c,q) features a uniformly proportional effect of transparency if there is some monotone φ\varphi such that

κ(c,q)=cφ(q).(1)\kappa(c,q)=c\varphi(q). \tag*{(1)}

The uniformly proportional case captures the benchmark in which lowering cc allows all consumers to learn quality vv more cheaply, but keeps relative heterogeneity unchanged. This assumption makes sense if consumer heterogeneity is orthogonal to the overall level of transparency – e.g., if relative consumer heterogeneity captures heterogeneous opportunity cost of time, and cc captures the amount of time it takes to navigate to a third-party opinion, then technology can be represented with a family satisfying (1).

The consumer’s decision proceeds as follows:

  1. Observe own cost parameter, κ(c,q)\kappa(c,q), and which product (SS or RR) is recommended. Choose whether the initial action a1a_1 is to Accept (AA), Reject (RR), or Explore (EE) the product. If a1∈{A,R}a_1 \in\{A,R\}, the decision is over and payoffs are realized.

  2. If a1=Ea_1=E, pay κ(c,q)\kappa(c,q) and observe a perfect signal of vv. Then, choose the post-exploration action a2∈{A,R}a_2 \in\{A,R\}.

Denote the consumer’s final decision between accept and reject (regardless of whether it is made in period 1 or 2) by dd. The consumer’s payoff is

uR=v⋅1d=A−κ(c,q)⋅1a1=E.u_R=v\cdot\mathbf{1}_{d=A}-\kappa(c,q)\cdot\mathbf{1}_{a_1=E}.

Indeed, some parts of the consumer’s decision are simple: upon seeing the safe product recommended (with vS>0v_S>0), the consumer will choose to accept the product. However, exploration may have value when the consumer sees the risky recommendation instead.

We will show that their optimal decision path upon a risky recommendation depends on the platform’s recommendation policy.

Recommendation policies Upon being shown the premium product, the consumer updates beliefs via Bayes Rule:

β(v∣π)=π(v)g(v)∫π(v)g(v) dv.\beta(v \mid\pi) = \frac{\pi(v)g(v)}{\int\pi(v)g(v)\,dv}.

The platform’s recommendation policy is a function π(v)∈[0,1]\pi(v) \in[0,1] that specifies the probability of recommending the premium product for each quality premium vv.

The consumer’s expected payoff under each initial action a1a_1, conditional on receiving a premium recommendation:

  • Accept:

UA=Eβ[v].U_A = \mathbb{E}_{\beta}[v].
  • Explore: After paying κ(c,q)\kappa(c,q) and learning vv, the consumer accepts the premium product iff v≥0v \ge0. Expected payoff:

UE=Eβ[v+]−κ(c,q),where v+=max⁡(v,0).U_E = \mathbb{E}_{\beta}[v^+] - \kappa(c,q), \qquad\text{where } v^+ = \max(v,0).
  • Reject: UR=0U_R = 0.

The comparison between acceptance and rejection is governed by the posterior mean

μ(π)=Eβ[v].\mu(\pi) = \mathbb{E}_{\beta}[v].

If μ(π)≥0\mu(\pi) \ge0, high-cost consumers weakly prefer accepting blindly to rejecting. In this region, the consumer population separates into explorers, those with κ(c,q)<aR(π)\kappa(c,q) < a_R(\pi), and blind accepters, those with κ(c,q)≥aR(π)\kappa(c,q) \ge a_R(\pi). This is the main region studied below.

Exploration is valuable because it allows the consumer to avoid the negative part of the recommendation. The value of exploration relative to blind acceptance is

UE−UA=Eβ[v+]−κ(c,q)−Eβ[v]=Eβ[(−v)+]−κ(c,q).U_E - U_A = \mathbb{E}_{\beta}[v^+] - \kappa(c,q) - \mathbb{E}_{\beta}[v] = \mathbb{E}_{\beta}[(-v)^+] - \kappa(c,q).

Define the posterior downside

aR(π)=Eβ[(−v)+].a_R(\pi) = \mathbb{E}_{\beta}[(-v)^+].

Then exploration is preferred to blind acceptance if and only if

κ(c,q)≤aR(π).(2)\kappa(c,q) \le a_R(\pi). \tag*{(2)}

(2) shows that verification is selected by low-cost consumers, while high-cost consumers accept blindly, provided that blind acceptance is preferred to rejection.

Optimal Recommendation under Commitment

In the commitment benchmark, the recommender commits to a measurable policy π\pi before consumers act. This section characterizes the optimal policy and then asks how it responds to the cost of verification.

Properties of the optimal policy

A recommendation policy is a function π:[v‾,v‾]→[0,1]\pi: [\underline{v},\overline{v}] \to[0,1] giving the probability that the risky product is recommended at quality vv. Write

L(π)=∫π(v)g(v) dv,H(π)=∫v≥0π(v)g(v) dv,B(π)=∫v<0π(v)g(v) dv.L(\pi)=\int\pi(v)g(v)\,dv,\qquad H(\pi)=\int_{v\geq0}\pi(v)g(v)\,dv,\qquad B(\pi)=\int_{v<0}\pi(v)g(v)\,dv.

for the frequency of risky recommendations, the mass of good products recommended, and the mass of bad products recommended. Denote by

A(π)=∫v<0(−v)π(v)g(v) dvA(\pi)=\int_{v<0}(-v)\pi(v)g(v)\,dv

the total posterior downside carried by the pool. The posterior downside per recommendation is a(π)=A(π)/L(π)a(\pi)=A(\pi)/L(\pi).

Acceptance/verification region. The case of commitment can be thought of as a Stackelberg game in which the platform first chooses a recommendation policy, and then the consumer chooses their response to the realized recommendation once the policy has been fixed. In every equilibrium of this Stackelberg game, the platform’s recommendation induces a weakly positive posterior mean,

μ(π)=∫vvπ(v)g(v) dvL(π)≥0,\mu(\pi)=\frac{\int_v v\pi(v)g(v)\,dv}{L(\pi)}\geq0,

and a consumer rejects a recommendation with probability 0.

Lemma 1. Under any optimal policy the consumer’s posterior expectation of recommended quality is weakly positive, μ(π)≥0\mu(\pi)\geq0, and the platform expects the consumer to never immediately reject a recommendation.

Proof. See Appendix.

Given that the consumer either accepts or verifies the recommendation of the risky product, total demand for the risky product is

D(π)=H(π)+(L(π)−H(π))[1−Fc(A(π)L(π))].D(\pi)=H(\pi)+(L(\pi)-H(\pi))\left[1-F_c\left(\frac{A(\pi)}{L(\pi)}\right)\right].

The platform’s optimal choice of π\pi maximizes

Π(π)=D(π)pR+(1−L(π))pS.(3)\Pi(\pi)=D(\pi)p_R+(1-L(\pi))p_S. \tag*{(3)}

Demand under a threshold. The platform’s objective is supermodular in the pairing between recommendation probability and quality, so the optimal coupling is monotone and takes the form of a threshold quality, v^\hat{v}, above which the recommendation is always made and below which it is never made.

Proposition 2 (Threshold optimality). Fix any recommendation frequency L∈(0,1)L \in(0,1). Among all policies with ∫πg=L\int\pi g = L, risky demand is maximized by recommending the top LL mass of qualities. Hence an optimal commitment policy may be taken to be π∗(v)=1v≥v^\pi^*(v) = \mathbb{1}_{v \ge\hat{v}} for some threshold v^\hat{v}.

Proof. In the acceptance–verification region every recommended good product is eventually adopted, while a recommended bad product is adopted only by consumers who accept blindly. Hence, D(π)D(\pi) is increasing in H(π)H(\pi) and decreasing in A(π)A(\pi) for fixed LL.

Suppose π\pi recommends some quality v1v_1 with positive probability while excluding some v2>v1v_2 > v_1. Shifting recommendation probability from v1v_1 to v2v_2 holds LL fixed, weakly raises HH, and weakly lowers AA, strictly so whenever the shift crosses zero or removes a worse bad product. Iterating such rearrangements yields the upper-contour policy 1v≥v^\mathbb{1}_{v \ge\hat{v}}.

For a threshold v^\hat{v} write

L(v^)=1−G(v^),B(v^)=G(0)−G(v^),H=1−G(0),L(\hat{v}) = 1 - G(\hat{v}), \qquad B(\hat{v}) = G(0) - G(\hat{v}), \qquad H = 1 - G(0),

so that HH is the mass of good products and B(v^)B(\hat{v}) the mass of bad products admitted into the pool. The posterior expected downside and the quantile of the marginal consumer are

a(v^)=∫v^0(−v)g(v) dv1−G(v^),qA(v^,c)=Fc(a(v^)),a(\hat{v}) = \frac{\int_{\hat{v}}^{0} (-v)g(v)\,\mathrm{d}v}{1 - G(\hat{v})}, \qquad q_A(\hat{v},c) = F_c(a(\hat{v})),

and risky demand is

D(v^)=H+B(v^)[1−Fc(a(v^))].(4)D(\hat{v}) = H + B(\hat{v})[1 - F_c(a(\hat{v}))]. \tag*{(4)}

(4) separates demand into an always-inframarginal component, the good mass HH, which is adopted whatever the standard, and a component that is affected by the marginal consumer, the bad mass B(v^)B(\hat{v}) multiplied by the share S(c,a)=1−Fc(a)S(c,a) = 1 - F_c(a) of consumers who accept blindly. Lowering the standard adds bad mass but raises the posterior downside, inducing more consumers to verify.

Remark 1 (Verification region). If xc>0x_c > 0, no consumer verifies when a(v^)≤xca(\hat{v}) \le x_c, so S(c,a(v^))=1S(c,a(\hat{v})) = 1 and (3) reduces to pR(1−G(v^))+pSG(v^)p_R(1 - G(\hat{v})) + p_SG(\hat{v}), which is strictly decreasing in v^\hat{v} because pR>pSp_R > p_S. The recommender therefore never stops in this region: every optimal threshold satisfies a(v^∗)≥xca(\hat{v}^*) \ge x_c. Write v^c\hat{v}_c for the threshold with a(v^c)=xca(\hat{v}_c) = x_c, with v^c=0\hat{v}_c = 0 when xc=0x_c = 0. The analysis below is carried out on the verification region {v^:v^0<v^<v^c}\{\hat{v} : \hat{v}_0 < \hat{v} < \hat{v}_c\}, on which a(v^)a(\hat{v}) lies in the interior of the support of FcF_c, so that SS, ψ\psi and their derivatives are well defined and smooth. At v^c\hat{v}_c itself ψ\psi jumps from 00 to xcfc(xc)x_c f_c(x_c), so the corner v^c\hat{v}_c can in principle be optimal; in all our examples the optimum is interior to the verification region, which we verify numerically.

