Bayesian Persuasion and the Economics of Disclosure: Reviews, Ratings, and Price Transparency

TL;DR: The clean theory of voluntary disclosure predicts that markets tell you everything. If a seller can prove a claim at no cost and chooses not to, a skeptical buyer should assume the worst, so every seller above the floor discloses and quality unravels into full transparency. Grossman (1981) and Milgrom (1981) proved it. In practice it does not happen. In a controlled experiment, Jin, Luca and Martin (2021) found that senders holding a middling draw disclosed it only about 44.5 percent of the time against a theoretical 100 percent, and profited by it, because receivers were not skeptical enough of silence. The gap between theory and behavior is where product design lives. A star average, the default sort on a review list, the choice between a mean and a distribution, and the moment a mandatory fee appears are all decisions about what to reveal. Kamenica and Gentzkow (2011) give the exact limit of what an honest sender can win by choosing the information structure, and the drip-pricing evidence shows how large the prize is: hiding fees until checkout raised StubHub revenue roughly 21 percent (Blake, Moshary, Sweeney and Tadelis, 2021). Regulators in the United States, California, the United Kingdom and the European Union moved against exactly this between 2022 and 2025. An operator should run a disclosure audit, measure post-purchase regret alongside conversion, and treat shrouding as a loan against future trust.

---

The Number You See Is a Decision Someone Made

Open any ticket page, hotel result, or software pricing table and count the choices that were made on your behalf before you read a single figure. A rating is shown as one averaged number rather than a distribution. Reviews arrive in an order, and the order was chosen. A fee is either inside the headline price or waiting three screens later. A comparison table lists the attributes on which this product wins and quietly omits the ones on which it loses. None of these is a lie, yet every one of them is a decision about what to reveal, and each decision moves what you will pay and whether you will buy.

Economists have a precise theory of that decision, and it starts from an uncomfortable prediction: in a world of rational buyers, this kind of selective revelation should not work at all. If a seller can costlessly prove a favorable fact and stays silent, a sharp buyer infers the fact is unfavorable and marks the seller down accordingly. The threat of that inference should force disclosure of everything, all the way down to the worst type still willing to trade. Markets, in the clean model, unravel into full transparency without any regulator lifting a finger.

The prediction is elegant and, on its own terms, proven. In practice I find it mostly wrong, and the reason is behavioral rather than logical. Buyers do not punish silence hard enough, and that single behavioral fact is the hinge of this essay: it converts what should be a solved problem into a permanent design surface. Wherever buyers under-read silence, a seller who chooses what to show can extract value from the shortfall, and a platform that chooses the defaults can either widen the gap or close it.

My argument runs in four moves. First, the unraveling result and why it is airtight on its own terms. Second, the experimental and field evidence that unraveling stalls, and the exact reason it stalls. Third, Bayesian persuasion as the theory that tells you the most an honest sender can win from a receiver's credulity, and where mandatory disclosure earns its keep. Fourth, the money side: shrouding, drip pricing, reviews as engineered information, and the wave of 2022 to 2025 regulation that treats fee timing as a consumer-protection question. I close with what an operator should build and measure, and with the harder question of when partial disclosure is legitimate rather than manipulative.

Unraveling: Why Silence Should Be Damning

Start with the mechanism, because the intuition is the whole thing. Suppose a seller privately knows a quality θ drawn from some range, can prove it costlessly if she chooses, and cannot lie. Buyers know only the distribution of θ, not the realization. A seller with a top-of-range product has every reason to prove it. Once she does, the pool of undisclosed sellers no longer includes the best type, so the average of "the ones staying quiet" falls. Now the best of the remaining silent sellers faces the same logic and discloses. The pool average falls again. The reasoning cascades until only the single worst type is left silent, and even a buyer who cannot see that type's quality knows exactly what silence means.

