Blog · Analysis · Last reviewed June 25, 2026

The Price Becomes a Personalized Prediction

Surveillance pricing turns the price tag into a model output: not only what a product costs, but what a system thinks a particular person can be made to pay.

The governed object is the effective offer: price, discount, fee, ranking, financing, shipping, urgency, consent state, agent access, and the non-personalized path the buyer could have used.

From Sticker to Signal

A price used to be a public fact, or at least it pretended to be one.

The price on a shelf, menu, catalog, ticket board, or website could be unfair, inflated, predatory, confusing, subsidized, or discriminatory. But it was usually visible as a social object. Buyers could compare it. Competitors could respond to it. Reporters could cite it. Regulators could inspect it. A person could say, "That is too expensive," and the statement referred to something other people could see.

Surveillance pricing changes the object. The price becomes a prediction about the buyer. It may depend on location, device, browsing history, purchase history, loyalty membership, search behavior, inferred income, inferred urgency, previous cart abandonment, referral channel, time of day, or the seller's estimate of willingness to pay. The same product can become many prices, each visible only inside a private interface session.

This is not the same as ordinary sales, coupons, regional pricing, or airline-style yield management. Those practices can still be controversial, but they often respond to public conditions: supply, demand, season, inventory, advance booking, or broad customer segment. The sharper AI-era issue is personalization based on behavioral surveillance. The price is no longer only about the market. It is about the model of the person.

For this essay, surveillance pricing means individualized or population-targeted pricing that uses personal data, behavioral traces, inferred traits, or vulnerability signals to estimate what a buyer will accept. The problem is not that prices move. The problem is that the movement can be driven by a hidden profile rather than by inspectable market conditions.

The audit object should be the effective offer, not only the base number beside the product name. A pricing system can change the displayed price, withhold a discount, reorder options, alter fees, shift financing terms, change shipping defaults, create artificial urgency, or make the cheaper path harder to find. A shopper may never see a higher nominal price and still receive a more expensive, more constrained, or less comparable market.

Current Context

As of June 25, 2026, surveillance pricing has moved from a speculative consumer worry into an active privacy, consumer-protection, and competition-policy issue. The Federal Trade Commission's 2024 orders and January 2025 staff perspective are not final enforcement findings, but they do establish a public record: intermediaries market tools that can combine customer data, behavioral signals, segmentation, and price optimization.

The legal field is splitting into two approaches. One is disclosure: New York's General Business Law Section 349-a tells covered businesses to disclose when a price was set by an algorithm using personal data, and EU consumer law requires pre-contract information when a price is personalized on the basis of automated decision-making. The other is substantive restriction: Maryland's Chapter 154, approved April 28, 2026 and effective October 1, 2026, limits certain personalized higher pricing in the food retail and third-party food delivery context while also adding a broader algorithm-or-personal-data disclosure rule.

New York's posture is also changing. Its existing Section 349-a is a disclosure law. By June 2026, the One Fair Price Act, A9349B/S8623B, had passed both the Assembly and Senate, and the official bill page listed the status as "Passed Senate & Assembly" rather than signed. If enacted, it would move beyond the disclosure model by prohibiting surveillance pricing, restricting data use for that purpose, preserving specified bona fide discounts, and requiring disclosure when automated pricing systems are used. That makes New York a live example of the policy ladder: first name personalized algorithmic pricing, then decide whether some forms should be barred.

Competition authorities are treating pricing algorithms as market structure, not just retail interface design. The OECD's 2025 G7 report focuses on AI-enabled algorithmic pricing, common third-party tools, monitoring, hub-and-spoke coordination, tacit collusion concerns, price discrimination, and exclusionary effects. The same software that personalizes a checkout page can also reshape how firms observe and react to each other.

The next frontier is agentic commerce. If a shopping agent compares offers or buys on a user's behalf, sellers may personalize not only to the person but also to the agent, wallet, browser, credential, or mandate. Price governance therefore has to cover machine-readable offers, receipts, ranking, sponsorship, and whether the agent can preserve a non-personalized path.

What Changed

The Federal Trade Commission put the issue into public view in 2024 and 2025. In July 2024, the FTC issued orders to eight companies offering surveillance-pricing products and services: Mastercard, Revionics, Bloomreach, JPMorgan Chase, Task Software, PROS, Accenture, and McKinsey & Co. The agency said these intermediaries advertised the use of advanced algorithms, artificial intelligence, and historical or real-time customer information to target prices for individual consumers.