An Elasticity Formula

The recommender’s problem can be seen as the following: they have access to a mass of good recommendations, which are used to subsidize the value of bad recommendations (v<0v < 0). The goal is to maximize the mass of bad recommendations accepted, since the quantity of true “good” recommendations, HH, is fixed and always accepted.

For bad recommendations, there is an effective demand curve described by the tradeoff between B(v^)B(\hat{v}) and S(c,a)S(c,a), the fraction of blind accepters. An increase in B(v^)B(\hat{v}) decreases S(c,a)S(c,a), since verification becomes more valuable as posterior expected downside increases. Here, B(v^)B(\hat{v}) serves as a quantity margin, and S(c,a)S(c,a) as a price, since if S(c,a)=0S(c,a)=0, then each false recommendation generates 0 profit for the recommender.

In contrast to standard pricing environments, the tradeoff depends not just on consumer variation, but also on the rate of dilution of the recommendation pool that corresponds to a given shift in the marginal consumer.

Define the defection elasticity

ψ(a,c)=afc(a)1−Fc(a)=−dln⁡S(c,a)dln⁡a,(5)\psi(a,c)=\frac{af_c(a)}{1-F_c(a)}=-\frac{d\ln S(c,a)}{d\ln a}, \tag*{(5)}

driven by the distribution of consumers FF, to be the proportional rate at which consumers “defect” to verification due to a change in the posterior downside aa, and the dilution elasticity,

η(v^)=dln⁡B(v^)dln⁡a(v^)=a(v^)(1−G(v^))(G(0)−G(v^))(∣v^∣−a(v^)),(6)\eta(\hat{v})=\frac{d\ln B(\hat{v})}{d\ln a(\hat{v})}=\frac{a(\hat{v})(1-G(\hat{v}))}{(G(0)-G(\hat{v}))(|\hat{v}|-a(\hat{v}))}, \tag*{(6)}

driven by quality distribution GG, to be the proportional rate of increase of the mass of bad recommendations due to the same change in the posterior downside. Their ratio is the elasticity of the inverse demand curve,

ξ(v^)=−dln⁡Sdln⁡B=ψ(a(v^),c)η(v^).(7)\xi(\hat{v})=-\frac{d\ln S}{d\ln B}=\frac{\psi(a(\hat{v}),c)}{\eta(\hat{v})}. \tag*{(7)}

Proposition 3 (Optimal recommendation threshold). Any optimum in the interior of the verification region, v^∗∈(v^0,v^c)\hat{v}^{*}\in(\hat{v}_0,\hat{v}_c), satisfies

ξ(v^∗)=ψ(a(v^∗),c)η(v^∗)=1−γS(c,a∗),(8)\xi(\hat{v}^{*})=\frac{\psi(a(\hat{v}^{*}),c)}{\eta(\hat{v}^{*})}=1-\frac{\gamma}{S(c,a^{*})}, \tag*{(8)}

where a∗=a(v^∗)a^{*}=a(\hat{v}^{*}); the RHS is a Lerner index with price S(c,a∗)S(c,a^{*}) and marginal cost γ\gamma.

Moreover,

Π′(v^)=pRg(v^)S(c,a(v^))[ξ(v^)−1+γS(c,a(v^))],(9)\Pi'(\hat{v})=p_Rg(\hat{v})S(c,a(\hat{v}))\left[\xi(\hat{v})-1+\frac{\gamma}{S(c,a(\hat{v}))}\right], \tag*{(9)}

so Π′\Pi' has the sign of the bracket, and the optimum is unique whenever ξ(v^)−1+γS(c,a(v^))\xi(\hat{v})-1+\frac{\gamma}{S(c,a(\hat{v}))} is strictly decreasing in v^\hat{v}.

Proof. Differentiating (3) with G′=gG'=g, B′=−gB'=-g and S′=−fcS'=-f_c,

Π′(v^)=pR(γg−gS(c,a(v^))+B(v^)fc(a(v^))∣a′(v^)∣)=−pR(g[S(c,a(v^))−γ]−B(v^)fc(a(v^))∣a′(v^)∣).\begin{aligned} \Pi'(\hat{v})=p_R\left(\gamma g-gS(c,a(\hat{v}))+B(\hat{v})f_c(a(\hat{v}))\lvert a'(\hat{v})\rvert\right) \\ &=-p_R\left(g\left[S(c,a(\hat{v}))-\gamma\right]-B(\hat{v})f_c(a(\hat{v}))\lvert a'(\hat{v})\rvert\right). \end{aligned}

By (7), B(v^)fc(a(v^))∣a′(v^)∣=gS(c,a(v^))ξ(v^)B(\hat{v})f_c(a(\hat{v}))\lvert a'(\hat{v})\rvert=gS(c,a(\hat{v}))\xi(\hat{v}), so Π′=pRg[Sξ−S+γ]\Pi'=p_Rg\left[S\xi-S+\gamma\right], which is (9). Setting Π′=0\Pi'=0 and dividing by S>0S>0 gives (8).

Note that since setting the optimal threshold is analogous to setting a monopoly price through false recommendations, the natural tradeoff is between marginal buyers and the deepening of the pool of bad recommendations. The term η(v^)\eta(\hat{v}) translates a shift in the marginal buyer to the impact on the mass of accepted false recommendations that the platform benefits from, so while the platform’s first order condition depends on consumer heterogeneity through ψ\psi as in standard monopoly pricing, it also depends on the product quality distribution GG through η\eta.

Remark 2 (Regularity). Uniqueness in Proposition 3 requires ξ(v^)−1+γ/S(c,a(v^))\xi(\hat{v})-1+\gamma/S(c,a(\hat{v})) to be strictly decreasing in v^\hat{v} where it crosses zero. This is analogous to Myerson-type regularity in monopoly pricing, except that here ψ\psi is the failure rate, evaluated at the endogenous verification cutoff a(v^)a(\hat{v}).

The Lerner term γ/S(c,a(v^))\gamma/S(c,a(\hat{v})) is always decreasing in v^\hat{v}: raising the standard lowers the posterior downside, so fewer consumers verify. But in addition, the demand elasticity ξ(v^)=ψ(a(v^),c)/η(v^)\xi(\hat{v})=\psi(a(\hat{v}),c)/\eta(\hat{v}) must be decreasing in v^\hat{v}. Since a(v^)a(\hat{v}) decreases with v^\hat{v}, a simple sufficient condition is that (i) ψ\psi is nondecreasing in aa, i.e. the consumer side satisfies IGFR, and (ii) the dilution elasticity η\eta is nondecreasing in v^\hat{v}.

Verification Costs and Recommendation Standards

This section discusses the effect of transparency on optimal recommendation thresholds. We think of transparency as being controlled by the verification technology cc: when cc increases, verification becomes more costly for every consumer.

The main result shows that increased transparency need not always discipline the recommender towards fewer bad recommendations. The effect of a proportional shift in transparency when the safe option is worth 0 to the platform (or the platform can recommend the risky product without excluding the safe one) depends on the elasticity of consumer-side elasticity ψ\psi; non-proportional shifts also affect the policy by reshaping the distribution of costs in the neighborhood of the marginal consumer.

To see this, note that the optimality condition can be summarized as

Φ(v^,c)=ξ(v^,c)−1+γS(a(v^))=0(10)\Phi(\hat{v},c)=\xi(\hat{v},c)-1+\frac{\gamma}{S(a(\hat{v}))}=0 \tag*{(10)}

where we assume that we are in the case where Φ(v^,c)\Phi(\hat{v},c) is strictly decreasing.

We can describe how unevenly a change in the technology affects the population by the cost-growth profile

e(c,q)≡∂ln⁡κ(c,q)∂ln⁡c≥0,(11)e(c,q) \equiv\frac{\partial\ln\kappa(c,q)}{\partial\ln c} \ge0, \tag*{(11)}

which records the elasticity with respect to cc of the cost borne by the consumer at quantile qq. A family of distributions is uniformly proportional if e(c,q)=1e(c,q)=1, i.e., every consumer’s cost rises with unit elasticity in cc. Write

e∗=e(c,qA),e′∗=∂e(c,q)∂q∣q=qAe^* = e(c,q_A), \qquad e^{\prime*} = \left.\frac{\partial e(c,q)}{\partial q}\right|_{q=q_A}

for the level and slope of the cost-growth profile (11) at the marginal quantile, and denote the elasticity of the defection elasticity by

Λ≡∂ln⁡ψ(a,c)∂ln⁡a.(12)\Lambda\equiv\frac{\partial\ln\psi(a,c)}{\partial\ln a}. \tag*{(12)}

Intuitively, Λ\Lambda describes the rate at which a shift in the marginal consumer aa thickens or thins out the density of near-marginal vs. inframarginal consumers. This affects the optimal standard because a decrease in transparency (increase in cc) shifts the marginal consumer independently of a shift in the recommendation policy and has a first-order impact on the platform’s incentive to capture vs. exploit consumers.

Lemma 4. At any v^\hat{v},

∂S∂c=e∗cψS≥0,∂ψ∂c=−ψc[e∗Λ+e′∗ψS].\frac{\partial S}{\partial c}=\frac{e^*}{c}\psi S\ge0,\qquad\frac{\partial\psi}{\partial c}=-\frac{\psi}{c}\left[e^*\Lambda+e^{\prime*}\psi S\right].

Proposition 5 (Opacity comparative static).

sign⁡(dv^∗/dc)=−sign⁡[e∗(Λ+γη(v^∗)S(c,a∗))+e′∗ψS(c,a∗)].(13)\operatorname{sign}(d\hat{v}^*/dc)=-\operatorname{sign}\left[e^*\left(\Lambda+\frac{\gamma\eta(\hat{v}^*)}{S(c,a^*)}\right)+e^{\prime*}\psi S(c,a^*)\right]. \tag*{(13)}

Proof. Since ∂Φ/∂v^<0\partial\Phi/\partial\hat{v}<0, the implicit function theorem gives sign⁡(dv^∗/dc)=sign⁡(∂Φ/∂c)\operatorname{sign}(d\hat{v}^*/dc)=\operatorname{sign}(\partial\Phi/\partial c). By Lemma 4,

∂Φ∂c=1η∂ψ∂c−γS2∂S∂c=−ψcη{e∗Λ+e′∗ψS+γe∗ηS}.(14)\frac{\partial\Phi}{\partial c}=\frac{1}{\eta}\frac{\partial\psi}{\partial c}-\frac{\gamma}{S^2}\frac{\partial S}{\partial c}=-\frac{\psi}{c\eta}\left\{e^*\Lambda+e^{\prime*}\psi S+\frac{\gamma e^*\eta}{S}\right\}. \tag*{(14)}

To understand the proposition, start with the simplest case of uniformly proportional shifts with γ=0\gamma=0, where direction of the policy response depends only on the sign of Λ\Lambda, which could be either positive or negative. We give examples of both in the subsection to follow, and show that when fcf_c is fat-tailed ψ\psi can be decreasing throughout the verification region, or non-monotone in aa with Λ<0\Lambda< 0 at a∗a^*.