The formal version is a fixed point. Let μ* be the highest quality that stays silent in equilibrium. A rational buyer facing non-disclosure must believe the product is worth the average of everything in the silent pool:

$$\mu^{*} \;=\; \mathbb{E}\!\left[\theta \mid \text{no disclosure}\right] \;=\; \frac{\displaystyle\int_{\underline{\theta}}^{\mu^{*}} \theta \, dF(\theta)}{F(\mu^{*})}.$$

The right-hand side is the mean of all types at or below μ*, and that mean is strictly less than μ* itself whenever the silent pool has any spread. A type sitting exactly at μ* is therefore being underpriced by staying silent, so it discloses, which lowers the pool and repeats the squeeze. The only self-consistent stopping point is the floor underlineθ. Grossman (1981, Journal of Law and Economics 24(3), 461-483) and Milgrom (1981, Bell Journal of Economics 12(2), 380-391) established this independently; Milgrom's "good news, bad news" framing is the one most product people have absorbed without knowing its source, because it is exactly the logic behind "if it were good, they would have told you."

Loading diagram...

Two assumptions carry the whole result, and naming them tells you exactly where it will break. Disclosure must be verifiable, so the seller cannot simply assert a good number, and it must be costless, so a good type is never deterred from proving itself by the price of proof. Relax verifiability and you are in the world of cheap talk, where words move nothing unless interests align. Add a disclosure cost and the lowest disclosing types drop out because proving a mediocre fact is not worth the fee, which reopens a silent pool with real spread. Milgrom and Roberts (1986, Journal of Political Economy 94(4), 796-821) worked through how competition among informed parties tightens the result, and how it loosens when proof is expensive.

The unraveling logic is not a fragile curiosity. The result is the reason nutrition panels, EPA mileage figures, and audited financials feel natural: once a few disclose, the rest are cornered. The interesting question for a product builder is not whether the logic holds. The logic holds. The real question is why, in the settings we actually operate, so much stays hidden and buyers let it.

Why Unraveling Stalls in the Wild

The cleanest test of unraveling is not a market, where a dozen forces tangle together, but a lab where the disclosure game is played bare. Jin, Luca and Martin (2021, American Economic Journal: Microeconomics 13(2), 141-173) built exactly that. A sender draws a whole number from 1 to 5, all equally likely, and decides only whether to reveal it. She cannot misreport. A receiver then guesses the number, the sender is paid more when the receiver guesses high, and the receiver is paid for accuracy. Unraveling makes a razor-sharp prediction here: every sender with a draw of 2 or more should reveal, and a receiver seeing silence should conclude the draw is 1.

The 324 subjects behaved differently. Senders with a 4 or a 5 disclosed roughly 95 percent of the time, close enough to the theory. But disclosure fell off a cliff below that. A draw of 3 was disclosed about 82 percent of the time, a draw of 2 only 44.5 percent, and a draw of 1 about 11 percent. Against a benchmark of 100 percent disclosure for every draw above the floor, senders were withholding middling news wholesale, and, crucially, they were right to. Withholding a 2 paid, because receivers seeing silence guessed a number well above the true average of the silent pool. The variance of the withheld state was almost a full unit, so silence was genuinely uninformative, and receivers treated it more kindly than the arithmetic warranted.

I keep returning to this study because the design point that matters most for practitioners is what did and did not close the gap. The authors varied feedback. When senders and receivers got no round-by-round feedback about the true state behind silence, behavior did not converge toward unraveling across 45 rounds. When receivers were debriefed after every round, told the true number and whether it had been hidden, disclosure of a 2 rose from about 42 percent to about 50 percent and receiver over-guessing shrank. Aggregate summaries handed out near the end, the kind of disclosure statistics a regulator or a press release provides, changed almost nothing. Skepticism is learned from direct, immediate, repeated experience of being fooled, not from being told in general that fooling occurs.

The field agrees with the lab. Mathios (2000, Journal of Law and Economics 43(2), 651-678) studied salad dressing before and after the Nutrition Labeling and Education Act made fat disclosure mandatory. Under the voluntary regime, every low-fat dressing carried a label and the higher-fat products mostly did not, exactly the top-down peeling unraveling predicts, but the peeling stopped short: wide variation in fat remained hidden among the silent, and when disclosure became mandatory, sales of the highest-fat dressings fell significantly. Partial unraveling had left real information on the table, and the mandate collected it. The behavioral gap is not a laboratory artifact; it survives into a real category with real labels and real money.