In January 2025, FTC staff released initial findings from that market study. The agency said details such as precise location, browser history, demographics, browsing patterns, shopping history, mouse movements, and items left unpurchased in an online cart can be used to tailor consumer pricing. Staff also said the intermediaries it examined worked with at least 250 clients, including consumer-facing businesses such as grocery, apparel, health and beauty, home goods, convenience, and building-material retailers.

The FTC's language matters because it shifts the frame from "dynamic pricing" to "surveillance pricing." Dynamic pricing can mean a price changes as demand changes. Surveillance pricing means the data exhaust of a person's life can become part of the pricing machine. That data exhaust may come from the retailer, a loyalty program, a data broker, an advertising identifier, a real-time bidding trace, or a platform partner. The object being optimized is no longer simply inventory, margin, or market clearing. It is the boundary of the buyer's tolerance.

The staff report was preliminary and based on aggregated or anonymized information because the 6(b) process protects confidential business material. That caveat is important. The public record does not prove every worst-case use in every sector. It does show that an intermediary ecosystem exists, that granular personal and behavioral data can feed pricing tools, and that regulators now see individualized algorithmic pricing as a privacy, competition, and consumer-protection problem.

The public-comment record is also narrower than the original January 2025 announcement suggested. A later FTC statement by then-former Chair Lina Khan said the surveillance-pricing request for public comment was withdrawn on January 22, 2025 and the docket was closed for further comments. That does not erase the staff perspective or the 6(b) orders, but it does mean the record should be cited as a preliminary market-study record, not as a completed rulemaking or final enforcement finding.

Offer Surfaces

The word "price" is too small for the modern checkout. A useful audit separates at least six offer surfaces.

Public dynamic pricing changes a visible price because of inventory, season, demand, capacity, or time. It can still be unfair or deceptive, but the driver is a market condition rather than a profile of a particular buyer.

Segment pricing shows different offers to groups: loyalty members, students, regions, referral channels, browser cohorts, app users, or predicted value tiers. The segment may be benign, discriminatory, or opaque depending on what data defines it and who is excluded.

Person-specific pricing changes the offer for a particular buyer or household based on purchase history, location, device, identity graph, browsing behavior, cart abandonment, urgency, or inferred willingness to pay. That is the core surveillance-pricing concern.

Personalized discounts can lower prices, but they still decide who receives relief. Withholding an ordinary discount from a profiled buyer can be functionally similar to raising the price.

Friction and fee personalization changes the path rather than the sticker: shipping fees, financing terms, subscription defaults, bundles, return windows, ranking, payment options, late-arriving fees, or the number of steps required to reach the cheaper option.

Agent-facing prices are offers shaped for shopping agents, wallets, comparison tools, or delegated payment systems. If a seller can infer the user's agent, mandate, budget, or switching costs, personalization can move from the human-facing page into machine-readable commerce.

This matters because a disclosure about "the price" can miss the effective offer. A regime that only watches the displayed base price can miss a personalized coupon, fee, ranking, financing offer, delivery window, return term, or agent-access rule that changes what the buyer actually receives.

The Effective Offer Receipt

The practical audit unit is the effective offer receipt. It is not merely a checkout receipt after payment. It is a record of the offer environment the buyer encountered before choice: the public reference price, the personalized displayed price, discounts shown or withheld, fees, shipping terms, financing terms, ranking position, urgency language, loyalty or consent state, agent credential, experiment cell, data categories, model or rule version, vendor role, non-personalized path, and final checkout price.

This record matters because personalized pricing often disappears when the session ends. A consumer screenshot can show one surface, but it usually cannot show the baseline, the profile inputs, the treatment group, the cheaper route not displayed, or the vendor system that produced the offer. A regulator needs enough structured evidence to reconstruct the decision without forcing every consumer to become a forensic investigator.

The receipt also has a privacy boundary. Keeping evidence for enforcement should not become a new dossier for marketing optimization. A serious design would separate consumer-access and regulator-access records from ad targeting, retention modeling, and future price training. It would minimize retention, restrict secondary use, and let the buyer obtain a human-readable explanation of the effective offer without expanding the surveillance system that produced it.

The Intermediary Market

The most important actor is often not the retailer whose name appears on the screen.