The reason is that opacity lowers the share of verifiers qA=Fc(a∗)q_A = F_c(a^*) at the current standard. When ψ\psi is increasing in aa, this lowers ψ\psi and makes it cheaper than before for the recommender to add bad recommendations; when ψ\psi is decreasing, it raises ψ\psi and makes adding bad recommendations more expensive (Figure 1). Assuming uniform proportionality and no opportunity cost of making a recommendation,3 this shift in the marginal consumer is the only strategically relevant change for the recommender4. When there is instead a positive cost of forgoing the safe recommendation (γ>0\gamma> 0), verifiers are particularly costly, and opacity is more prone to increase false recommendations since it decreases the number of verifiers. This is reflected in the lower, strictly negative threshold in (15).

Figure 1. Opacity and the defection elasticity. Notes: Proportional families, γ=0\gamma= 0. Verification becomes costlier (c0→c1c_0 \to c_1) while the standard v^\hat{v}, and hence the posterior downside aa, is held fixed. Top: density of ln⁡x\ln x; its height at the cutoff is afc(a)a f_{c}(a) and the shaded area is the blind-accepter share SS, so ψ=afc(a)/S\psi= a f_{c}(a)/S. Bottom: ψ\psi against a/ca/c, with η(v^)\eta(\hat{v}) unaffected by cc. (a) Exponential costs (Example 1): ψ\psi falls below η\eta and v^\hat{v} falls. (b) Log-Pareto costs (Example 3, ν=1\nu= 1): ψ\psi rises above η\eta and v^\hat{v} rises. Cutoffs are illustrative.

Corollary 1 (Uniformly proportional transparency). If the family FcF_c features a uniformly proportional effect of transparency, then e∗=1e^* = 1 and e′∗=0e^{\prime*} = 0, so

dv^∗dc<0⟺Λ>−γη(v^∗)S(c,a∗)≤0.(15)\frac{\mathrm{d}\hat{v}^*}{\mathrm{d}c} < 0 \quad\Longleftrightarrow\quad\Lambda> -\frac{\gamma\eta(\hat{v}^*)}{S(c,a^*)} \le0. \tag*{(15)}

Outside of the proportional case, there is an additional spread effect, which describes the relative bunching or spreading of the cost distribution just left of the marginal buyer induced by the shift in the verification technology cc. If e′∗<0e^{\prime*} < 0, then costs rise faster for consumers marginally more attentive than the marginal one and the cost distribution is locally compressed around a(v^)a(\hat{v}). Equivalently, the density fc(a)f_c(a) rises, defection becomes more elastic, and the recommender gains a stronger incentive to persuade marginal buyers to accept the product without verification. If e′∗>0e^{\prime*} > 0 the distribution is locally spread out, which means the recommender becomes able to more cheaply flood the pool with false recommendations.

This effect is fully responsible for the direction of the difference between the recommender’s response to a general transparency shift vs. the comparable proportional shift if the rate of change of costs for the marginal buyer is normalized to be the same in both cases (e∗=1e^* = 1). To see how this might be useful, take the case of comparing the uniformly proportional cost-reduction (e.g., time-savings, as explained in § 2) to a policy intervention that results in a uniform additive reduction in verification costs, e.g., reducing the payment for a third-party information source, that affects the marginal buyer equally. Since e∗<0e^* < 0 for the additive intervention, it induces a change in the platform’s behavior that is more beneficial to the consumer (though this does not account for the effect of the costs themselves).

More generally, suppose that a policymaker wishes to compare the marginal effect of any two smooth transformations of the cost distribution, parametrized by families Fc1F_c^1 and Fc2F_c^2 with local cost-growth profiles e1∗e_1^* and e2∗e_2^* respectively.

Corollary 2 (Comparison with proportional transparency). Fix the initial cost distribution and compare a shift with profile e1∗e_1^* to a shift with profile e2∗e_2^*. The former

yields fewer false recommendations if and only if

(e1∗−e2∗)[Λ+γηS]+(e1′∗−e2′∗)ψS<0.(e_1^* - e_2^*)\left[\Lambda+ \frac{\gamma\eta}{S}\right] + (e_1^{\prime*} - e_2^{\prime*})\psi S < 0.

Examples

We now give a couple of example parametrizations, under the assumption of uniformly proportional families FcF_c, that illustrate the role of consumer heterogeneity in FcF_c. We also set γ=0\gamma= 0 throughout so that the sign of the effect of a shift in opacity cc is the same as that of Λ\Lambda.

Example 1 (Light-tailed types: opacity lowers standards). Let quality be uniform, v∼U[−1,1]v \sim U[-1,1], and let verification costs be exponential, Fc(x)=1−e−x/cF_c(x) = 1 - e^{-x/c} on (0,∞)(0,\infty), so that κ(c,q)=−cln⁡(1−q)\kappa(c,q) = -c\ln(1-q). The defection elasticity is ψ(a,c)=a/c\psi(a,c) = a/c, so Λ=1>0\Lambda= 1 > 0. The first-order condition ψ(a(v^),c)=η(v^)\psi(a(\hat v),c) = \eta(\hat v) reduces to the cubic

v^2(2+∣v^∣)=2c(1+∣v^∣)2,(16)\hat v^2(2+|\hat v|) = 2c(1+|\hat v|)^2, \tag*{(16)}

which has a unique root, interior if and only if c<3/8c < 3/8, with dv^∗/dc<0d\hat v^*/dc < 0. Note that as c→0c \to0, v^→0\hat v \to0 as well: vanishing costs result in no false recommendations.

Example 2 (Pareto tail: knife-edge case). For a Pareto tail with survival 1−Fc(x)=(c/x)α1-F_c(x) = (c/x)^\alpha on x≥cx \ge c, so that xc=cx_c = c, ψ(a,c)=α\psi(a,c) = \alpha is constant on the verification region and Λ=0\Lambda= 0. Local changes in cc have no first-order effect through the consumer-side elasticity.

Example 3 (Heavy-tailed types: opacity raises standards). Let bad qualities have a thin tail: on [−1,0)[-1,0), the density is exponential in ∣v^∣|\hat v| with scale τ\tau, g(v)=κev/τg(v) = \kappa e^{v/\tau}, and it is uniform, g=12g = \frac{1}{2}, on [0,1][0,1].5 Let verification costs be log-Pareto, 1−Fc(x)=(1+ln⁡(x/c))−ν1-F_c(x) = (1+\ln(x/c))^{-\nu} on [c,∞)[c,\infty), so that xc=cx_c = c and, on the verification region a>ca > c, ψ(a,c)=ν/(1+ln⁡(a/c))\psi(a,c) = \nu/(1+\ln(a/c)) is strictly decreasing in aa, with Λ=−1/(1+ln⁡(a/c))<0\Lambda= -1/(1+\ln(a/c)) < 0; for a<ca < c nobody verifies and ψ=0\psi= 0, and Remark 1 shows the optimum never lies there. When the standard is already sufficiently low, the average downside saturates at the tail mean, a(v^)→τ/2a(\hat v) \to\tau/2, while the marginal bad product keeps deteriorating, so ψ(a,c)\psi(a,c) is decreasing. Note that for sufficiently low cc, v^=v\hat v = v: vanishing costs lead to maximal false recommendations.

Example 4 (Bimodal types: two groups of consumers). Keep the quality distribution of Example 3. Let a fraction pp of consumers (“sophisticated”) have exponential costs with mean cs1cs_1 and the rest (“naive”) exponential costs with mean cs2>cs1cs_2 > cs_1, so that $1-F_c(x) = p e^{-x/(cs_1)} + (1-p)e^{-x/(cs_2)}$. The family is uniformly proportional, satisfies Assumption 2, and both components are light-tailed. Writing λi=1/(csi)\lambda_i = 1/(cs_i) and $w(a)

for the share of sophisticated consumers among blind accepters, the hazard rate is h(a)=wλ1+(1−w)λ2h(a) = w\lambda_{1} + (1-w)\lambda_{2}, ψ=ah(a)\psi= a h(a), and

Λ(a,c)=1−aw(1−w)(λ1−λ2)2h(a).(17)\Lambda(a,c) = 1 - \frac{a w(1-w)(\lambda_{1}-\lambda_{2})^{2}}{h(a)}. \tag*{(17)}

Since w(1−w)w(1-w) is the variance of group membership among blind accepters, Λ<0\Lambda< 0 exactly where that pool is tipping from mostly sophisticated to mostly naive: a higher downside drives out the sophisticated, while those who remain are increasingly insensitive to it. With p=0.6p = 0.6, s1=0.05s_{1} = 0.05, s2=1s_{2} = 1 and τ=0.15\tau= 0.15 the optimum is unique and lies in this region for c∈(0.26,0.44)c \in(0.26, 0.44), where dv^∗/dc>0\mathrm{d}\hat{v}^{*}/\mathrm{d}c > 0: as cc falls from 0.400.40 to 0.300.30, v^∗\hat{v}^{*} falls from −0.230-0.230 to −0.465-0.465.

Comparison. Both the light- and heavy-tailed examples feature interior thresholds at c=0.05c = 0.05 and c=0.06c = 0.06, so we can compare the same increase in the verification cost:

Example 1 (Λ>0\Lambda> 0)Example 3 (Λ<0\Lambda< 0)
c=0.05c = 0.05c=0.06c = 0.06c=0.05c = 0.05c=0.06c = 0.06
threshold v^∗\hat{v}^{*}−0.266-0.266−0.296-0.296−0.727-0.727−0.590-0.590
posterior downside a∗a^{*}0.0280.0280.0340.0340.0720.0720.0690.069
Λ\Lambda+1.00+1.00+1.00+1.00−0.73-0.73−0.88-0.88
false recs B(v^∗)B(\hat{v}^{*})0.1330.1330.1480.1480.4970.4970.4910.491
response to c↑c \uparrowv^∗↓\hat{v}^{*} \downarrow/false recs ↑\uparrowv^∗↑\hat{v}^{*} \uparrow/false recs ↓\downarrow

Table 1.

Note that the same increase in cc moves the two thresholds in opposite directions. An even starker illustration may be what happens as c↓0c \downarrow0: in this case, the recommender has very little power to persuade the consumer, as the consumer will verify recommendations unless they are a massive cost-outlier or the recommendation is almost perfect. However, under the fat- and thin-tailed distributions the recommender is pushed to opposite extremes: in Example 1, with thin tails, there are few outliers and persuasion means that false recommendations disappear from the policy. In Example 3, as cc vanishes, taking advantage of a fat tail becomes magnitudes more important than persuading the marginal consumer, and so the optimal policy in the limit is full exploitation.

