From Personal Data to a Personal Price

Personalized pricing becomes easier to judge once the word “algorithm” is taken out of the mystery. This chapter follows the path from a consumer or device signal to a displayed price, and asks which links in that path must be established before the result can be called personalized algorithmic pricing. The practical value of that path is simple: it separates a price that changes because the market changed from one that changes because a business learned something about the person seeing it.

The chain from data to a price

A pricing system usually begins with an estimate of demand. The business wants to know not only what a product costs to supply, but how likely a particular customer is to buy at different prices. That estimate can be crude or sophisticated. It may use information supplied during registration, a purchase history, a location, a loyalty account, browsing behavior, or characteristics inferred from a device. The important point is not that one particular kind of data must be present. It is that the business uses information connected to a consumer or device as an input to the pricing decision.

The next step is to estimate heterogeneous price response—the fact that different consumers may react differently to the same price. One customer may abandon a purchase after a small increase; another may continue because the product is urgent or alternatives are inconvenient. A firm can represent this difference as a predicted willingness to pay, a predicted price sensitivity, or a probability of conversion at several candidate prices. The prediction does not need to be perfectly accurate to be commercially useful. A small improvement across many transactions can be worth more than a precise prediction for a few individuals.

The business then turns predictions into segments or price rules. A segment is a group of consumers who look sufficiently alike, for the purpose at hand, that the system assigns them the same price or offer. With enough features, the groups can become very fine-grained. In the economic sense, this is an extreme form of third-degree price discrimination: prices vary across estimated consumer types. It can look individual even when the system is technically assigning a customer to one of many groups rather than calculating a unique number from scratch.

It helps to separate three operations that are often collapsed into the word “personalization.” The first is measurement: collecting or receiving a signal about a person, household, account, or device. The second is inference: using that signal, usually alongside many others, to estimate a response to price. The third is application: using the estimate to change the amount, discount, offer, or fee that reaches the consumer. Data can be present in the first operation without being used in the third. Conversely, a business may use a previously collected signal at the moment of pricing even if the consumer did not provide new information on that screen. The causal and legal question is about the link to the pricing decision, not merely the existence of a large database.

The output need not be a unique number for the practice to have a personalized effect. A system might keep a base price unchanged and vary a discount, show a promotion to one account, change the reference price used to describe a sale, or add a fee. From the consumer’s point of view, these pieces can combine into the amount paid. From an evidentiary point of view, they are separate outputs that may be controlled by different actors. That is why the New York Attorney General asked about prices, original prices, discounts, offers, promotions, and loyalty programs rather than assuming that only a changing base price could matter.

Imagine, as a simplified example, two customers requesting the same service at the same time. The system may charge both the same amount because the relevant demand estimate is similar. It may show different amounts because one customer’s data predicts a lower likelihood of leaving, or it may show different discounts because the customers belong to different test cohorts. The visible result is the same kind of fact—a different number—but the mechanism, evidence, and policy consequences are different. A useful explanation must preserve that difference all the way from the input to the screen.

Machine learning matters here because it lowers the cost of finding patterns in many variables and applying them repeatedly. It does not remove the need to choose an objective. A business may optimize immediate revenue, expected profit over several months, conversion, retention, or the performance of a promotion. Different objectives can produce different prices from the same data. Nor does machine learning turn a prediction into a fact about the person. A model still contains assumptions, estimation error, and information it cannot observe.

The ZipRecruiter study illustrates this structure without making the system sound omniscient. The researchers modeled a customer’s purchase decision as a response to price and observable business features. They used self-reported features to estimate how price sensitivity varied, then used those estimates to recommend uniform and personalized prices. The model could identify useful differences among prospective employer customers, but it still had to account for unobserved demand shocks. Two businesses with the same recorded features could make different choices because the model did not know everything that mattered to them.

That limitation separates personalized pricing from perfect price discrimination. Perfect discrimination would let a seller charge each customer exactly the highest amount that customer would pay. A practical personalized system generally cannot do that. It sees signals, not the whole person; it predicts a response, rather than reading a settled willingness to pay. Its uncertainty can leave money with consumers, cause it to set a price too high and lose a sale, or lead it to offer a lower price to a customer it believes is more price-sensitive.

New York’s statutory language, as quoted by the state Attorney General, captures the relevant connection in a compact way. It defines personalized algorithmic pricing as dynamic pricing set by an algorithm that uses personal data. “Dynamic pricing” concerns a price that fluctuates depending on conditions. “Personal data” concerns data that identifies, or could reasonably be linked directly or indirectly, to a specific consumer or device. The legal distinction is therefore not simply between a human and a machine. It is between a fluctuating price that responds to conditions and a price whose setting uses personal data tied to the person or device receiving it.

sequenceDiagram participant C as Consumer or device participant P as Platform or retailer participant M as Pricing model participant D as Price display C->>P: Request product or service P->>M: Combine personal data and market conditions M->>M: Estimate demand and assign a price rule M-->>D: Return price or promotion opt Personal-data pricing D-->>C: Show price and relevant disclosure end D-->>C: Present purchase choice C->>P: Buy, compare, correct, or leave

The sequence shows where the word “personalized” enters. Market conditions may be part of the same calculation, but personal data is a separate input. The model mediates between the input and the displayed amount; the price display is the point at which the consumer can learn that the mediation occurred. A notice cannot make the prediction accurate, but without notice the consumer may not know that accuracy, data use, or an alternative purchasing route is worth investigating.

The sequence also explains why a model’s sophistication is not the main issue. A simple rule such as “show a different discount to customers associated with a particular account history” can have the relevant structure. A complex model may be more accurate but still fail to produce a covered price if it uses only shared inventory or demand conditions. In practice, several models can operate in succession: one estimates demand, another selects an offer, and a platform assembles a final screen. The consumer needs to know which stage changed the price, while an investigator needs records that connect the stages.

Why a different screen is not proof

The most common mistake is to reason backward from an unequal outcome. If two people see different prices, it is tempting to conclude that one person was profiled. That conclusion may be right, but the visible difference does not identify its cause. A rideshare fare can vary because supply and demand are different in two nearby areas. A hotel rate can reflect availability and timing. Insurance and credit prices may depend on individualized risk as part of the product itself. These are not equivalent to a retailer using a customer’s personal information to estimate how much more they can pay.

Even within one online store, several mechanisms can be layered together. A retailer may run a randomized price test, placing customers into cohorts to measure how different offers affect sales. It may offer a loyalty discount visible only to account holders. A brand may fund a promotion while the retailer controls a base price. A platform may show a different “original price,” discount, delivery fee, or service charge even when the final sale price matches. The customer sees the final screen, but the business may have several systems producing the pieces that appear on it.

This is why “algorithmic” is not a sufficient explanation. An automated system can select a promotion for reasons that are not personal pricing. It can assign a test group randomly. It can update an inventory-sensitive price for everyone exposed to the same local conditions. Conversely, a business could use a relatively simple rule and still be using personal data. The relevant question is what the system does, not how advanced it sounds.

A supported classification requires a causal chain with at least four links. First, the business must identify the input: data that is specific to, or reasonably linked with, the consumer or device. Second, the evidence must show that an algorithm or automated rule used that input in setting a price, discount, offer, or fee. Third, the resulting amount or pricing statement must be the one shown or published to the consumer. Fourth, the practice must be evaluated in the market and legal context that applies to that display. A correlation between a customer characteristic and a price can suggest a hypothesis, but it does not by itself prove each link.

Each link has its own failure mode. A personal-data record may be stored but never passed to the pricing service. A model may use the record to predict demand but a separate experiment may randomly determine which price is shown. A price may differ because a promotion is account-specific, while the product’s underlying price is common. Or an analyst may compare screens from different stores, times, delivery zones, or inventory states and attribute the difference to the customer when the market conditions were not held constant. The further the observation is from the actual pricing log, the more carefully the conclusion must be stated.

The strongest comparison is not simply “customer A paid more than customer B.” It asks what was held constant, what varied, who assigned the variation, and which data fields were available at the decision point. That is why records of cohort assignment, model inputs, experiment dates, adjusted variables, and the final screen are more probative than a collection of screenshots alone. Screenshots establish what was displayed. They do not, by themselves, establish why it was displayed.

The New York Attorney General’s requests to Instacart make this standard concrete. The letter sought descriptions of pricing tools, the variables changed in experiments, the way customers were selected, the data used for different values, the effect on individual and basket-level prices, and screen flows showing where disclosures appeared. Those requests are not mere paperwork. They are an attempt to reconstruct the path that the final number conceals. Without the experiment records and model description, an observer may know that prices differed but not whether personal data caused the difference.

The distinction also matters for the consumer’s expectations. The FTC’s proposed enforcement statement says people generally expect a shelf price or a product listing to be the same price that another person would see at the same place and time, while recognizing that variation is more familiar in some markets. A consumer’s expectation is not proof of what happened. It helps determine whether a hidden pricing basis could be material or misleading. If a price looks common but was individualized, the omission may affect whether the person compares another seller, changes a purchasing route, challenges inaccurate data, or declines the transaction.

That is the role of disclosure. New York’s quoted rule requires a clear and conspicuous statement when its specified conditions are met. The FTC proposal would generally expect disclosure not only of personalization, but also of its basis and the types of data used where consumers reasonably expect common pricing. Both approaches treat notice as a way to restore a choice that opacity removes. Neither approach says that a label proves the data is correct, resolves the distributional effects of the price, or makes every disclosed practice lawful.

Disclosure can also fail in more than one way. A notice may exist but be buried behind fine print, separated from the price, or written so vaguely that the consumer cannot tell whether the amount, discount, or fee was individualized. It may accurately say that personal information is used while omitting the feature that materially changes the price. It may appear after the consumer has committed to a transaction. Visibility is therefore not a decorative feature of compliance; it affects whether the notice can change behavior.

There is a practical difference between disclosure that merely names a practice and disclosure that makes the practice usable knowledge. “Personal information may affect your experience” leaves the consumer unable to tell whether the price changed, what kind of information mattered, or what response is available. A more informative notice identifies the pricing basis at the moment the amount is presented. It still cannot reveal every model parameter or prove that the data is correct, but it gives the consumer a reason to compare the price, ask for correction, or decide whether to proceed.

The chain from input to price gives us a disciplined way to approach the next case. We should ask what shoppers actually observed, what the company says its tests and promotions did, what personal-data link the regulator is investigating, and what the screen communicated at the moment of choice. An unequal price is an important signal. It is not yet an explanation.

References