The Same Product, a Different Price

Personalized pricing is the subject of this series: the practice of using data and algorithms to set different prices for different consumers. The opening question is not whether two people can ever see different prices. They plainly can. The useful question is what caused the difference, what the consumer can know about it, and how the practice changes the balance between a firm’s revenue and consumers’ welfare. Those distinctions make it possible to judge the issue without treating every price change as either harmless commerce or an automatic legal violation.

A price normally looks like a simple fact. A product is on a shelf, a service appears on a screen, and the number beside it seems to be the number anyone else would pay at that place and time. That expectation is not a law of nature. Prices can change because supply and demand change, because taxes or local conditions differ, or because the product itself—insurance, for example—depends on the buyer’s risk. But a different kind of change occurs when a business uses information about a particular consumer or device to decide what that consumer will see. The price is no longer only a response to the market. It is also a response to a data-based estimate of the person.

The core theme is visibility and contestability. A consumer cannot compare, challenge, or avoid a pricing practice that they cannot recognize. At the same time, a visible price difference does not reveal its own cause. The rest of this series uses economic evidence, a live regulatory dispute, and emerging disclosure rules to examine that gap between what appears on the screen and what the pricing system is doing underneath.

A Price Can Differ Without Being Personalized

The first distinction is between ordinary dynamic pricing and personalized pricing. Dynamic pricing means that a price fluctuates with conditions. A rideshare fare may rise when demand is high in one neighborhood. A hotel rate may change with timing or availability. The change can affect many customers who face the same conditions; it does not necessarily depend on a profile built about one customer.

Personalized pricing adds a consumer-specific input. A business may use past purchases, browsing behavior, location, income-related information, loyalty data, or other features to estimate how sensitive a person is to price. An algorithm then assigns that person to a segment or calculates a price for that person. The economic literature calls this a highly granular form of third-degree price discrimination. It is “personalized” because the segments can be fine enough to approximate individual treatment, not because the system has perfect knowledge of the individual.

That last qualification matters. Perfect, or first-degree, price discrimination would let a seller charge each buyer exactly their maximum willingness to pay. Real systems do not have that power. They rely on imperfect signals, model assumptions, and patterns that leave important preferences unobserved. Two people who look alike to a model may still respond differently. A firm can therefore gain useful predictive power without being able to extract every dollar of value from every buyer.

This is why the number on the screen is only the beginning of the inquiry. Suppose two shoppers see different prices for the same item. The difference might reflect different store inventories, a promotion assigned to a randomized test group, a retailer’s ordinary markup, or a fee calculated from delivery conditions. It might also reflect personal data used to set the price. The outcome alone cannot distinguish those explanations.

flowchart LR A["Shared market conditions"] --> C["Price-setting rule"] B["Consumer-specific data"] --> C C --> D["Displayed price"] D --> E["Consumer choice and response"] C --> F["Firm profit"] E --> G["Consumer welfare"] D --> H["Disclosure question"]

The diagram captures the basic problem. Shared conditions and consumer-specific data can enter the same pricing system, but they raise different questions. The displayed price affects a consumer’s choice; the rule behind it affects the firm’s profit; and the relationship among the two determines whether disclosure is needed. A consumer who sees an unusual number cannot infer the full path from the result alone.

The distinction also keeps the subject from becoming a debate about algorithms in the abstract. An algorithm is a method for turning inputs into an output. The relevant issue is which inputs it uses, what output it controls, and whether the consumer reasonably expected that output to be common rather than individualized. Those are questions about a pricing practice, not about whether software was involved.

The Economic Tension Is Real

Businesses have a reason to learn which customers are more price-sensitive. If one uniform price is too low for some buyers and too high for others, a firm may raise revenue by charging different prices. The NBER study of ZipRecruiter provides a controlled example. Researchers randomly varied monthly prices for prospective employer customers in 2015, estimated how reported business characteristics were related to price response, and tested their recommendations with a new group of customers.

The study found a strong incentive to move away from the company’s $99 status-quo price. Its estimated personalized pricing raised expected revenue per consumer by 19 percent over the model’s optimized uniform price. Even after imposing a $499 cap because some unconstrained recommendations reached far beyond the tested price range, expected revenue remained more than 8 percent above uniform pricing. A second field experiment broadly matched the predicted revenue pattern. These results do not describe grocery delivery, and they do not establish a universal profit gain. They do show why a firm might pursue personalization even when the practice attracts scrutiny: the commercial payoff can be substantial in a setting where demand varies across customers.