Regime switches. With a larger sophisticated share the two-group family lowers standards discontinuously. The objective (4) splits into a persuasion term pB(v^)e−a(v^)/(cs1)p B(\hat{v}) e^{-a(\hat{v})/(c s_{1})}, which peaks at a shallow standard, and an exploitation term (1−p)B(v^)e−a(v^)/(cs2)(1-p) B(\hat{v}) e^{-a(\hat{v})/(c s_{2})}, which peaks at a deep one, so Π\Pi can have two local maxima. With p=0.7p = 0.7, s1=0.05s_{1} = 0.05, s2=1s_{2} = 1 and τ=0.15\tau= 0.15 the shallow branch is optimal for c>0.170c > 0.170 and the deep branch for c<0.170c < 0.170 (Figure 2): at the switch the threshold jumps from −0.073-0.073 to −0.382-0.382, false recommendations rise from 0.1930.193 to 0.4610.461, and consumer welfare falls by 6%.

Figure 2. Regime switch in the two-group example. Optimal threshold. Consumer welfare along each branch. Verification-cost scale cc (lower = more transparent). Consumer welfare WW. Persuade branch (shallow standard). Exploit branch (deep standard). Notes: p=0.7p = 0.7, s1=0.05s_1 = 0.05, s2=1s_2 = 1, τ=0.15\tau= 0.15, γ=0\gamma= 0. Left: optimal threshold against cc; solid segments are the global optimum, dashed segments the dominated local optimum, the dotted line the switch at c≈0.170c \approx0.170. Right: consumer welfare along each branch. Lower cc means cheaper verification.

Since Λ>0\Lambda> 0 on both branches, Proposition 5 does not directly apply to this change; the reversal comes from comparing the two regimes. Once transparency makes the sophisticated consumers verify regardless of the recommendation standard, only the naive consumers remain on the margin. Then the recommender stops maintaining a costly standard it no longer needs.

Value of Transparency for Welfare

For consumers who would choose to verify vv, transparency also has a more direct positive effect: it directly lowers the disutility of acquiring information. The previous discussion showed that the behavioral effect on the platform of making vv more cheaply visible could, in some cases, hurt the consumer; we show here that the effect of deteriorating standards can outweigh the consumer’s direct cost-savings.

Let

C(v^,c)=∫0qAκ(c,q) dqC(\hat{v},c)=\int_{0}^{q_A}\kappa(c,q)\,\mathrm{d}q

be expected verification cost per risky recommendation, integrating over the consumers who explore. Let us consider the case in which the safe product yields zero consumption utility, vS=0v_S = 0,6 so that consumer welfare is generated entirely by risky recommendations. Under a threshold v^\hat{v} such that all consumers either accept or verify the recommendation, ex ante consumer welfare is

W(v^,c)=∫0vˉvg(v) dv−S(a(v^))a(v^)L(v^)−C(v^,c)L(v^).(18)W(\hat{v},c)=\int_{0}^{\bar{v}}v g(v)\,\mathrm{d}v-S(a(\hat{v}))a(\hat{v})L(\hat{v})-C(\hat{v},c)L(\hat{v}). \tag*{(18)}

For recommended good products, all consumers adopt the risky product, but explorers pay verification costs. For recommended bad products, blind accepters adopt and receive negative value, while explorers pay verification costs and reject.

Define the cost-weighted average growth rate among verifiers to be

eˉ≡∫0qAe(c,q) κ(c,q) dq∫0qAκ(c,q) dq≥0;\bar{e}\equiv\frac{\int_{0}^{q_A}e(c,q)\,\kappa(c,q)\,\mathrm{d}q}{\int_{0}^{q_A}\kappa(c,q)\,\mathrm{d}q}\geq0;

a uniformly proportional technology has eˉ=1\bar{e}=1.

Lemma 6 (Envelope properties). For any verification technology,

∂W∂v^=g(v^)[S∣v^∣+C]>0,∂W∂c=−L(v^)Ceˉc.\frac{\partial W}{\partial\hat{v}}=g(\hat{v})\left[S|\hat{v}|+C\right]>0,\qquad\frac{\partial W}{\partial c}=-L(\hat{v})\frac{C\bar{e}}{c}.

Proof. See Appendix § 9.

Here, ∂W∂v^\frac{\partial W}{\partial\hat{v}} captures the impact of standards on consumer welfare, while ∂W∂c\frac{\partial W}{\partial c} is the direct effect. For the net impact on consumer welfare, we combine the direct effect with the effect on v^\hat{v}, scaled by its impact on welfare.

Proposition 7 (Opacity and consumer welfare). Let vS=0v_S=0 and let v^∗(c)\hat{v}^{*}(c) be the platform’s optimal threshold. Then

ddcW(v^∗(c),c)=g(v^∗)[S(a∗)∣v^∗∣+C]dv^∗dc⏟standards effect−L(v^∗)Ceˉc⏟direct effect.\frac{\mathrm{d}}{\mathrm{d}c}W(\hat{v}^{*}(c),c)=\underbrace{g(\hat{v}^{*})\left[S(a^{*})|\hat{v}^{*}|+C\right]\frac{\mathrm{d}\hat{v}^{*}}{\mathrm{d}c}}_{\textit{standards effect}}-\underbrace{L(\hat{v}^{*})\frac{C\bar{e}}{c}}_{\textit{direct effect}}.

Proof. By Lemma 6 and the chain rule.

Whenever standards respond positively to transparency, the two effects lie in the same direction, and transparency yields higher consumer welfare. However, when transparency makes the platform lower standards, the standards effect can dominate. In the economy of Example 3 (τ=0.15\tau=0.15, ν=0.3\nu=0.3), the following numerical example demonstrates:

cc0.040.050.06
v^∗(c)\hat{v}^{*}(c)−0.863-0.863−0.727-0.727−0.590-0.590
W(v^∗(c),c)W(\hat{v}^{*}(c),c)0.17920.17940.1823
W(v^∗(0.04),c)W(\hat{v}^{*}(0.04),c) (threshold held fixed)0.17920.17800.1769

Table 2.

Consumer welfare decreases with decreasing cost cc at the optimum: the platform’s endogenous lowering of the standard outweighs the effect of cheaper verification.7 By contrast, in the light-tailed economy of Example 1, Λ>0\Lambda>0, so welfare falls monotonically in cc.

Which Technologies Help Consumers? Outside the uniformly proportional case, Prop. 7 also gives some insight into which technologies are best. Clearly, if cc is allowed to go to 0, consumers will obtain their first-best welfare, even if false recommendations abound. So, let us consider instead when increasing opacity (a positive FOSD shift in cc) is not so bad, or even helpful.

Proposition 8 (Comparison.). Fix the initial cost distribution. A technology with profile e1e_1 leaves consumers better off after a marginal increase in cc than one with profile e2e_2 does if and only if

(eˉ1−eˉ2)LC+ψg(v^∗)[S(a∗)∣v^∗∣+C]η∣∂Φ∂v^∣{(e1∗−e2∗)[Λ+γηS]+(e1∗−e2∗)ψS}<0.(\bar{e}_1-\bar{e}_2)LC+\frac{\psi g(\hat{v}^{*})\left[S(a^{*})\lvert\hat{v}^{*}\rvert+C\right]}{\eta\left\lvert\frac{\partial\Phi}{\partial\hat{v}}\right\rvert}\left\{(e_1^{*}-e_2^{*})\left[\Lambda+\frac{\gamma\eta}{S}\right]+(e_1^{*}-e_2^{*})\psi S\right\}<0.

Proof. See Appendix B.

For comparable shifts, we can assume that e1∗=e2∗e_1^{*}=e_2^{*}. Note first that both eˉ\bar{e} and e∗e^{*} have a negative effect on consumer welfare, but there is a tension in minimizing both: when costs grow more quickly for lower-cost consumers than high-cost ones (e∗e^{*} decreases), then some verifiers must pay more (eˉ\bar{e} must be positive, assuming costs don’t decrease for anyone). A small increase in cc therefore benefits consumers the most if the increase in costs is sharply concentrated in near-marginal verifiers.

Recommendations without Commitment

In the commitment benchmark, the recommender chooses its policy before it observes quality. In this section, we consider the case when the recommender cannot commit to a policy. The recommender now observes vv first and then chooses which product to recommend. In equilibrium, consumers correctly anticipate the recommendation policy.

Equilibrium

The difference from the commitment case is that the recommender no longer accounts for how its recommendations affect consumers’ beliefs. Under commitment, a false recommendation raises the downside posterior and leads more consumers to verify, and the recommender weighs this cost when it chooses π\pi. Without commitment, the recommender decides at each vv and takes consumers’ beliefs as given. Therefore, the recommender’s payoff from recommending the risky product depends on vv only through its sign. A good product is adopted by all consumers, whereas a bad product is adopted only by those who do not verify, however bad it is.

Formally, a strategy for the recommender is a policy π\pi as in Section 3, where π(v)\pi(v) is now the probability of recommending the risky product after observing vv.

Consumers’ behavior is unchanged. They form the posterior β(⋅∣π)\beta(\cdot\mid\pi); when μ(π)≥0\mu(\pi) \ge0, a share S(c,a(π))S(c, a(\pi)) of them accepts without verifying.

An equilibrium is a policy π\pi such that, for every vv, π(v)\pi(v) maximizes the recommender’s expected payoff given consumers’ response to π\pi.8

In an equilibrium with μ(π)≥0\mu(\pi) \ge0, a false recommendation earns pRS(c,a(π))p_R S(c, a(\pi)), and the safe product earns pSp_S. The recommender is therefore indifferent between the two when the share of consumers who accept without verifying equals the normalized margin γ\gamma. Intuitively, a false recommendation earns pRp_R from each consumer in this share, and this must match the pSp_S that the safe product earns. Since consumers whose cost lies below the posterior downside verify, the indifference condition holds if and only if the posterior downside equals the verification cost of the consumer at quantile 1−γ1-\gamma. We denote this cost by

aγ(c)≡κ(c,1−γ).(19)a_\gamma(c) \equiv\kappa(c, 1-\gamma). \tag*{(19)}

It depends on γ\gamma and FcF_c, but not on the discution GG.

Lemma 9 (Structure of equilibria). In every equilibrium, consumers never reject, and every good product is recommended. Moreover, either every bad product is recommended (pooling), or

a(π)=aγ(c)(20)a(\pi) = a_\gamma(c) \tag*{(20)}

and the recommender earns the truthful profit Π(0)=pRH+pSG(0)\Pi(0) = p_R H + p_S G(0).

The indifference condition (20) pins down the posterior downside of the pool. However, it does not pin down the composition of the pool. As a result, Lemma 9 allows many equilibria.9 We focus on threshold policies, π=1v≥v^\pi= \mathbb{1}_{v\ge\hat{v}}, which have the same form as the commitment policy. Note that full pooling is the threshold v^=v\hat{v}=v. We write aˉ≡∫v‾0(−v)g(v) dv\bar{a} \equiv\int_{\underline{v}}^0(-v)g(v)\,dv for the posterior downside under full pooling.

