From a Helpful Signal to a Higher Price
This chapter follows the path from a personal signal to a personalized price. The useful question is not whether a seller knows everything about a buyer, but how a small predictive advantage can change the offer; tracing that path lets us distinguish an observed feature, a seller’s inference, and the buyer’s actual willingness to pay.
A personal feature is something a system can observe or receive: a choice, a browsing action, a context signal, or information the buyer provides. An inference is the conclusion drawn from that feature—for example, that the buyer is likely to value a product highly or is unlikely to walk away from a higher price. Willingness to pay is the buyer’s actual maximum acceptable price in the situation. These three things may be related, but they are not interchangeable.
A seller can be wrong about the buyer. The important point is that the seller does not need certainty. A prediction that is better than a guess can still change the seller’s decision. Personalized pricing begins when that prediction enters the offer itself.
The Price Is Built From an Estimate
Imagine that two people are considering the same product. Each person values it at 100 in the relevant situation, but the seller does not observe that value directly. Under a uniform price of 60, each buyer keeps a potential surplus of 40 if they purchase. Under a personalized system, the seller uses information about one buyer to predict that they would accept 90 and offers that buyer 85. The product is unchanged. The buyer’s potential surplus has fallen from 40 to 15 because the seller’s estimate changed the division of value.
This is a simple illustration, not a claim about a particular retailer’s pricing rule. It isolates the mechanism described in the research: behavioral or contextual information becomes evidence about a buyer’s valuation, and the seller uses that evidence to set a price. The seller is not changing what the product does. The seller is changing how much of the product’s value the buyer must surrender to obtain it.
The difference between the three layers matters for judgment. The feature might be a rough proxy rather than a direct measure. The inference might reflect a model’s assumptions or a seller’s objective. The buyer’s actual willingness to pay may depend on facts the system cannot see, such as a change in need, a hard budget, or a decision to postpone the purchase. Calling all three “the buyer’s value” makes the pricing process seem more objective than it is.
The same personal signal can also be used for two different jobs. A recommendation system may use it to choose a product, bundle, or explanation that better fits the buyer. That is service personalization: the system changes what it shows or how it assists. A pricing system may use it to decide what to charge. That is price personalization: the system changes the economic terms. One system can perform both jobs, but the consequences are different.
A better recommendation can create value by helping the buyer find something they genuinely want. A higher buyer-specific price can transfer part of that value to the seller. The first question is therefore not whether data were used. It is whether the use of data changed the buyer’s experience, the seller’s estimate, the economic offer, or several of these at once.
The pricing step can be explicit or hidden. An offer might display a different number. It might present the same list price but vary a discount, shipping charge, bundle, or eligibility condition. In a bargaining setting, the same mechanism can appear as a different division of the surplus rather than a different sticker price. The research material uses both kinds of language: a seller may set a personalized price, or an algorithm may use leaked information to make a more aggressive bargaining offer.
That last distinction is important because the buyer often sees the outcome but not the inference behind it. They may know the price, yet not know which signal moved the price, whether the signal was accurate, or whether the same offer would have been available without the data. This is the basic information asymmetry: the seller can act on a prediction that the buyer cannot inspect or challenge easily.
A prediction can be useful without being fair, and it can be unfair without being accurate. If a seller’s estimate is too low, the buyer may receive a discount or an opportunity they would not otherwise have received. If it is too high, the buyer may lose part of the surplus they would have kept under a common price. If the signal is a poor proxy, the buyer may receive a worse offer for a reason that says little about their actual valuation. The mechanism does not decide which of these outcomes is acceptable; it tells us where to look for the decision.
That decision can be tested with four questions. First, what exactly was observed: a preference, a behavior, a demographic fact, or merely a context? Second, what did the seller infer from it? Third, did that inference change the price, the bargaining split, or only the recommendation? Fourth, could the buyer discover or contest the connection? The first three questions trace the causal path. The fourth determines whether the buyer can respond to it. Without that last piece, a buyer may experience a changed deal without knowing whether the cause was personal information, ordinary demand conditions, or an error in the system.
The seller’s use of a signal also need not be visible at the moment of collection. Information can be gathered while the buyer is choosing a product and used later by another pricing algorithm. In the experimental setting supplied for this series, the later algorithm receives access to leaked information and makes a more aggressive bargaining offer. In a dynamic model, a seller receives observations over time and updates its estimate. The distance between collection and pricing is part of the mechanism: a harmless-looking exchange can become economically relevant at a later stage.