Surveillance pricing depends on a stack: data collection, identity resolution, analytics, customer segmentation, experimentation platforms, recommendation systems, price optimization tools, marketing automation, payment data, loyalty programs, ad-tech traces, data clean rooms, and retail software. A merchant may buy the capability from a vendor rather than build it from scratch. A consumer may experience the result as a normal product page.

That makes accountability hard. If a shopper sees a higher price, who set it? The retailer? The pricing vendor? The marketing platform? The loyalty system? The data broker? The model trained on similar users? The A/B test? The rule engine? The human pricing manager who approved an automated recommendation?

AI intensifies the ambiguity because the interface can personalize more than the number. It can personalize ranking, urgency language, discount timing, checkout friction, displayed alternatives, delivery fee framing, bundle suggestions, financing offers, and return prompts. The nominal price may be only one part of an adaptive commercial environment. A user might never be charged a different base price, yet still be steered toward a higher effective price through what the system chooses to show first, hide, emphasize, or make tedious.

This is why a pricing audit has to follow both the data path and the decision path. A cookie banner may produce the permission record, a clean room may produce the segment, a broker may provide enrichment, a pricing vendor may recommend the offer, and a payment or commerce agent may execute the transaction. Accountability fails if each participant treats the consequential price as someone else's downstream use.

This is why pricing belongs beside search, recommendation, feeds, hiring, insurance, lending, and public-service triage. A model does not need to issue a formal denial to change a person's life. It can alter the offers available at the point of action.

Comparison Breaks

Markets rely on comparison. Surveillance pricing weakens the conditions that make comparison meaningful.

If each user sees a different price, the buyer cannot easily know whether the offer is fair, inflated, discounted, or punitive. If the price changes after login, after a cookie is set, after a location is inferred, or after a cart is abandoned, the buyer may not know which behavior changed the system's estimate. If the system creates urgency or hides alternatives, the buyer may not know whether waiting, clearing cookies, using a different device, asking a friend, or shopping elsewhere would produce a better result.

The deeper harm is epistemic. The interface turns the market into a private experiment. Each person becomes an observed subject inside a pricing environment whose rules are not visible. The user sees an offer. The system sees a profile, a treatment condition, a conversion probability, and an expected margin. The same point appears in physical commerce when the smart cart becomes the checkout witness: the store can observe the path to purchase, not only the final purchase.

The baseline problem is central. A person cannot know whether a personalized offer is favorable without a reference point: public shelf price, non-personalized price, lowest available price, recent average, legally defined prior price, or a comparable offer available without behavioral profiling. Without that baseline, "personalized" can mean either a discount or an extraction event.

That does not make every personalized price exploitative. There are benign or mixed cases: targeted discounts, income-sensitive offers, student pricing, retention offers, accessibility-related accommodation, or lower prices for genuinely price-sensitive customers. But the same technical capacity can also identify desperation, loyalty, ignorance, fatigue, isolation, or high switching costs. Without disclosure and limits, the user cannot tell which world they are in.

The Law Enters

New York became the clearest U.S. disclosure test case. General Business Law Section 349-a requires covered entities doing business in New York to include a clear and conspicuous disclosure when they set the price of a good or service using personalized algorithmic pricing and use personal data specific to the consumer. The statutory text requires the statement: "THIS PRICE WAS SET BY AN ALGORITHM USING YOUR PERSONAL DATA."

That is a disclosure model rather than a ban. It does not say every personalized algorithmic price is illegal. It says the consumer must be told when the price has been set this way. The law also defines personalized algorithmic pricing as dynamic pricing derived from or set by an algorithm using consumer data and varying among individual consumers or consumer populations.

Maryland adds a different model. Chapter 154, approved by the Governor on April 28, 2026 and effective October 1, 2026, restricts covered food retailers and third-party food delivery service providers from using dynamic pricing or surveillance personal data to set a higher price for specified food for a single consumer or group of consumers. It also bars use of protected-class data where that use withholds or denies an accommodation, advantage, or privilege. Separately, Maryland's broader disclosure rule tells merchants not covered by the food-specific section to use the statement: "THIS PRICE WAS SET BY AN ALGORITHM OR BY USING YOUR PERSONAL DATA."

The European Union had already moved in a similar transparency direction. Directive (EU) 2019/2161 added a consumer information requirement for cases where a price is personalized on the basis of automated decision-making. EU guidance distinguishes this from dynamic or real-time pricing that changes in response to market demand without personalization based on automated decision-making.