From now on, we assume that the safe product is profitable for the recommender, γ>0\gamma> 0, and that the risky product is worth purchasing on average, E[v]≥0\mathbb{E}[v] \ge0. If γ=0\gamma= 0, a false recommendation is never worse than the safe product, so every bad product is recommended. The second condition ensures that every threshold policy, including full pooling, has a nonnegative posterior mean, so consumers never reject it. Under these two conditions, a threshold equilibrium exists for every cc.

Proposition 10 (Threshold equilibrium). There is exactly one threshold equilibrium. If S(c,aˉ)≥γS(c,\bar{a}) \ge\gamma, it is full pooling, that is, v^nc=v\hat{v}^{nc}=v. Otherwise, its threshold v^nc∈(v,0)\hat{v}^{nc}\in(v,0) solves a(v^nc)=aγ(c)a(\hat{v}^{nc})=a_\gamma(c).

The proofs for are in Appendix § 11. Proposition 10 says that there are two cases. Suppose that verification is so costly that many consumers accept without verifying even when every product is recommended (when S(c,aˉ)≥γS(c,\bar{a}) \ge\gamma). Then every false recommendation remains profitable, so verification does not discipline the recommender, and every product is recommended. Otherwise, verification disciplines the recommender exactly up to the point of indifference.

The threshold equilibrium is also the natural one to study. Suppose that a false recommendation carries a small cost that increases with how bad the product is, for example, because worse products generate more complaints, refunds, or legal liability. Such a cost break the recommender’s indifference condition in favor of the mildest bad products and therefore eliminates other equilibria.

Remark 3 (Equilibrium selection). Suppose that a false recommendation of quality vv costs the recommender εh(∣v∣)\varepsilon h(\lvert v\rvert), where hh is continuous and strictly increasing, with h(0)=0h(0)=0. Then, for every ε>0\varepsilon> 0, the equilibrium is unique and is a threshold policy. As ε→0\varepsilon\to0, it converges to the threshold equilibrium of Proposition 10.

In every non-pooling equilibrium, the recommender earns exactly the truthful profit Π(0)\Pi(0), whatever FcF_c and cc (Lemma 9). As long as more than a share γ\gamma of consumers accept without verifying, a false recommendation is more profitable than the safe product. The recommender therefore adds bad products until the share falls to γ\gamma, at which point each bad product earns pSpS whether or not it is recommended. Blind acceptance thus generates no rent for the recommender: the profit it could earn from consumers’ trust is fully dissipated.

Figure 3 illustrates this logic. Under a threshold policy v^<0\hat{v} < 0, the recommender’s profit in excess of the truthful profit is

Figure 3. Commitment versus no commitment. Notes: Exponential costs (Example 1) with γ=1/4\gamma= 1/4 and pR=1p_R = 1. Each curve plots the share of consumers who accept without verifying, S(c,a(v^))S(c, a(\hat{v})), against false recommendations B(v^)B(\hat{v}) as the threshold v^\hat{v} varies. The recommender’s profit in excess of the truthful profit equals the area B(S−γ)B(S-\gamma) of the rectangle under the curve and above the dashed line S=γS=\gamma. Panel (a), c=0.10c=0.10: the commitment optimum (circle) maximizes this area; the no-commitment equilibrium (square) lies at S=γS=\gamma, where the area is zero. Panel (b): a decrease in cc from 0.100.10 to 0.050.05 shifts the curve down; the no-commitment equilibrium moves along S=γS=\gamma toward fewer false recommendations. Lower cc means cheaper verification.

Π(v^)−Π(0)=pRB(v^)(S(c,a(v^))−γ),(21)\Pi(\hat{v})-\Pi(0)=p_R B(\hat{v})\bigl(S(c,a(\hat{v}))-\gamma\bigr), \tag*{(21)}

that is, pRp_R times the area of the rectangle under the curve in panel (a) and above the line S=γS=\gamma. Commitment chooses the point on the curve that maximizes this area. Without commitment, the equilibrium lies where the curve crosses S=γS=\gamma, so the area is zero.

Transparency without commitment

Under commitment, the recommender chooses the location of the marginal verifier. Its response to transparency therefore depends on the density of consumers near that margin. Without commitment, however, the marginal verifier is the consumer at quantile 1−γ1-\gamma, whose cost rises with cc at the rate

dln⁡aγ(c)dln⁡c=e(c,1−γ)>0.(22)\frac{\mathrm{d}\ln a_{\gamma}(c)}{\mathrm{d}\ln c}=e(c,1-\gamma)>0. \tag*{(22)}

The response of the equilibrium to cc thus depends only on how fast the cost of the consumer at quantile 1−γ1-\gamma grows, a rate that is positive for every technology. Neither Λ\Lambda nor e′∗e^{\prime*} matters. As a result, the shape of FcF_c affects the size of the response, but not its sign. Panel (b) of Figure 3 illustrates the consequence: a lower cc shifts the curve S(c,a(v^))S(c,a(\hat{v})) down, and the equilibrium moves along the fixed line S=γS=\gamma toward fewer false recommendations.

Proposition 11 (Transparency without commitment). The threshold v^nc(c)\hat{v}^{nc}(c) and consumer welfare W(v^nc(c),c)W(\hat{v}^{nc}(c),c) are both nonincreasing in cc. Moreover, both are strictly decreasing where the threshold equilibrium is not pooling.

Proposition 11 shows that, without commitment, cheaper verification never increases false recommendations and never lowers consumer welfare, whatever the shape of FcF_c.10 In the fat-tailed case of Example 3, the commitment optimum converges to recommending every product as c↓0c \downarrow0. Without commitment, however, aγ(c)=cφ(1−γ)→0a_\gamma(c) = c\varphi(1-\gamma) \to0 for any uniformly proportional family. As a result, the threshold equilibrium converges to recommending only good products. Therefore, the perverse effects of transparency identified in § 3 also depend on the recommender’s commitment power.

The value of commitment

Commitment is necessary for the perverse effects of transparency, but it nonetheless benefits both the recommender and consumers. To compare the two benchmarks, let ΠC=max⁡v^Π(v^)\Pi^{C} = \max_{\hat v}\Pi(\hat v) denote the commitment profit and Πnc=Π(v^nc)\Pi^{nc} = \Pi(\hat v^{nc}) the profit in the threshold equilibrium.

Proposition 12 (Value of commitment). Every commitment optimum satisfies v^≥v^nc\hat v \geq\hat v^{nc}, ΠC≥Πnc\Pi^{C} \geq\Pi^{nc}, and W(v^∗,c)≥W(v^nc,c)W(\hat v^{*},c) \geq W(\hat v^{nc},c). All three inequalities are strict where the threshold equilibrium is not pooling.

The intuition follows from (21). Commitment lets the recommender stop while the share of non-verifiers still exceeds γ\gamma. Without commitment, by contrast, that share falls to γ\gamma. As a result, the pool under commitment contains fewer bad products, so consumers also gain (Lemma 6). Commitment therefore benefits consumers in levels. However, it is also what makes their welfare respond ambiguously to transparency.

Discussion and Conclusion

Before concluding, we discuss the assumption that recommendations are binary.

General recommendations The restriction to a binary recommendation is motivated by what is observed in practice: products tend to be shown or not shown, or alternatively, recommended or not, but rarely given a complex rating. Indeed, in our main model, we assume that the risky product can only be accepted if it is recommended.

Nevertheless, the question of the optimal flexible recommendation policy is interesting. In contrast to the case of persuading a single consumer with known preferences, when consumers have heterogeneous attention costs, the restriction to binary recommendations is not without loss. When recommendations can involve as many messages as necessary, the form of the optimal policy involves a single non-recommendation message and, in many cases, multiple different recommendation messages. We know that such a policy can perform strictly better than a binary one because it better couples the proportion of accepting consumers (SS) to the proportion of recommendations under a message that are false (BB) than what can be done with a single threshold. Formalizing such a setup is beyond our current scope, but we conjecture that the optimal policy is one in which intervals of positive and negative qualities are paired “assortatively”: recommendation messages admit a ranking in which they pool progressively shrinking intervals of more deeply negative qualities with progressively expanding intervals of milder positive qualities.

Conclusion

This paper studies the content of recommendations issued by a self-interested intermediary when the consumers who receive them can verify quality at a private cost. The recommender is better informed, but may make false recommendations when some consumers are uninformed due to high attention costs. We have shown that an optimal binary recommendation can, perversely, respond negatively to easier verification by the consumer, via a tradeoff somewhat familiar from monopoly pricing: the consumer’s cost may indeed ”pass through negatively” to the recommender. Thus, starkly, recommendation quality may be worst precisely when consumers’ verification capabilities are strongest. Moreover, in some regimes, the effect of deteriorating recommendations can be enough to make consumers worse-off in the presence of better verification technology.

A practical takeaway is that transparency interventions have subtle effects: at the least, estimating their effect on recommendation quality requires knowing the shape of the attention-cost distribution near the verification cutoff, which may be nontrivial. A number of open questions remain. One, as slightly foreshadowed, is whether these results carry over to non-binary recommendation environments. Another interesting direction would be to consider the optimal design of verification technologies explicitly under practical constraints. Finally, our verification technology is stark in that it is perfectly revealing; although we reasonably believe the results carry through under other well-behaved cost and signal structures, this would be useful (and not immediate) to confirm.

Appendix

Proof of Lemma 1. First, consider the case in which E[v]>0\mathbb{E}[v] > 0. In this case, μ(π)>0\mu(\pi) > 0, and acceptance dominates rejection (but exploration is still possible), for any recommendation policy.

Next, consider the case where E[v]≤0\mathbb{E}[v] \le0. Define v0v_0 to be the unique threshold such that if the platform uses the recommendation policy π0\pi_0 that recommends the product if and only if v>v^0v > \hat{v}_0, then μ(π0)=0\mu(\pi_0) = 0.

Now consider an arbitrary policy π\pi. If μ(π)<0\mu(\pi) < 0, then all types of the consumer either reject the product or explore it. The platform’s payoff is upper-bounded by (1−G(0))pR+G(0)pS(1-G(0))p_R + G(0)p_S. The policy that simply recommends the product only when v>0v > 0 replicates this upper bound, and under it μ(π)>0\mu(\pi) > 0 so there is no rejection.

Finally, suppose there is a policy π\pi such that μ(π)=0\mu(\pi) = 0 and that, conditional on receiving the risky recommendation, the consumer rejects with probability p>0p > 0. We show that π\pi cannot be optimal.

Because μ(π)=0\mu(\pi) = 0,

Eβ[v+]=Eβ[(−v)+]=aR(π).\mathbb{E}_{\beta}[v^{+}] = \mathbb{E}_{\beta}[(-v)^{+}] = a_R(\pi).