Bayesian Persuasion: The Exact Limit of Honest Spin

Once you accept that receivers are not perfectly skeptical, the interesting question flips. Instead of asking why sellers fail to disclose, ask how much a sender can gain by choosing the information structure in advance, still without ever lying. Kamenica and Gentzkow (2011, American Economic Review 101(6), 2590-2615) answered this exactly, and their answer is the analytical core of everything that follows.

The setup separates two ideas everyday language runs together. The sender cannot misrepresent the state, and the receiver is fully rational and knows the sender's incentive. Yet the sender still gains, because she commits to a signal, a rule mapping states to messages, before the state is drawn. The canonical example is a prosecutor facing a judge. Suppose the prior probability that a defendant is guilty is 0.3, and the judge convicts only when her posterior belief in guilt is at least 0.5. With no investigation, the judge sees a 0.3 and acquits every time, so the prosecutor's conviction rate is zero. The prosecutor cannot fabricate evidence, but she can design an investigation whose outcomes she commits to report in full.

The trick is to pool the guilty with just enough of the innocent. Design a signal that says "convict" for every guilty defendant and for some fraction of the innocent, calibrated so that a "convict" message leaves the judge with a posterior of exactly 0.5, her indifference point, at which she is willing to convict. The math of that calibration is forced by the requirement that beliefs average back to the prior. Formally, the sender chooses a distribution τ over posteriors, subject to the beliefs being Bayes-consistent, and her value is the concave closure of her payoff as a function of the posterior:

$$V(\mu_0)=\max_{\tau}\ \mathbb{E}_{\tau}\!\left[\hat v(\mu)\right]\ \ \text{s.t.}\ \ \mathbb{E}_{\tau}[\mu]=\mu_0, \qquad V(\mu_0)=\operatorname{cav}\,\hat v(\mu_0).$$

For the prosecutor, the payoff is a step function, v̂(μ)=1μ ≥ 0.5, worth 1 when the judge convicts and 0 otherwise. The concave closure of that step evaluated at the prior μ₀ = 0.3 equals 0.6. The prosecutor cannot lie, faces a rational judge, and still moves the conviction rate from zero to 60 percent purely by choosing what to reveal and committing to it. The concavification is not decoration; it is the exact ceiling. No honest information structure does better, and the optimum always induces the coarsest beliefs that just clear the receiver's action threshold.

Rayo and Segal (2010, Journal of Political Economy 118(5), 949-987) reached the same frontier from the seller's angle, characterizing the optimal way to bundle and reveal product attributes when the sender wants to steer choice. Their result is the theoretical license for the comparison table that shows the three specs where you win and buries the two where you lose: partial, truthful, structured disclosure that is optimal for the sender and fully anticipated by the theory. The persuasion literature does not moralize about this; it measures it, and the measurement tells you the exact size of the surplus a clever information design can capture from a rational counterparty, before we even add the behavioral credulity the experiments found.

An honest sender who chooses the information structure in advance can move a rational receiver as far as the concave closure of her own payoff allows, and no further.

, paraphrasing Kamenica and Gentzkow (2011)

Put the two results together and we have the operator's map. The concavification tells you the ceiling against a perfectly rational buyer. The disclosure experiments tell you how far above that ceiling a real buyer's credulity lets you reach, and for how long, before feedback erodes it. The distance between the two is the difference between persuasion that survives scrutiny and shrouding that is living on borrowed time.

Mandatory Disclosure When the Market Will Not Self-Reveal

If unraveling stalls, the policy question is whether forcing disclosure changes real behavior, or whether it merely papers a wall nobody reads. The single most convincing natural experiment is Los Angeles County's 1998 rule requiring restaurants to post a hygiene grade card, an A, B, or C, in the window. Jin and Leslie (2003, Quarterly Journal of Economics 118(2), 409-451) tracked what happened when a previously buried inspection score became a large letter at the door.

Three things moved together. Inspection scores rose as restaurants competed for the visible grade. Consumer demand became sensitive to hygiene, which it had not been when scores were filed away at the health department. And hospitalizations for foodborne illness in Los Angeles fell by about 20 percent, a discrete drop timed to the grade cards and significant at 99 percent confidence. The revenue mechanism is stark in the numbers: after grade cards, an A restaurant earned about 5.7 percent more revenue than before, a B about 0.7 percent more, and a C about 1 percent less. Before the cards, variation in the underlying, unposted scores had no measurable effect on revenue at all, because consumers had no cheap way to read it.