Disclosure is useful, but it is not sufficient. A label can tell the user that a model is involved without explaining what data mattered, whether the price is higher or lower than the baseline, how to opt out, whether protected-class proxies were used, whether the same product is cheaper elsewhere, or whether ranking and friction also changed. A disclosure can become a ritual of transparency: true enough to satisfy the rule, too thin to restore bargaining power.

That is why the legal question should not stop at the warning sentence. Enforcement also needs the reference price, the non-personalized path, the data categories used, the discount or fee logic, the protected-class proxy review, and the session record showing what the consumer actually received. Without those records, a disclosure law can name the machine while leaving its treatment invisible.

The Competition Problem

Surveillance pricing is also a competition problem.

The OECD's 2025 report on algorithmic pricing in G7 jurisdictions describes pricing algorithms that can set or recommend prices in near real time, using granular data and goals such as revenue, profit, and market share. It identifies risks including algorithm-facilitated collusion, hub-and-spoke schemes through common third-party tools, resale-price monitoring, tacit collusion concerns, and unilateral strategies that can harm consumers through price discrimination or exclusionary effects.

The common-vendor problem is especially important. If many firms rely on the same pricing intermediary, the market may become easier to coordinate even without an old-fashioned agreement. Nonpublic information, shared software, automated monitoring, and similar optimization goals can reshape competitive behavior. The same pattern appears in rent-setting software: an optimization tool can become market infrastructure. The law is still learning how to interpret that stack.

At the consumer level, personalized pricing can also reduce the disciplining effect of competition. A seller with enough data does not need to offer one attractive public price. It can offer each buyer the most profitable version of the deal that buyer is predicted to accept. That can make price competition less visible and less general. Bargains become private. Penalties become private. The public signal weakens.

The same concern applies when the personalization is expressed through discounts instead of surcharges. A retailer can say the public price is stable while privately deciding who receives a coupon, who sees a cheaper bundle, who gets a financing prompt, or who is routed to a higher-margin substitute. Competition law and consumer law therefore need records of the whole offer environment, not only proof that two people saw the same shelf price.

This is one reason the issue belongs in AI governance rather than only consumer advice. The question is not merely whether users should clear cookies or comparison shop harder. The question is whether markets remain inspectable when prices become adaptive outputs of opaque systems.

The Governance Standard

A serious governance regime for algorithmic personalized pricing should meet a higher standard than "the consumer was shown a label."

First, distinguish public dynamic pricing from individualized surveillance pricing. A price that changes because inventory is low is different from a price that changes because the buyer has been profiled. Regulation should preserve that distinction.

Second, require baseline comparison where feasible. If a personalized price is shown, the user should know whether it is higher, lower, or different from a non-personalized reference price. Otherwise disclosure cannot tell the user what happened.

Third, make data categories visible. Location, browsing behavior, purchase history, loyalty status, device signals, inferred income, search behavior, and third-party segments are different inputs. A useful notice should identify the categories that influenced pricing, not merely say "personal data."

Fourth, prohibit sensitive and proxy-based exploitation. Pricing systems should not use protected traits, vulnerability indicators, health status, emergency need, disability proxies, or hardship signals to raise prices or withhold ordinary offers.

Fifth, audit ranking and friction, not only the displayed number. The effective price includes what appears first, what is hidden, which fees arrive late, how discounts are framed, which financing terms appear, what return window applies, and whether the user is pushed away from cheaper alternatives.

Sixth, provide a non-personalized path. A user should be able to buy ordinary goods and services without consenting to behavioral personalization. That path should not be made artificially expensive, slow, or incomplete. This belongs with data minimization, not only price disclosure.

Seventh, preserve records for enforcement. Regulators need access to model versions, data inputs, pricing rules, experiment logs, treatment groups, vendor contracts, human approval records, disclosures shown, reference prices, non-personalized paths, and final effective prices. Without records, individualized prices disappear into session history.

Eighth, audit the vendor stack. The retailer, pricing vendor, data broker, clean-room partner, ad-tech platform, loyalty provider, and payment intermediary can each touch the conditions that shape a price. A serious review belongs in vendor and platform governance, not only in marketing compliance.

Ninth, require impact assessment for consequential pricing. If pricing affects necessities, housing, health, education, transportation, insurance, finance, employment, emergency services, or public benefits, it should trigger an algorithmic impact assessment, disparate-impact testing, appeal channels where feasible, and limits on protected-class proxies.