Apart from the measure-zero marginal type, consumers with κ(c,q)<a\kappa(c,q) < a explore, while consumers with κ(c,q)>a\kappa(c,q) > a are indifferent between accepting and rejecting. Let

s=1−Fc(a)s = 1 - F_c(a)

denote the mass of consumers who do not explore. By assumption, a mass pp of consumers rejects and a mass s−ps-p accepts.

We construct a strictly more profitable policy by discarding a sufficiently small set of particularly bad false recommendations. Under the posterior induced by π\pi, there is a positive probability of recommended qualities satisfying

v<0and−v>a.v < 0 \qquad\text{and} \qquad-v > a.

Indeed, the recommended pool has positive probability of v>0v > 0, on which (−v)+=0<a(-v)^{+} = 0 < a, while the posterior average of (−v)+(-v)^{+} equals aa. Hence some recommended negative qualities must have downside strictly greater than aa.

Choose a measurable set EE of such qualities whose conditional probability under the risky recommendation is δ∈(0,p)\delta\in(0,p). Let

yˉ=Eβ[−v∣v∈E]>aR(π),\bar{y} = \mathbb{E}_{\beta}[-v \mid v \in E] > a_R(\pi),

and define π~\tilde{\pi} to be identical to π\pi, except that the risky product is no longer recommended when v∈Ev \in E. The original policy has total downside L(π)a(π)L(\pi)a(\pi). The modified policy therefore has posterior downside

a~=aR(π~)=a(π)−δyˉ1−δ<aR(π).\tilde{a} = a_R(\tilde{\pi}) = \frac{a(\pi)-\delta\bar{y}}{1-\delta} < a_R(\pi).

Moreover, because the original posterior mean is zero and the discarded qualities have conditional mean −yˉ-\bar{y}, the posterior mean under the modified policy is

μ(π~)=δyˉ1−δ>0.\mu(\tilde{\pi}) = \frac{\delta\bar{y}}{1-\delta} > 0.

Every consumer who rejected under π\pi had verification cost κ(c,q)≥a\kappa(c,q) \ge a. Under π~\tilde{\pi}, such a consumer strictly prefers blind acceptance to rejection because μ(π~)>0\mu(\tilde{\pi}) > 0, and strictly prefers blind acceptance to exploration because

UE−UA=a~−κ(c,q)<aR(π)−κ(c,q)≤0.U_E-U_A = \tilde{a}-\kappa(c,q) < a_R(\pi)-\kappa(c,q) \le0.

Thus, every consumer either accepts or verifies the new policy.

We have that D(π~)−D(π)≥p(1−δ)L(π)−(s−p)δL(π)=L(π)(p−sδ)D(\tilde{\pi})-D(\pi) \ge p(1-\delta)L(\pi)-(s-p)\delta L(\pi)=L(\pi)(p-s\delta), and moreover, the platform now receives pSp_S from the safe recommendations made on the discarded states, so

Π(π~)−Π(π)≥pRL(π)(p−sδ)+pSδL(π)>0.\Pi(\tilde{\pi})-\Pi(\pi) \ge p_RL(\pi)(p-s\delta)+p_S\delta L(\pi)>0.

Therefore, any policy with μ(π)=0\mu(\pi)=0 under which the consumer rejects with positive probability is strictly dominated; having ruled out all the cases above, we know that any optimal policy induces rejection with probability zero.

Lemma 13 (Derivative of posterior downside). For v^<0\hat{v}<0,

a′(v^)=−g(v^)1−G(v^)(∣v^∣−a(v^))<0.a'(\hat{v})=-\frac{g(\hat{v})}{1-G(\hat{v})}\left(\lvert\hat{v}\rvert-a(\hat{v})\right)<0.

Proof. We have a=A/La=A/L. Since A′(v^)=v^g(v^)A'(\hat{v})=\hat{v}g(\hat{v}) and L′(v^)=−g(v^)L'(\hat{v})=-g(\hat{v}),

a′(v^)=A′L−AL′L2=g(v^)(v^L+A)L2=g(v^)(v^+a)L=−g(v^)L(∣v^∣−a).a'(\hat{v})=\frac{A'L-AL'}{L^2}=\frac{g(\hat{v})(\hat{v}L+A)}{L^2}=\frac{g(\hat{v})(\hat{v}+a)}{L}=-\frac{g(\hat{v})}{L}\left(\lvert\hat{v}\rvert-a\right).

Because a(v^)a(\hat{v}) is an average downside over the pool while ∣v^∣\lvert\hat{v}\rvert is the largest downside included at the margin, ∣v^∣>a(v^)\lvert\hat{v}\rvert>a(\hat{v}) in the interior.

Proof of Lem. 4. Differentiating κ(c,Fc(a))=a\kappa(c,F_c(a))=a in cc at fixed aa gives ∂cFc(a)=−κc/κq\partial_cF_c(a)=-\kappa_c/\kappa_q, so ∂cS=κc(c,qA)fc(a)\partial_cS=\kappa_c(c,q_A)f_c(a); since κ(c,qA)=a\kappa(c,q_A)=a we have κc(c,qA)=e∗a/c\kappa_c(c,q_A)=e^{*}a/c and afc(a)=ψSaf_c(a)=\psi S, giving the first expression. For the second, fc(a)=1/κq(c,qA)f_c(a)=1/\kappa_q(c,q_A) and κqc=[e′κ+eκq]/c\kappa_{qc}=[e'\kappa+e\kappa_q]/c yield ∂cfc=(fc/c)[e∗(ψ−Λ)−e′∗ψS]\partial_cf_c=(f_c/c)[e^{*}(\psi-\Lambda)-e'^{*}\psi S], where we used (κqq/κq)ψS=1+ψ−Λ(\kappa_{qq}/\kappa_q)\psi S=1+\psi-\Lambda, itself obtained by differentiating (5) in ln⁡a\ln a. Substituting into ∂cψ=(a/S)[∂cfc−fc2κc/S]\partial_c\psi=(a/S)[\partial_cf_c-f_c^2\kappa_c/S] gives the result.

Proof of Lemma 6 and Proposition 7

Write Wc(v^,c)=∫v^vˉvg(v) dv−S(a)A(v^)−C(v^,c)L(v^)W^c(\hat{v},c)=\int_{\hat{v}}^{\bar{v}}vg(v)\,\mathrm{d}v-S(a)A(\hat{v})-C(\hat{v},c)L(\hat{v}), where A(v^)=∫v^0(−v)g(v) dvA(\hat{v})=\int_{\hat{v}}^0(-v)g(v)\,\mathrm{d}v is the total downside of the pool, L(v^)=1−G(v^)L(\hat{v})=1-G(\hat{v}) the frequency of risky recommendations, a=a(v^)=A/La=a(\hat{v})=A/L, and qA=Fc(a)q_A=F_c(a) so that κ(c,qA)=a\kappa(c,q_A)=a and C(v^,c)=∫0qAκ(c,q) dqC(\hat{v},c)=\int_0^{q_A}\kappa(c,q)\,\mathrm{d}q.

Effect of the threshold. Since ∂v^qA=fc(a)a′(v^)\partial_{\hat{v}}q_A=f_c(a)a'(\hat{v}), we have ∂v^S=−fc(a)a′(v^)\partial_{\hat{v}}S=-f_c(a)a'(\hat{v}) and ∂v^C=κ(c,qA)∂v^qA=afc(a)a′(v^)\partial_{\hat{v}}C=\kappa(c,q_A)\partial_{\hat{v}}q_A=a f_c(a)a'(\hat{v}). Differentiating in v^\hat{v},

∂Wc∂v^=−S(a)A′(v^)−CL′(v^)+fc(a)a′(v^)A−afc(a)a′(v^)L.\frac{\partial W^c}{\partial\hat{v}}=-S(a)A'(\hat{v})-C L'(\hat{v})+f_c(a)a'(\hat{v})A-a f_c(a)a'(\hat{v})L.

The last two terms equal fc(a)a′(v^)[A−aL]=0f_c(a)a'(\hat{v})[A-aL]=0, because aL=AaL=A by the definition of aa. This is the envelope property of the consumer’s exploration cutoff: the consumer at quantile qAq_A is exactly indifferent between exploring and accepting blindly, so the composition shift in who verifies has no first-order welfare effect. With A′(v^)=v^g(v^)A'(\hat{v})=\hat{v}g(\hat{v}) and L′(v^)=−g(v^)L'(\hat{v})=-g(\hat{v}),

∂Wc∂v^=g(v^)[S(a)∣v^∣+C]>0.\frac{\partial W^c}{\partial\hat{v}}=g(\hat{v})\left[S(a)|\hat{v}|+C\right]>0.

Effect of the cost. Differentiating in cc at fixed v^\hat{v}, we have ∂cS=κc(c,qA)fc(a)\partial_cS=\kappa_c(c,q_A)f_c(a) and

∂cC=∫0qA∂κ(c,q)∂c dq+κ(c,qA)∂cqA=K−aκc(c,qA)fc(a),\partial_cC=\int_0^{q_A}\frac{\partial\kappa(c,q)}{\partial c}\,\mathrm{d}q+\kappa(c,q_A)\partial_cq_A=\mathcal{K}-a\kappa_c(c,q_A)f_c(a),

using ∂cqA=∂cFc(a)=−κc/κq=−κc(c,qA)fc(a)\partial_cq_A=\partial_cF_c(a)=-\kappa_c/\kappa_q=-\kappa_c(c,q_A)f_c(a). The same cancellation applies: the two terms in κc(c,qA)fc(a)\kappa_c(c,q_A)f_c(a) cancel against A=aLA=aL, leaving only the direct increase in verification payments by the consumers who continue to verify,

∂Wc∂c=−L(v^)K=−L(v^)Ceˉc≤0.\frac{\partial W^c}{\partial c}=-L(\hat{v})\mathcal{K}=-\frac{L(\hat{v})C\bar{e}}{c}\leq0.

Combining. Under the regularity condition of Remark 2, v^∗(c)\hat{v}^*(c) is differentiable, and the chain rule gives the decomposition in Proposition 7. Substituting dv^∗/dc=(∂Φ/∂c)/∣∂Φ/∂v^∣\mathrm{d}\hat{v}^*/\mathrm{d}c=(\partial\Phi/\partial c)/|\partial\Phi/\partial\hat{v}|, with ∂Φ/∂c\partial\Phi/\partial c given by (14), and multiplying by c>0c>0 gives the condition in Proposition 8. If the bracket in (13) is nonnegative, then dv^∗/dc≤0\mathrm{d}\hat{v}^*/\mathrm{d}c\leq0, so both terms are nonpositive and consumer welfare weakly decreases in cc. A welfare increase therefore requires a negative bracket together with a standards effect large enough to outweigh the direct effect.