In my reading the lesson is not "disclosure is good." The real lesson is more specific and more useful. Forcing a fact into a cheap, standardized, at-the-point-of-decision format did what the voluntary market had failed to do, because the format collapsed the buyer's inference cost to near zero. Unlike a score filed in a database, a grade in the window is read. Dranove and Jin (2010, Journal of Economic Literature 48(4), 935-963) survey the wider record of quality disclosure and certification and find the same conditional pattern: mandated disclosure works when the measure is salient, comparable, and hard to game, and backfires or does nothing when it is coarse, noisy, or easily manipulated on the reported dimension while quality drifts on the unreported ones. In practice the design of the disclosure, not the mere fact of it, decides the outcome.

Shrouding and Drip Pricing: The Money in the Gap

Turn the same theory toward price and the stakes become directly financial. A price is not one number; it is a base plus a set of add-ons, fees, shipping, and taxes, and the seller chooses which parts are salient at the moment of decision. Gabaix and Laibson (2006, Quarterly Journal of Economics 121(2), 505-540) formalized this as shrouded attributes: in a market with some inattentive consumers, competition does not clean up hidden fees the way unraveling promises, because a firm that unshrouds a rival's add-on mostly educates customers who then substitute to the base good rather than switching firms. Shrouding can be a stable equilibrium, sustained by competition rather than despite it.

The behavioral engine underneath is limited attention to non-salient charges. Write the price a buyer actually responds to as a base p plus an attention-weighted fee:

$$\hat p \;=\; p + \theta f, \qquad \theta \in [0,1],$$

where f is the mandatory fee and θ is how much of it the buyer registers at decision time. A fully salient fee has θ = 1 and the buyer responds to the true total; a shrouded fee drives θ toward 0 and the buyer responds mostly to the base. Chetty, Looney and Kroft (2009, American Economic Review 99(4), 1145-1177) pinned θ well below 1 in a grocery experiment: posting tax-inclusive prices on about 750 products for three weeks cut their demand by roughly 8 percent relative to controls, even though the tax was public, unchanged, and added at the register either way. Making an already-known charge salient was enough to move quantity. Ellison and Ellison (2009, Econometrica 77(2), 427-452) documented the online version, where retailers use loss-leader base prices and costly add-on upgrades to obfuscate, and estimated the steep implied price sensitivity that obfuscation is designed to blunt.

Brown, Hossain and Morgan (2010, Quarterly Journal of Economics 125(2), 859-876) ran the field test on eBay: splitting a price into a low item cost plus a shrouded shipping charge raised revenue relative to an all-in price when the shrouded portion was large, and buyers under-reacted to the part listed separately. The controlled scale test came from ticketing. Blake, Moshary, Sweeney and Tadelis (2021, Marketing Science 40(4), 619-636) ran a field experiment on StubHub across millions of visits, showing about 15 percent in fees either upfront during browsing or held back until checkout. In the back-end condition, buyers were roughly 14 percent more likely to purchase and total revenue rose about 21 percent, and a meaningful part of the revenue gain came not from more sales but from buyers choosing more expensive tickets once the fee was out of view at the moment of comparison.

A 21 percent revenue swing from the timing of a fee is not a rounding error, and it is exactly the surplus the persuasion frontier says is available when the receiver under-weights part of the state. The uncomfortable part for anyone running a checkout is that the incentive is real and the effect reproduces, yet most dashboards never register the downstream cost. The comfortable story that "customers would see through it" is the same story the disclosure experiments demolished. I have yet to see a funnel where that story survived contact with the refund data.

Reviews as Information Design

Ratings and reviews are where these mechanics touch the most product decisions, because a review system is a sender in the persuasion sense even when every individual review is honest. The platform chooses the summary statistic, the ordering, the default filter, and the badge, and each choice is a signal structure applied before any particular buyer arrives.