Tenth, protect agentic commerce. As AI shopping agents compare prices and buy on behalf of users, sellers may optimize for the agent as much as the person. Governance has to cover machine-readable offers, hidden sponsorship, discriminatory agent treatment, delegated-payment receipts, and attempts to infer user vulnerability through delegated shopping behavior. The same issue appears when the payment agent becomes the cashier.

Eleventh, preserve price-session evidence. The record should include the base price, reference price, discounts shown or withheld, fees, shipping terms, financing terms, ranking position, urgency messages, loyalty status, experiment cell, data categories used, model or rule version, vendor involved, and final receipt. Without that evidence, enforcement sees only the final number after the personalization machinery has disappeared.

Twelfth, test with controlled comparison shoppers. Regulators, auditors, and civil-society researchers need lawful ways to compare prices across accounts, devices, locations, consent states, agent credentials, and data profiles. Otherwise individualized pricing becomes difficult to prove precisely because the practice is individualized.

Thirteenth, set stricter rules for necessity markets. In food, medicine, housing, transportation, energy, insurance, credit, education, emergency travel, and other low-choice markets, a system should not use urgency, scarcity of alternatives, health need, disaster context, hardship indicators, or protected-class proxies to raise price, withhold ordinary discounts, or add friction. The governance standard should look more like an adverse-action explanation interface than a marketing experiment.

Fourteenth, keep receipts without building a new dossier. Enforcement evidence should be retained long enough to audit the offer and short enough to avoid becoming another commercial memory layer. Access controls, purpose limits, deletion schedules, and separation from ad-tech and price-optimization systems are part of the safety case, not afterthoughts.

Source Discipline

Surveillance-pricing claims need careful source separation. An FTC 6(b) staff perspective is evidence about an ongoing market study, not a final adjudication that every named firm violated the law. A state statute is binding text, but only within its jurisdiction and scope. A governor's announcement can explain policy intent, but the operative language is in the law. EU directive and guidance materials describe transparency duties, not a general ban on personalized prices. OECD competition reports identify market risks and enforcement patterns, not proof that a particular retailer colluded.

Pending legislation needs the same discipline. New York's existing Section 349-a disclosure rule is enacted law; A9349B/S8623B is a passed bill unless and until the official record shows gubernatorial signature and effective-date status. Likewise, the FTC surveillance-pricing staff perspective should not be described as a completed enforcement case, and the withdrawn RFI should not be treated as an open public-comment process.

Good evidence should identify the jurisdiction, date, sector, product, data categories, baseline price, effective price, and decision point. It should say whether the claim concerns a higher base price, a withheld discount, a personalized coupon, a ranking change, a fee, a financing offer, or checkout friction. Those are different harms and require different records.

The strongest public-interest record combines statute, regulator document, vendor documentation, observed user interface, independent audit, and transaction log. A screenshot alone can show what one person saw. It cannot prove what the system knew, which treatment group they were in, or whether a cheaper non-personalized path existed.

Internal links on this page provide adjacent site vocabulary and examples; they are not evidence for legal status. Legal and current-source claims should rest on the FTC materials, state bill and chapter records, EU directive and guidance materials, and OECD competition report listed below. Current-source claims were rechecked against primary or official sources on June 25, 2026.

What This Changes

The price tag is becoming an interface of belief.

It tells the user what the world costs. It tells the seller what the user might tolerate. It tells the market what demand looks like. When the price is public, that belief can be contested socially. When the price is personalized by hidden systems, each user receives a private theory of what the world is worth and what they are worth inside it.

That is the recursive danger. The system watches behavior, predicts willingness to pay, changes the offer, observes the response, and then treats the response as evidence about the person. A buyer who accepts a high price may teach the system that they are high value. A buyer who searches aggressively may be classified as price sensitive. A buyer in a hurry may be shown urgency. A buyer with fewer alternatives may become more profitable because the model has learned their constraint.

This is model-mediated knowledge at the cash register. The machine does not only answer questions or generate text. It decides what version of the market appears to whom.

The humane rule is simple: people should not have to pass through an unseen psychological auction to buy ordinary goods and services. If a price is personalized, the personalization should be named, bounded, contestable, and inspectable. If a model is using a person's data to decide what they will pay, the person should not be the last party to know.

Sources

Return to Blog