Details for the Examples

Example 1 (light-tailed types)

Quality side. With g=12g=\frac{1}{2}, G(0)=12G(0)=\frac{1}{2}, and 1−G(v^)=1+∣v^∣21-G(\hat{v})=\frac{1+|\hat{v}|}{2},

a(v^)=∫0∣v^∣t12 dt(1+∣v^∣)/2=∣v^∣22(1+∣v^∣),B(v^)=G(0)−G(v^)=∣v^∣2,η(v^)=1+∣v^∣2+∣v^∣.a(\hat{v})=\frac{\int_0^{|\hat{v}|}t\frac{1}{2}\,\mathrm{d}t}{(1+|\hat{v}|)/2}=\frac{|\hat{v}|^2}{2(1+|\hat{v}|)},\qquad B(\hat{v})=G(0)-G(\hat{v})=\frac{|\hat{v}|}{2},\qquad\eta(\hat{v})=\frac{1+|\hat{v}|}{2+|\hat{v}|}.

where the last expression follows from substituting a(v^)a(\hat{v}) into the definition of η\eta. Note μ(v^)=1−∣v^∣2≥0\mu(\hat{v}) = \frac{1-|\hat{v}|}{2} \ge0 for all ∣v^∣≤1|\hat{v}| \le1, so the entire range lies in the acceptance–verification region and v^0=v=−1\hat{v}_0 = v = -1.

Consumer side. With Fc(x)=1−e−x/cF_c(x) = 1 - e^{-x/c} we have κ(c,q)=−cln⁡(1−q)\kappa(c,q) = -c\ln(1-q), so the family is uniformly proportional, and ψ(a,c)=a/c\psi(a,c) = a/c, strictly increasing in aa, with Λ=1\Lambda= 1.

Uniqueness and interiority. The ratio σ(v^)=ψ(a(v^),c)/η(v^)=∣v^∣2(2+∣v^∣)2c(1+∣v^∣)2\sigma(\hat{v}) = \psi(a(\hat{v}),c)/\eta(\hat{v}) = \frac{|\hat{v}|^2(2+|\hat{v}|)}{2c(1+|\hat{v}|)^2} satisfies

dln⁡σd∣v^∣=2∣v^∣+12+∣v^∣−21+∣v^∣=4+3∣v^∣+∣v^∣2∣v^∣(1+∣v^∣)(2+∣v^∣)>0,\frac{\mathrm{d}\ln\sigma}{\mathrm{d}|\hat{v}|} = \frac{2}{|\hat{v}|} + \frac{1}{2+|\hat{v}|} - \frac{2}{1+|\hat{v}|} = \frac{4+3|\hat{v}|+|\hat{v}|^2}{|\hat{v}|(1+|\hat{v}|)(2+|\hat{v}|)} > 0,

so (16) has a unique root and the single-crossing condition of Remark 2 holds. The root is interior (∣v^∗∣<1|\hat{v}^*| < 1) if and only if σ(−1)>1\sigma(-1) > 1, that is, c<3/8c < 3/8. Implicit differentiation of (16) gives

d∣v^∗∣dc=2(1+∣v^∗∣)2(4∣v^∗∣+3∣v^∗∣2)−4c(1+∣v^∗)>0,\frac{\mathrm{d}|\hat{v}^*|}{\mathrm{d}c} = \frac{2(1+|\hat{v}^*|)^2}{(4|\hat{v}^*|+3|\hat{v}^*|^2)-4c(1+|\hat{v}^*)} > 0,

the denominator being positive by the same single-crossing property.

Numerical solutions.

cc0.050.100.150.20
v^∗\hat{v}^*−0.266-0.266−0.405-0.405−0.526-0.526−0.638-0.638
a∗a^*0.0280.0580.0910.124
false recs B(v^∗)B(\hat{v}^*)0.1330.2030.2630.319
demand D(v^∗)D(\hat{v}^*)0.5760.6130.6440.671

Table 3.

Each optimum is verified to dominate the corner policies v^=0\hat{v} = 0 (demand 12\frac{1}{2}) and v^=−1\hat{v} = -1.

Example 3 (heavy-tailed types)

The bad-side density is g(v)=κev/τg(v) = \kappa e^{v/\tau} on [−1,0)[-1,0) with κ=[2τ(1−e−1/τ)]−1\kappa= [2\tau(1-e^{-1/\tau})]^{-1}, so bad products have total mass 12\frac{1}{2} concentrated within depth τ\tau of zero; the good side is uniform, g=12g = \frac{1}{2} on [0,1][0,1]. Verification costs satisfy x≥cx \ge c, so on the region a(v^)<ca(\hat{v}) < c nobody verifies and demand is simply 1−G(v^)1-G(\hat{v}). The optima reported below are verified to lie in the interior verification region a(v^)>ca(\hat{v}) > c.

Quality side. Writing χ(v^)=e−∣v^∣/τ\chi(\hat{v}) = e^{-|\hat{v}|/\tau} and N=1−e−1/τN = 1-e^{-1/\tau},

B(v^)=1−χ(v^)2N,a(v^)=τ[1−(1+∣v^∣τ)χ(v^)]N[1+2B(v^)],η(v^)=τ[1−(1+∣v^∣τ)χ(v^)]2NB(v^)(∣v^∣−a(v^)).B(\hat{v}) = \frac{1-\chi(\hat{v})}{2N}, \qquad a(\hat{v}) = \frac{\tau\left[1-\left(1+\frac{|\hat{v}|}{\tau}\right)\chi(\hat{v})\right]}{N[1+2B(\hat{v})]}, \qquad \eta(\hat{v}) = \frac{\tau\left[1-\left(1+\frac{|\hat{v}|}{\tau}\right)\chi(\hat{v})\right]}{2NB(\hat{v})(|\hat{v}|-a(\hat{v}))}.

For ∣v^∣|\hat{v}| beyond a few multiples of τ\tau the exponential terms vanish: a(v^)→τ/2a(\hat{v}) \to\tau/2 and η(v^)≈τ/(∣v^∣−τ/2)\eta(\hat{v}) \approx\tau/(|\hat{v}|-\tau/2), strictly decreasing in ∣v^∣|\hat{v}|. The economics: the average downside saturates at the tail mean τ/2\tau/2, while the marginal bad product at depth ∣v^∣|\hat{v}| keeps deteriorating, so the pool’s tolerance for dilution falls as the standard deepens. The posterior mean satisfies μ(v^)≥0.17\mu(\hat{v}) \ge0.17 for all ∣v^∣≤1|\hat{v}| \le1, so the acceptance–verification region is again the entire range.

Uniqueness. On the verification region the first-order condition is

ν1+ln⁡(a(v^)/c)=η(v^).\frac{\nu}{1+\ln(a(\hat{v})/c)}=\eta(\hat{v}).

Because a(v^)a(\hat{v}) saturates, the left side is nearly flat in ∣v^∣|\hat{v}| (its slope is exponentially small in ∣v^∣/τ|\hat{v}|/\tau), while the right side is strictly decreasing; hence Δ(v^)=ψ−η\Delta(\hat{v})=\psi-\eta is strictly increasing where it crosses zero, which is the second-order condition for a maximum, and the crossing is unique. Substituting the saturated values gives the approximate closed form reported in the text.

Numerical solutions (τ=0.15\tau=0.15, ν=0.3\nu=0.3):

cc0.040.050.06
v^∗\hat{v}^{*}−0.864-0.864−0.727-0.727−0.590-0.590
a∗a^{*}0.0740.0720.069
false recs B(v^∗)B(\hat{v}^{*})0.4990.4970.491
demand D(v^∗)D(\hat{v}^{*})0.9330.9530.973

Table 4.

Each a∗>ca^{*}>c, so Λ=−1/(1+ln⁡(a∗/c))<0\Lambda=-1/(1+\ln(a^{*}/c))<0 at every optimum, and each solution is verified to be the global maximum (against the corners and against a fine grid over v^\hat{v}). The false-recommendation mass falls only slightly because the bad tail is thin beyond depth τ\tau; the economically meaningful movement is in the marginal quality admitted.

Proofs for Section 5

Proof of Lemma 9. Fix an equilibrium π\pi with L(π)>0L(\pi)>0.

*Step 1: μ(π)≥0.∗Supposeinsteadthat\mu(\pi)\ge0.* Suppose instead that\mu(\pi)<0. Then consumers who do not verify reject the risky recommendation. Hence a recommended bad product is adopted by nobody, and recommending it yields $0<p_S. Therefore, no bad product is recommended, which implies that μ(π)>0\mu(\pi)>0, a contradiction. Since μ(π)≥0\mu(\pi)\ge0, no consumer rejects.

Step 2: good products. Every consumer adopts a recommended good product. Verifiers adopt it because v≥0v\ge0, and the others adopt it blindly. Recommending a good product therefore yields pR>pSp_R>p_S, so π(v)=1\pi(v)=1 for all v≥0v\ge0.

Step 3: bad products. Recommending a bad product yields pRS(c,a(π))p_R S(c,a(\pi)), which does not depend on vv. If S(c,a(π))>γS(c,a(\pi))>\gamma, then π(v)=1\pi(v)=1 for almost every v<0v<0, so π≡1\pi\equiv1. If S(c,a(π))<γS(c,a(\pi))<\gamma, then π(v)=0\pi(v)=0 for almost every v<0v<0. In that case, a(π)=0a(\pi)=0. However, S(c,0)=1>γS(c,0)=1>\gamma because FcF_c is atomless with Fc(0)=0F_c(0)=0, a contradiction. Otherwise, S(c,a(π))=γS(c,a(\pi))=\gamma, which is (20). Step 4: profit. In a non-pooling equilibrium, the recommender is indifferent at every bad vv. Hence each bad state yields pSp_S in expectation, whatever the realized recommendation. Each good state yields pRp_R. Summing over states gives Π(0)=pRH+pSG(0)\Pi(0) = p_R H + p_S G(0).

Proof of Proposition 10. Under π≡1\pi\equiv1, the posterior is the prior, so μ=E[v]≥0\mu= \mathbb{E}[v] \ge0 and a=aˉa = \bar{a}. The recommender does not deviate at a bad vv if and only if pRS(c,aˉ)≥pSp_R S(c,\bar{a}) \ge p_S, that is, S(c,aˉ)≥γS(c,\bar{a}) \ge\gamma.

Next, consider a threshold v^∈(v‾,0]\hat{v} \in(\underline{v},0]. It truncates the prior from below, so μ(v^)≥E[v]≥0\mu(\hat{v}) \ge\mathbb{E}[v] \ge0. By Lemma 9, it is an equilibrium if and only if a(v^)=aγ(c)a(\hat{v}) = a_\gamma(c). Indeed, indifference makes 1v≥v^\mathbb{1}_{v \ge\hat{v}} optimal at every vv. By Lemma 13, a(⋅)a(\cdot) is continuous and strictly decreasing on [v‾,0][\underline{v},0], from aˉ\bar{a} to 0. Since aγ(c)>0a_\gamma(c) > 0, a solution in (v‾,0)(\underline{v},0) therefore exists if and only if aγ(c)<aˉa_\gamma(c) < \bar{a}. When it exists, it is unique.