The consumer side cannot be reduced to the same revenue number. In the ZipRecruiter case, a majority of consumers were predicted to receive a personalized price below the optimal uniform price. Yet total consumer surplus fell under the usual arithmetic measure because a smaller group faced much larger losses. Both statements can be true. “Most people paid less” describes the distribution of changes; “total surplus fell” describes their aggregate value. A welfare measure that gives more weight to reducing inequality can prefer the allocation that the simple total rejects.

The study also warns against assuming that a data restriction has one predictable effect. Across modeled combinations of feature groups, some narrower information sets produced lower consumer surplus than broader ones. The result does not mean that collecting more personal data is generally beneficial. It means that welfare depends on how particular signals sort consumers into prices, not simply on the count of variables available to the firm.

The Instacart episode brings the same tension into a consumer-goods setting, but with a different kind of evidence. The New York Attorney General said a study of 437 shoppers across four cities found that nearly three-quarters of grocery items appeared at multiple prices, with some prices up to 23 percent higher for the same product in the same store at the same time and an average high-to-low difference of 13 percent. The Attorney General’s letter also described instances in which shoppers saw as many as five sale prices for one item, as well as different original prices and discounts.

Those observations make the issue concrete, but they do not by themselves prove the mechanism. Instacart said customers had been assigned to price-testing cohorts and that retail partners might use behavioral data for discounts or promotions. The company announced that it was ending item price tests while continuing to say that retailers would set their own prices and that promotions, marketing incentives, and loyalty integrations would remain. The Attorney General then requested agreements, experiment records, descriptions of pricing tools, data categories, and screen flows. That response is important: the official record treats the price gap as a reason to investigate, not as a completed finding that every difference was personalized algorithmic pricing.

Why “Not Simply Banned” Is the Right Question

The economic uncertainty and the evidentiary gap help explain the policy direction in the supplied material. New York’s Algorithmic Pricing Disclosure Act, as quoted by the Attorney General, addresses a business that uses an algorithm and personal data to set a price shown to a New York consumer. It requires a clear and conspicuous disclosure near the relevant price. That is a direct intervention in visibility: the consumer should not have to discover a hidden pricing policy by comparing accounts, taking screenshots, or reading buried fine print.

The New York approach is therefore not the same as a categorical ban on price differences. It does not say that every dynamic price, promotion, or individualized offer is prohibited merely because it varies. It says that when the quoted statutory conditions are met, the consumer must be told that personal data helped set the price. The Attorney General’s concern about Instacart’s disclosure was not only whether a notice existed, but whether it was prominent enough and close enough to the displayed prices to serve its purpose.

The FTC’s proposed enforcement policy takes a related but broader approach. It says the Commission lacks authority to prohibit personalized pricing in every circumstance, while stating that hidden or misleading practices may violate the existing ban on unfair or deceptive acts. Where consumers reasonably expect a common price, the proposal says a business should disclose not only that a price is personalized, but also the basis for personalization and the types of data used. The proposal is not a binding rule, and it does not declare that a fully disclosed price is always lawful. It describes how existing enforcement standards may apply.

Disclosure matters because it can change the consumer’s options. A person who knows that a price was individualized may compare another retailer, use a different account or purchasing route where lawful, correct inaccurate information, decide not to complete the transaction, or question whether the data was collected for that purpose. Notice cannot undo a price already paid, and it cannot settle whether a redistribution of surplus is fair. It can, however, make a hidden choice visible enough to contest.

A responsible judgment about personalized pricing therefore asks more than whether the final numbers differ. What condition changed? Did the pricing rule use data linked to a particular consumer or device? Did an algorithm or a test determine the displayed price, discount, or fee? Was the practice disclosed where the consumer could reasonably see it? Who gained, who lost, and could the affected consumer avoid, correct, or challenge the result? The answers will not always produce the same legal or economic conclusion.

That is why personalized pricing is not simply banned—and why invisibility is not a trivial detail. A universal prohibition would ignore differences among markets, the limits of the evidence, and the possibility that some consumers receive lower prices. A hands-off approach would ignore the firm’s incentive to hide a material pricing basis and the consumer’s inability to respond without knowing it. The next step is to follow the data-to-price path closely enough to tell those cases apart.

References