The evidence that reviews move money is old and solid. Chevalier and Mayzlin (2006, Journal of Marketing Research 43(3), 345-354) compared the same books across two retailers and found that an improvement in a book's reviews at one site raised its relative sales there, and that one-star reviews moved sales more than five-star reviews did, an asymmetry that matters enormously for how you display the tail. Luca (2016, Harvard Business School Working Paper 12-016) used a regression discontinuity around Yelp's rounding thresholds and Washington State revenue data to show that a one-star increase in a restaurant's Yelp rating raised revenue by 5 to 9 percent, with the effect concentrated among independents and absent for chains, which already carry their own reputation. Reviews substitute for brand where brand is thin.

With reviews moving money like that, every display choice is a persuasion lever. Showing a single averaged star hides the shape of the distribution, and a 4.3 built from mostly fives and a few ones is a different product from a 4.3 built from steady fours, yet the average erases the difference the negativity asymmetry says buyers care about most. A default sort of "most helpful" is a curation rule, and if "helpful" votes accumulate on early, positive, lengthy reviews, the default quietly foregrounds a flattering sample while remaining, review by review, entirely truthful. A verified-purchase badge changes the meaning of silence: once some reviews are marked verified, an unmarked review reads as weaker evidence, which is unraveling operating inside the review list itself.

I treat the review surface the same way I treat a pricing page's information architecture: as a set of defaults that determine what a low-effort reader concludes. The on-page trust elements that a conversion team A/B tests, badges, counts, testimonials, are the visible layer, and their measured lift is real, as the field-test record shows. But the deeper design decision is upstream of any single element: it is what the system chooses to summarize and surface at all. A platform that optimizes the summary purely for conversion is running a persuasion policy whether or not anyone on the team has used the word.

Regulation Caught Up With the Theory, 2022 to 2025

For two decades the shrouding literature was a warning without teeth. Between 2022 and 2025 that changed on both sides of the Atlantic, and the interesting thing for an operator is how closely the rules track the economics rather than mere sentiment.

The European Union moved first on reviews. The Omnibus Directive, Directive (EU) 2019/2161, had to be in force across member states by 28 May 2022, and it targets exactly the silence problem: a trader may not claim reviews come from real purchasers unless it takes reasonable, proportionate steps to verify that, and platforms must disclose whether and how they check. In persuasion terms, the directive forces the provenance badge that makes silence informative, and it bans the fake positive review that would otherwise poison the pool.

California addressed price shrouding directly. Senate Bill 478, in force from 1 July 2024, makes it unlawful to advertise a price lower than what the consumer will actually pay, with only government taxes and genuine shipping excluded, which is a statutory ban on drip pricing rather than a disclosure nudge. The United Kingdom's Digital Markets, Competition and Consumers Act 2024 received Royal Assent in 2024, and its consumer provisions, in force from 6 April 2025, require mandatory fees to sit inside the headline price and, where a charge genuinely cannot be computed up front, to be given the same prominence as the total; the same Act attacks fake reviews, with penalties reaching up to 10 percent of global turnover.

The United States followed in the same window. The Federal Trade Commission announced its Rule on Unfair or Deceptive Fees on 17 December 2024, effective 12 May 2025, covering live-event ticketing and short-term lodging. The rule does not ban any particular fee or forbid dynamic pricing; it requires the total price, inclusive of mandatory charges, to be shown up front and more prominently than other pricing, with civil penalties of up to about $51,700 per violation. The regulatory theory is precisely the salience model: do not outlaw the fee, force θ toward 1 by making the total unavoidable at the moment of decision.

What an Operator Should Actually Do

In my advisory work the theory and the evidence converge on a short list of decisions, not tips. The first is to run a disclosure audit: enumerate every surface where the product chooses what to reveal, the star summary, the review sort, the default filter, the comparison table, the fee sequence, and the cancellation flow, and for each one write down what a low-effort buyer concludes and what a fully informed one would. The gaps between those two columns are your persuasion inventory, and some of them are loans you did not know you had taken.

The second decision is to stop trusting conversion alone and instead instrument the cost it hides. A shrouded surface almost always lifts the immediate conversion rate; that is why it survives in so many funnels, yet the immediate rate is the least informative number in the whole flow. The number that tells you whether the lift is real is downstream: refund and return rates, chargebacks and disputes, support contacts that mention surprise, and the retention difference between customers who felt informed and customers who felt tricked. In both engagements above, the shrouded design won on conversion and lost on the fuller ledger. If your dashboard stops at conversion, you are measuring the size of the loan and calling it profit.