Finally, aγ(c)<aˉa_\gamma(c) < \bar{a} if and only if S(c,aˉ)<γS(c,\bar{a}) < \gamma. This is because S(c,⋅)S(c,\cdot) is nonincreasing and S(c,aγ(c))=γS(c,a_\gamma(c)) = \gamma. Moreover, S(c,⋅)S(c,\cdot) is strictly decreasing to the right of aγ(c)a_\gamma(c), which lies in the interior of the support. Hence exactly one of the two threshold equilibria exists.

Proof of Remark 3. With the cost εh(∣v∣)\varepsilon h(|v|), recommending the risky product at v<0v < 0 yields pRS(c,a(π))−εh(∣v∣)p_R S(c,a(\pi)) - \varepsilon h(|v|). This payoff is strictly decreasing in ∣v∣|v|. Good products still yield pR>pSp_R > p_S, and Step 1 of the proof of Lemma 9 applies unchanged. Hence every equilibrium is, up to a null set, a threshold policy. Define

Dε(v^)≡S(c,a(v^))−γ−εpRh(∣v^∣).D_\varepsilon(\hat{v}) \equiv S(c,a(\hat{v})) - \gamma- \frac{\varepsilon}{p_R}h(|\hat{v}|).

A threshold v^∈(v‾,0)\hat{v} \in(\underline{v},0) is an equilibrium if and only if Dε(v^)=0D_\varepsilon(\hat{v}) = 0. The threshold v‾\underline{v} is an equilibrium if and only if Dε(v‾)≥0D_\varepsilon(\underline{v}) \ge0. Note that DεD_\varepsilon is continuous and strictly increasing in v^\hat{v}, because S(c,a(v^))S(c,a(\hat{v})) is nondecreasing and h(∣v^∣)h(|\hat{v}|) is strictly decreasing in v^\hat{v}. Moreover, Dε(v^)→1−γ>0D_\varepsilon(\hat{v}) \to1-\gamma> 0 as v^↑0\hat{v} \uparrow0. Therefore, there is exactly one equilibrium. It is v‾\underline{v} if Dε(v‾)≥0D_\varepsilon(\underline{v}) \ge0, and the unique zero of DεD_\varepsilon otherwise.

As ε→0\varepsilon\to0, Dε→D0D_\varepsilon\to D_0 uniformly on [v‾,0][\underline{v},0]. Moreover, D0D_0 is strictly increasing near its zero v^nc\hat{v}^{nc}, because aγ(c)a_\gamma(c) lies in the interior of the support. Hence the equilibrium threshold converges to v^nc\hat{v}^{nc}. If D0(v‾)>0D_0(\underline{v}) > 0, the equilibrium is full pooling for all small ε\varepsilon.

Proof of Proposition 11. First, suppose that the threshold equilibrium is not pooling. Then a(v^nc)=aγ(c)=κ(c,1−γ)a(\hat{v}^{nc}) = a_\gamma(c) = \kappa(c,1-\gamma), and (11) gives (22). This elasticity is strictly positive because ∂Fc/∂c<0\partial F_c/\partial c < 0 on the interior of the support. Since a(⋅)a(\cdot) is strictly decreasing (Lemma 13), it follows that

dv^ncdc=−e(c,1−γ)aγ(c)L(v^nc)c g(v^nc)(∣v^nc∣−aγ(c))<0.\frac{\mathrm{d}\hat{v}^{nc}}{\mathrm{d}c} = -\frac{e(c,1-\gamma)a_\gamma(c)L(\hat{v}^{nc})} {c\,g(\hat{v}^{nc})(|\hat{v}^{nc}|-a_\gamma(c))} < 0.

Next, S(c,aˉ)S(c,\bar{a}) is nondecreasing in cc by Assumption 2. Hence the values of cc with S(c,aˉ)≥γS(c,\bar{a}) \ge\gamma form an upper interval, on which the equilibrium is pooling and v^nc=v‾\hat{v}^{nc} = \underline{v}. At the lower end of this interval, aγ(c)=aˉa_{\gamma}(c)=\bar{a}, so v^nc\hat{v}^{nc} is continuous in cc. Therefore, v^nc\hat{v}^{nc} is nonincreasing everywhere.

For welfare, the chain rule gives ddcW(v^nc(c),c)=∂v^W⋅dv^nc/dc+∂cW\frac{\mathrm{d}}{\mathrm{d}c}W(\hat{v}^{nc}(c),c)=\partial_{\hat{v}}W\cdot\mathrm{d}\hat{v}^{nc}/\mathrm{d}c+\partial_{c}W. By Lemma 6, ∂v^W>0\partial_{\hat{v}}W>0 and ∂cW≤0\partial_{c}W\leq0. Both terms are therefore nonpositive. Moreover, the first term is strictly negative where the equilibrium is not pooling.

Proof of Proposition 12. Suppose first that the threshold equilibrium is not pooling, so that Πnc=Π(0)\Pi^{nc}=\Pi(0) by Lemma 9. As v^↑0\hat{v}\uparrow0, a(v^)→0a(\hat{v})\to0 and S(c,a(v^))→1>γS(c,a(\hat{v}))\to1>\gamma. Moreover, B(v^)>0B(\hat{v})>0 for every v^<0\hat{v}<0. By (21), it follows that Π(v^)>Π(0)\Pi(\hat{v})>\Pi(0) for v^\hat{v} close to zero, so ΠC>Πnc\Pi^{C}>\Pi^{nc}. Since Π(v^∗)>Π(0)\Pi(\hat{v}^{*})>\Pi(0), (21) implies that B(v^∗)[S(c,a∗)−γ]>0B(\hat{v}^{*})[S(c,a^{*})-\gamma]>0, so that S(c,a∗)>γ=S(c,aγ(c))S(c,a^{*})>\gamma=S(c,a_{\gamma}(c)). Because S(c,⋅)S(c,\cdot) is nonincreasing, a∗<aγ(c)=a(v^nc)a^{*}<a_{\gamma}(c)=a(\hat{v}^{nc}). Because a(⋅)a(\cdot) is strictly decreasing, v^∗>v^nc\hat{v}^{*}>\hat{v}^{nc}.

Suppose instead that the threshold equilibrium is pooling. Then v^nc=v≤v^∗\hat{v}^{nc}=v\leq\hat{v}^{*}. Moreover, ΠC≥Π(v)=Πnc\Pi^{C}\geq\Pi(v)=\Pi^{nc}, because full pooling is feasible under commitment.

In both cases, W(v^∗,c)≥W(v^nc,c)W(\hat{v}^{*},c)\geq W(\hat{v}^{nc},c), because W(⋅,c)W(\cdot,c) is strictly increasing (Lemma 6). The inequality is strict when v^∗>v^nc\hat{v}^{*}>\hat{v}^{nc}.

References

  1. [1]Simon P. Anderson and Régis Renault. Advertising content. American Economic Review, 96(1):93–113, 2006.
  2. [2]Mark Armstrong and Yongmin Chen. Inattentive consumers and product quality. Journal of the European Economic Association, 7(2–3):411–422, 2009.DOI
  3. [4]Jeremy I. Bulow and Paul Pfleiderer. A note on the effect of cost changes on prices. Journal of Political Economy, 91(1):182–185, 1983.DOI
  4. [5]Laura Doval. Whether or not to open pandora’s box. Journal of Economic Theory, 175:127–158, 2018. ISSN 0022-0531. doi: https://doi.org/10.1016/j.jet.2018.01.005. URL https://www.sciencedirect.com/science/article/pii/S0022053118300073.
  5. [6]Glenn Ellison and Sara Fisher Ellison. Search, obfuscation, and price elasticities on the internet. Econometrica, 77(2):427–452, 2009.
  6. [8]Xavier Gabaix and David Laibson. Shrouded attributes, consumer myopia, and information suppression in competitive markets. Quarterly Journal of Economics, 121 (2):505–540, 2006.
  7. [10]Justin P. Johnson and David P. Myatt. On the simple economics of advertising, marketing, and product design. American Economic Review, 96(3):756–784, June 2006. doi: 10.1257/aer.96.3.756. URL https://www.aeaweb.org/articles?id=10.1257/aer.96.3.756.DOI
  8. [11]Emir Kamenica and Matthew Gentzkow. Bayesian persuasion. American Economic Review, 101(6):2590–2615, 2011.
  9. [12]Martin A. Lariviere. A note on probability distributions with increasing generalized failure rates. Operations Research, 54(3):602–604, 2006.DOI
  10. [13]Martin A. Lariviere and Evan L. Porteus. Selling to the newsvendor: An analysis of price-only contracts. Manufacturing and Service Operations Management, 3(4):293–305, 2001.DOI
  11. [14]Tracy R. Lewis and David E. M. Sappington. Supplying information to facilitate price discrimination. International Economic Review, 35(2):309–327, 1994.DOI
  12. [16]Georgy Lukyanov and Samuel Safaryan. Public persuasion with endogenous fact-checking. Technical report, arXiv:2508.19682, 2025. Working paper.
  13. [17]Ludmila Matysková and Alfonso Montes. Bayesian persuasion with costly information acquisition. Journal of Economic Theory, 2023.
  14. [18]Monika Mrázová and J. Peter Neary. Not so demanding: Demand structure and firm behavior. American Economic Review, 107(12):3835–3874, 2017.DOI
  15. [19]Martin Peitz and Anton Sobolev. Biased recommendations and differentially informed consumers. Economic Theory Bulletin, 13:245–261, 2025.DOI
  16. [20]Luciano Pomatto, Philipp Strack, and Omer Tamuz. The cost of information: The case of constant marginal costs. American Economic Review, 113(5):1360–93, May 2023. doi: 10.1257/aer.20190185. URL https://www.aeaweb.org/articles?id=10.1257/aer.20190185.DOI
  17. [21]Anne-Katrin Roesler and Balázs Szentes. Buyer-optimal learning and monopoly pricing. American Economic Review, 107(7):2072–80, July 2017. doi: 10.1257/aer.20160145. URL https://www.aeaweb.org/articles?id=10.1257/aer.20160145.DOI
  18. [22]Steven Salop and Joseph Stiglitz. Bargains and ripoffs: A model of monopolistically competitive price dispersion. Review of Economic Studies, 44(3):493–510, 1977.DOI
  19. [23]Hal R. Varian. A model of sales. American Economic Review, 70(4):651–659, 1980.
  20. [24]Martin L. Weitzman. Optimal search for the best alternative. Econometrica, 47(3):641–654, 1979.
  21. [25]E. Glen Weyl and Michal Fabinger. Pass-through as an economic tool: Principles of incidence under imperfect competition. Journal of Political Economy, 121(3):528–583, 2013.DOI

Paper details

vlnc:W_xcgdgwb5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents