What a Fair Pricing Rule Must Make Visible
Personalized pricing is not simply banned because the same visible outcome can come from different mechanisms, the economic effects are not one-directional, and the legal tools in the supplied record are designed for more than one kind of price difference. This final chapter brings those pieces together. It uses the data-to-price chain, the Instacart inquiry, the ZipRecruiter experiments, and the New York and federal approaches to give readers a practical standard: identify what changed, trace who or what caused it, measure who bears the effect, and then ask what the consumer was told and could reasonably do.
Unequal prices are an observation, not a conclusion
The most important connection across the evidence is a distinction between an outcome and its explanation. Two consumers can see different prices for the same product, but the difference may reflect a shared market condition, a randomized test, a promotion, a loyalty rule, a partner’s markup, or personal data used to estimate one consumer’s response. These mechanisms can coexist on the same platform. A displayed number is therefore a starting observation, not a complete account of personalized pricing.
The Instacart record shows why that distinction matters in a consumer market. The New York Attorney General summarized a report in which 437 shoppers in four cities observed multiple prices for many of the same grocery items. The summary described prices up to 23 percent higher for the same product in the same store at the same time, an average 13 percent gap between the highest and lowest price for a single good, and differences in original prices and discounts even when the sale price matched. Those observations establish a serious pattern as officially reported. They do not establish which system produced each price or whether personal data caused every difference.
The ZipRecruiter study answers a different question. Researchers randomly varied prices for prospective employer customers, estimated heterogeneous demand, and tested personalized recommendations with a new sample. That design is stronger for learning how prices affected conversion and revenue in that enterprise market. It does not identify what happened in Instacart’s grocery experiment, because the firms, customers, products, dates, and pricing systems differ. The two cases can be connected at the level of mechanism and evidence: both concern a firm learning that customers respond differently to price. They cannot be combined into a single estimate of how common personalized pricing is or how much it costs households.
A careful assessment moves through four links. It identifies the price component that changed, whether that was a base price, a discount, an original price, a promotion, or a fee. It identifies the actor or tool that controlled the component. It asks what data and assignment rule were available at the decision point and whether they actually affected the output. It then checks what was displayed to the consumer before the choice. Each link matters because a failure at one point changes the claim that can responsibly be made.
This standard also explains why the Attorney General requested agreements, experiment records, data categories, algorithms, customer-selection rules, and screen flows. Screenshots can show that two people saw different amounts. They cannot show whether a retailer, brand, platform, or random test assigned those amounts. A privacy policy can mention data collection. It cannot prove that a particular field entered the calculation that produced a particular price. Pricing logs, model inputs, experiment documentation, partner agreements, and time-matched displays answer different parts of the question.
The word “personalized” should therefore be reserved for a meaningful data-to-price connection. A system that changes a delivery charge because local conditions changed is not automatically the same as a system that uses information linked to a consumer to infer that they are less likely to switch. A randomly assigned discount is not automatically the same as a price chosen from a customer’s predicted willingness to pay. A loyalty offer can be individualized in one sense while leaving the underlying price common in another. The classification must follow the mechanism rather than the emotional force of the word “algorithm.”
The same rule protects against an opposite error.