The third decision is the honest one: when is partial disclosure legitimate rather than manipulative? The persuasion literature is amoral about this, so the line has to come from somewhere else, and the workable test is relevance versus concealment. A comparison table that leads with the attributes a buyer in this segment actually weights, and omits specifications nobody in that segment uses, is relevance; it lowers the buyer's effort without changing the conclusion a fully informed buyer would reach. A fee held back specifically because salience would change the decision is concealment; its entire value comes from the buyer reaching a conclusion they would reverse if better informed. The same distinction separates a review summary designed to inform from one designed to flatter. If the design's payoff depends on the buyer staying uninformed, it is shrouding, and shrouding now carries a legal tail as well as a trust cost.

The strategic frame that ties it together is the one from the persuasion frontier. Against a rational buyer, the concavification is the ceiling on what selective disclosure can win, and that ceiling is legitimate to pursue: structure the information, lead with your strengths, summarize honestly. Above that ceiling sits the extra margin that only a buyer's credulity provides, and that margin is temporary by nature, because feedback, competition, and regulation all erode it, in that order. In my experience, building a business on the durable part is slower and more defensible. Building it on the credulity is the discount-engineering mistake in another currency: a real short-run number that quietly mortgages the long-run relationship. The signal a firm sends by choosing honest salience is itself a form of costly signaling, and the review and rating rules a platform sets are a mechanism design problem in which users will play whatever the defaults reward. Choose the defaults as if a skeptical regulator and a well-fed feedback loop were both coming, because by 2025 they are.

Key Takeaways

1. Unraveling is real but conditional. Costless, verifiable disclosure to skeptical buyers forces full transparency top-down (Grossman 1981; Milgrom 1981), but the result rests entirely on buyer skepticism that the evidence shows is usually missing.

2. Silence pays because buyers under-read it. In a controlled game, senders disclosed a middling draw of 2 only 44.5 percent of the time against a 100 percent prediction, and profited, because receivers over-guessed the silent pool (Jin, Luca & Martin 2021). Only immediate, repeated, personal feedback narrowed the gap; aggregate disclosure statistics did not.

3. Bayesian persuasion sets the exact ceiling on honest spin. An honest sender who commits to an information structure can move a rational receiver up to the concave closure of her payoff (Kamenica & Gentzkow 2011): the prosecutor lifts convictions from 0 to 60 percent at a prior of 0.3 without lying.

4. Mandatory disclosure works when it is salient at the point of decision. Los Angeles hygiene grade cards raised A-grade revenue about 5.7 percent and cut foodborne-illness hospitalizations about 20 percent (Jin & Leslie 2003), because a letter in the window collapses inference cost that a filed score does not.

5. Shrouding and drip pricing are the price version, and the money is large. Making a known grocery tax salient cut demand about 8 percent (Chetty, Looney & Kroft 2009); holding StubHub fees until checkout raised revenue about 21 percent and purchase likelihood about 14 percent (Blake et al. 2021).

6. Reviews are information design even when every review is honest. The averaged star, the default sort, and the verified badge each move belief; one-star reviews move sales more than five-star (Chevalier & Mayzlin 2006), and a one-star Yelp gain lifts independent-restaurant revenue 5 to 9 percent (Luca 2016).

7. Regulation now tracks the salience economics. The EU Omnibus Directive (2022), California SB 478 (2024), the UK DMCC Act (2025), and the FTC fee rule (2025) force the total price and review provenance to the point of decision rather than banning fees outright. Build the salient surface before the deadline, and measure refunds and retention, not just conversion.

Further Reading

• Trust Signals and Their Measurable Lift, the on-page elements this essay treats as the visible layer of information design, with their field-tested effect sizes.
• Pricing Pages as Information Architecture, the pricing-surface counterpart to fee timing, where what you choose to show sets the buyer's default conclusion.
• Discount Engineering: When Promotions Destroy LTV, the same short-run-number, long-run-cost trade-off that shrouding faces, in the currency of price.