What the Instacart Case Actually Shows
The Instacart episode is useful because it puts personalized pricing in front of ordinary shoppers, but it does not answer every question that the screenshots seem to raise. This chapter reconstructs what the New York Attorney General said was observed, what Instacart said about its pricing tests and promotions, and why the state asked for deeper records. The aim is to separate a documented price pattern from a proven data-to-price mechanism and from a final legal finding.
A striking pattern, with a careful attribution
The episode began with a report by Groundwork Collaborative and Consumer Reports, released on December 9, 2025. The New York Attorney General’s January 8, 2026 release summarized the report as an experiment involving 437 shoppers in four cities. The shoppers added items to Instacart carts and compared the prices displayed to them. According to the Attorney General, nearly three-quarters of the grocery items in the experiment appeared at multiple prices. Some shoppers saw prices up to 23 percent higher for the exact same product in the same store at the same time. Across a single good, the average difference between the highest and lowest observed prices was 13 percent.
Those are large enough differences to make the issue more than a technical debate. The Attorney General said the report estimated that the gap could translate into as much as $1,200 a year in additional costs for a typical family of four. That estimate is a projection from the reported experiment, not a measurement of every family’s actual annual bill. The proper claim is therefore that the reported price differences could have a meaningful household effect, not that every family using Instacart paid that amount.
The Attorney General’s letter supplied another detail: the report observed as many as five different sale prices for the exact same grocery item in the exact same store at the exact same time. It also reported variation in the “original price” shown to shoppers and in the discount applied to reach a sale price, even when shoppers received the same sale price. This distinction matters. A different final price changes what a shopper pays. A different original price can change what a shopper believes they saved. The two outcomes may share a cause, but they are not the same measurement.
Reference-price differences can also complicate a consumer’s ability to detect a problem. If one shopper sees a sale price of $8 with an original price of $10 and another sees the same $8 sale price with an original price of $12, their immediate payment is identical, but the claimed discount is not. If the two sale prices are different as well, the observer must separate the effect of the underlying price from the effect of the promotional framing. A study that records only the final charge may miss a difference in how the offer was represented; a study that records only the advertised discount may miss the amount actually paid.
The distinction is not a technical excuse for ignoring the pattern. It is a way to identify the consumer consequence precisely. Different final prices may change access or household cost. Different reference prices may change perceived value and comparison behavior. Both can matter to a regulator, but the evidence needed to connect either result to personal data and an algorithm is different.
The record also contains a question about what the shoppers were actually comparing. A controlled assignment to different test cohorts can produce unequal displays without the firm using personal data to estimate individual willingness to pay. A loyalty offer can be account-specific without changing the underlying base price. A retailer or brand can fund a promotion while a platform controls how it is displayed. A delivery fee can reflect an address or local condition. These possibilities do not disprove personalized pricing; they show why the observed gap cannot identify its cause on its own.
The same-store, same-time comparison is valuable because it holds important context more constant than a casual comparison between two shoppers in different places. It still does not reveal the entire pricing system. The comparison tells us that the displayed outcomes differed under closely matched conditions. It does not tell us which inputs were available to the platform, which partner set each component, whether a price experiment was random or targeted, or whether personal data changed the amount.
The denominator matters as much as the headline. “Nearly three-quarters of the grocery items” refers to the items observed in the reported experiment, not to three-quarters of all products on Instacart or to three-quarters of every shopper’s annual purchases. “Up to 23 percent higher” describes the largest reported difference, not the typical difference. The 13 percent figure is an average gap between the highest and lowest prices for a single good in the report as summarized by the Attorney General. Keeping those units separate prevents a striking result from silently becoming a broader prevalence claim than the evidence supports.
The comparison also does not tell us whether the price was set by Instacart, a retailer, a food brand, or a combination of systems. A platform may display a retailer’s base price, apply a promotion funded by a brand, calculate a delivery-related charge, and present the result through one interface. The shopper experiences one checkout, but responsibility for the inputs and outputs may be distributed. That is why the case cannot be resolved by asking only whether “Instacart” changed a number. The relevant question is which entity or tool controlled which part of the price.
The chronology is important because hindsight can make the case look more settled than it was. The report established a pattern that deserved explanation. Instacart’s statements described one part of the response. The Attorney General’s letter then sought evidence about the systems behind both the old tests and the practices that would remain. At the point covered by the supplied record, the central causal question was still open.
What Instacart said, and what the state needed to know
Instacart’s response, as quoted in the Attorney General’s letter, had two parts. The company said it had been randomly assigning customers to price-testing cohorts. It also said that its retail partners might use behavioral data to inform discounts or promotional offers. Random assignment and behavioral targeting are not interchangeable. The first describes how customers may have been allocated to an experiment; the second describes a possible use of consumer behavior in setting an offer. A complete account would need to show which practice applied to which price component and at what point.
There is a further distinction between using behavior to learn whether a promotion works and using behavior to decide which price a particular shopper receives. In the first case, customers can be placed into groups by a random process and the business can compare group outcomes. In the second, the business can use a customer’s observed characteristics to predict a response and then tailor the offer. Both practices can be automated and both can be described as testing, but the personal-data connection and the consumer’s expectation may differ. A platform’s public description of randomized cohorts therefore does not resolve whether another stage used personal information.
The roles of the platform and its partners matter for the same reason. The letter says retailers and food brands could influence prices, discounts, offers, or promotions through agreements and tools. A customer may reasonably see one Instacart interface without knowing which partner supplied the price or which system generated the offer. That lack of visibility makes it difficult to assign responsibility from the screen alone, and it explains why the Attorney General requested agreements as well as technical descriptions.
The letter connected the issue to Instacart’s Eversight platform, which the company promoted for revenue optimization and automated, continuous testing of prices with shoppers. It also mentioned the Caper Cart smart cart and its capabilities in gamification, personalization, and advertising. These descriptions establish that Instacart and its partners had tools capable of changing how prices or offers reached shoppers. They do not, by themselves, establish that a particular tool used personal data to set a particular grocery price.
On December 22, 2025, Instacart announced that it was ending all “item price tests” on its platform, effective immediately. The announcement said retailers would no longer be able to use Eversight technology to run item price tests through Instacart. But the company also said that retail partners would continue to set their own prices and that Instacart would continue offering promotions, discounts, marketing incentives, and loyalty integrations. The letter further quoted reporting that grocery retailers and food brands could still test different types of promotions and discounts through the platform.
That combination makes the word “ending” narrower than it might first sound. It describes the discontinuation of a category the company called item price tests. It does not mean that every kind of price variation, discount experiment, loyalty offer, or automated pricing tool disappeared. Nor does it establish that the discontinued tests were legally the same as personalized algorithmic pricing. Instacart said those tests were not dynamic pricing or surveillance pricing. The Attorney General treated that characterization as something to examine, not as the last word.
The Attorney General’s own concern focused on disclosure as well as price variation. The letter said that, on some parts of the platform, a notice could be reached only by clicking fine-print text such as “View pricing policy,” “Higher than in-store prices,” or “Pricing & fees.” The notice described randomized tests, personal information used to calculate fees, and personalized incentives, but the Attorney General said it was not clearly displayed near prices. The letter also said there was no disclosure on certain category pages or individual product pages displaying prices. Whether that assessment ultimately establishes a statutory violation is not resolved in the supplied materials. It does show what “visibility” means in practice: a notice hidden on another page may not reach a shopper at the moment they decide whether the displayed number is comparable to someone else’s.
To investigate the mechanism, the Attorney General requested several kinds of information. The request covered agreements with retailers and food brands, descriptions of Eversight and Caper Cart, records of price experiments, the dates and selection rules for those experiments, the number of affected customers, and the variables adjusted. It asked whether the change involved a sale price, original price, discount, or promotion; how personal data was used; and what happened to product and basket prices for individual shoppers and for the affected group. It also sought a statement about whether Instacart or its partners used dynamic or personalized algorithmic pricing in New York, the categories of data involved, an explanation of how an algorithm set the price, and screenshots and screen flows showing disclosures from the first display of pricing information.
This is a roadmap for moving from suspicion to proof. First, investigators need to identify the relevant price or offer. Then they need to identify the actor and tool that controlled it. They need to know whether customers were assigned randomly or selected using personal information, and whether the data affected the output or merely accompanied the transaction. Finally, they need to see what the consumer was told before deciding. Each step can change the classification.
The request for experiment dates and affected-customer counts addresses scope. A price difference observed once may be an error or a local event; a repeated pattern across a defined cohort may reveal a systematic rule. The request for adjusted variables addresses the object of the test: changing a base price, a discount, an original price, or a promotion can produce different economic effects. The request for individual and basket-level outcomes addresses incidence: a small item-level difference may accumulate across a basket, while a visible item difference may be offset by a promotion elsewhere.
The request for screen flows addresses timing. A disclosure that appears before a customer begins browsing is not equivalent to one that appears only after a price is selected, and a notice on a store landing page may not reach a shopper on a category or product page. The Attorney General’s concern was therefore about the relationship between the notice and the act of choosing, not simply whether some text existed somewhere in the application.
For example, suppose the same milk item appears at two prices. If a retailer randomly assigned shoppers to two experimental prices, the result is a price test. If a loyalty account receives a discount based on a program’s rules, the result may be a personalized offer, but the relevant legal and economic analysis still depends on how the rule uses data and how the offer is displayed. If the price changes because a local store’s inventory or delivery conditions change, the cause is different again. If the platform predicts that a particular shopper is less able or willing to switch and raises the price on that basis, the personal-data link becomes central.
The official record does not tell us which of these explanations accounts for every Instacart observation. It does not include the underlying Groundwork Collaborative and Consumer Reports report, Instacart’s response to the January letter, the requested agreements or model records, or a later enforcement decision. The absence of those materials is not evidence that no personal-data pricing occurred. It is evidence that the supplied record cannot settle the question.
Several kinds of new evidence could change that assessment. Pricing logs could show that a customer-specific field entered the decision that generated a particular amount. Experiment documentation could show that the same difference came from random cohort assignment without personal targeting. Partner agreements could identify whether a retailer or brand, rather than the platform, supplied the relevant rule. Screen captures tied to the same transaction could show whether a disclosure was visible before the choice. None of these records would automatically decide the legal question, but each would close a different gap between observation and explanation.
What the case does establish is narrower and more useful. An official state release reported substantial price variation under closely matched shopping conditions. Instacart acknowledged price variation, described randomized testing and possible behavioral-data use by partners, and announced an end to item price tests while preserving other pricing and promotion practices. The Attorney General then sought the records needed to connect a displayed price to an algorithm, personal data, and a disclosure. That is the difference between seeing that prices differ and knowing why.
The next comparison is with a study designed to answer a different question. Instead of beginning with screenshots and an unresolved platform mechanism, researchers working with ZipRecruiter deliberately varied prices, modeled heterogeneous demand, and tested their recommendations. That controlled setting cannot prove what happened at Instacart, but it can show why a firm might have an incentive to personalize prices in the first place.