The sequence shows why “the algorithm changed the price” is incomplete. The price is the visible endpoint of a chain: the buyer reveals something, a system interprets it, the seller chooses an offer, and the buyer decides whether to continue. Each step creates a possible failure or contest. The signal may be misleading, the inference may be inaccurate, the offer may extract too much surplus, or the buyer may not have enough information to know which step caused the result.
Every Extra Interaction Changes Two Information States
The chain becomes harder to judge over time because the buyer and seller may learn different things from the same interaction. A buyer may browse longer to understand a product’s quality, fit, or value. The seller may observe more behavior and improve its estimate of the buyer’s reservation price—the highest price the buyer would accept in that situation. More learning for one side can mean more leverage for the other.
Consider a buyer comparing two versions of a product. At the beginning, the buyer is uncertain about which version is worth choosing. A few more interactions may reveal useful information: the buyer can inspect features, compare trade-offs, and discover whether the product meets a need. If the buyer is cautious, reducing that uncertainty can improve the expected value of purchasing.
But those interactions may also reveal preferences to the seller. The sequence of products inspected, the time spent comparing, or the decision to return to a particular option can provide clues about what matters to the buyer. A seller that receives increasingly informative observations may refine its pricing estimate. The buyer is learning about the product while the seller is learning about the buyer.
The 2020 stopping-time paper formalizes this as a dynamic decision. At each moment, the consumer can purchase, reject the good, or continue browsing. Continuing preserves the possibility of learning more and receiving a better future opportunity. It also permits more information to accumulate on both sides. The paper treats the seller’s pricing strategy as fixed and credibly committed in advance so that the consumer’s timing problem can be studied separately from a strategic response by the seller.
That assumption does not make the problem trivial. Even with a fixed strategy, the current offer may be worth accepting for one of two reasons. The price may be good enough now, or the buyer may be sufficiently uncertain about their own valuation that waiting has little expected benefit. Conversely, a positive current payoff does not prove that buying now is best. Waiting may reduce uncertainty or lead to a lower future price in some possible paths.
Risk aversion makes the distinction sharper. In the paper’s Gaussian example, the consumer’s expected valuation can remain the same while uncertainty about that valuation falls over time. A risk-averse buyer can then prefer a later decision because the range of possible outcomes is narrower. The seller, meanwhile, receives noisy observations and updates an estimate of the buyer’s valuation. The buyer’s reason to wait and the seller’s reason to keep observing are operating simultaneously.
The asymmetry is easiest to miss when the buyer experiences the interaction as learning rather than disclosure. A product page teaches the buyer something about the good; the same page may teach the seller something about the buyer’s urgency or preferences. The buyer’s new knowledge can improve the decision, while the seller’s new knowledge can narrow the buyer’s room to negotiate. Neither effect is guaranteed on every visit, but both belong in the same calculation.
This also explains why a price signal is difficult to interpret after the fact. Suppose a later offer rises. It could reflect a general change in demand, a different product or seller, a pricing error, or a more informative estimate of this buyer. The model does not give the buyer a magic test for distinguishing these causes. It supplies a disciplined suspicion: when an interaction collects information and the economic terms later change, ask whether the information crossed from service design into price setting.
This is why “share less” is not a complete explanation of the consumer’s problem. The buyer may need information from the interaction, and refusing every form of personalization can sacrifice a real service benefit. Nor is “keep browsing until the price falls” a safe rule. A later offer may be higher, the opportunity may disappear, or the seller’s new information may make the buyer easier to price. The relevant comparison is between the benefit of one more interaction and the information advantage it may give the seller.
The distinction between service and price personalization also helps locate the buyer’s practical leverage. If extra information only improves the recommendation, the buyer can evaluate whether the improved match is worth sharing the signal. If the information also changes the price or bargaining position, the buyer must ask a second question: whether the service benefit is being financed by giving the seller a better way to claim the resulting surplus. A better match does not automatically justify a less favorable deal.
There is a further boundary. In the stopping-time model, the seller’s strategy is treated as fixed rather than redesigned in response to the consumer’s behavior. That lets the paper characterize a consumer-side defense, but a real seller might adapt prices, data collection, or timing once it notices that buyers are trying to limit disclosure. The stopping logic remains useful as a way to think about option value; its simulated performance is not a guarantee against an intelligent opponent.
The mechanism now gives us a precise object to test. A buyer’s decision can depend not only on the size of a possible loss, but on how clearly that loss is described. A known probability can be weighed against a known benefit. An uncertain range forces the buyer to supply a belief about what may happen. The next question is therefore behavioral: when the same data-to-price mechanism is presented as measurable risk or as an ambiguous possibility, do people respond in the same way?