Read this first
This is the short form of a long research paper. The full paper is called Receipt Required: The State of Retail Media in 2026. It has the sources, the tables and the math. This page has the same findings in fewer and shorter words. The counts come from the same data file as the paper. The other numbers are copied from it. Each part links to the place in the paper that backs it up.
The research runs to 3 November 2026. That is when facts that can change, like who owns a firm, were checked. It looks mostly at the US.
What retail media is
Big stores and shop apps now sell ads. Amazon, Walmart, Target, Kroger and Instacart all do. The ads show up on their own sites and apps. They also use what they know about their shoppers to place ads elsewhere: on other sites, on TV and on screens in their stores. Banks, card firms and travel firms sell ads this way too.
Their pitch is the till. A store knows what you bought. So it can tell a brand: your ad ran, and then these people bought your goods. That is the result the brand gets back. By default, the store works it out.
A shop receipt lets you check what you paid for. This paper asks the same of an ad result. Who made the number? What kind of number is it? Can the buyer check it? Has anyone outside checked it? The paper does not test if any number is right. It tests what each firm shows in public.
The paper keeps two kinds of result apart:
- Matched sales: a shopper saw or clicked an ad, then bought the goods. The store counts that sale for the ad. No one asks if the shopper would have bought it anyway.
- Lift: the store keeps some shoppers from seeing the ad, then compares them with those who saw it. If the test is sound, the gap estimates what the ad caused. The guides the trade wrote say to pick the two groups at random, like a coin toss, when you can.
Stores also give a ratio called return on ad spend: matched sales for each dollar spent on ads. It is not profit. Chapter 1 sets out the terms.
What we found
1. No one can say how big it is in the US
One trade report puts US ad sales of this kind at $63.4 billion in 2025. But it comes from a survey, and partly from estimates. And it does not split ads on the store's site from ads elsewhere or in stores. A year before, the same numbers had a different name. Firms report their own ad sales in ways that don't line up. Amazon's ad sales alone are bigger than that US total. But they cover the whole world, and they include ads sold to publishers and authors. So the numbers can't be added up or ranked. Chapter 2 has the numbers.
Who owns what keeps changing too. Since the start of 2023, eleven deals that changed who owns a firm have closed. The last was on 2 November 2026, when Publicis closed its purchase of LiveRamp. Three more were agreed or announced, with no dated close by 3 November 2026. In one, DoorDash agreed on 28 October 2026 to buy Instacart. The deal is due to close in late 2027 or early 2028. Chapter 2 has the deals.
2. The store counts matched sales, not lift
We looked at 49 ad services run by stores and online shops. 37 of them say what their standard report gives. 36 of those 37 give matched sales. Just one gives lift as a rule: the screens in Albertsons stores. Even there, it does not say how the test stores were picked.
The rules also differ from store to store. Amazon counts a sale up to 7 days after a click for some sellers, and up to 14 days for others. Walmart changed its rules on 1 June 2026. It did not redo its old numbers. So old and new do not line up. Chapter 3 has the details.
3. Lift is an extra, and few say how the groups were picked
22 services offer a lift test. At one, the screens in Albertsons stores, lift is the standard report. The rest sell it as an extra. 21 of the 22 do not say how they picked who sees the ad and who does not. Just one says it picks at random: Instacart. And it says so only on pages from 2022 and 2023. Only two state a rule for who can get a test. None of the sales tests says how many shoppers it needs, or how sure a result must be. Chapter 3 shows how lift is sold.
4. The seller makes the number, and may bill on it
Where a service shows how it reports results, the firm that sold the ads also makes the number. It uses its own rules. No service says it keeps the team that sells the ads apart from the team that counts the results.
At some firms the count is also the bill. Amazon's ad contract says its fees are "based solely on our measurements." Rules for fixing a wrong number are the seller's own too. Amazon updates its sales counts 1, 7 and 28 days after a sale. Only 7 of 69 services could be scored on how they fix and change their numbers. None scored above 3 out of 5. For most, we could not read the terms they sign with buyers. Chapter 4 has the details.
5. Of six big services, only Amazon lets the buyer run its own sums
We looked closely at six big services. At Amazon, a brand that buys ads for its products in search can write its own queries in Amazon Marketing Cloud. That is a "clean room": a place where data can meet under set rules. At Amazon's, only totals can leave, not a record of each shopper. And the sales it sees there are still Amazon's matched sales. So it still can't tell what the ad caused.
None of the six says it gives a record of each time an ad was shown. Not to the buyer, and not to an outside firm the buyer hires. At the other five, the buyer gets the seller's totals, in a dashboard, a file or a data feed. Chapter 5 has the details.
6. Ads away from the store: most don't say how they link the ad to the sale
Stores also sell ads on other sites, on social media and on TV. Then the ad and the sale happen in two places. Someone has to link them.
13 store ad services say they measure results for these ads. Only 3 say the ad is linked to the sale through a match made outside the store's own site, such as in a clean room. All three name the same firm, LiveRamp. None says what it matches on. Two more, Kroger and Target's Roundel, name a clean room but not a sale. The other 8 do not say how they link the ad to the sale. This test is about what the services write down in public, not what they do. To back the claim that most services document such a match, the test needed 9 of 13. To reject it, it needed 4 or fewer. At 3, the claim is rejected. Read more strictly, the count falls to 1 or 2.
We screened six market reports. None we could read splits the money by where the ads run. So we can't tell if ads away from the store's site are a growing share.
Screens in stores have their own gaps. Walmart says its counts of who saw a screen are estimates, and it does not show how it makes them. Albertsons compares stores with screens to like stores without them. But it does not say how it chose them. Chapter 6 has the details.
7. The US ad-count body has approved counts of ads, not sales
The Media Rating Council (MRC) is the US body that checks how ad numbers are made. On 3 November 2026, seven ad services in this field held its full approval. All seven approvals cover counts of ads shown and clicked, not sales.
One entry on its list, named just "Epsilon", covers a kind of sales link. The list does not say which Epsilon business it is. We think it is not Epsilon's store ad unit. But that is a guess, based on a 2021 press release. So for now this test can't say yes or no. If the guess is right, no approval covers a sales number.
Kroger's sales numbers are at an early stage of MRC review. That is not a pass. Another body, the Alliance for Audited Media, approved how DoorDash measures lift in September 2026. It does not say if DoorDash picks its groups at random. Chapter 7 has the record.
8. The best tests were not run on store ads
Firms like Facebook and eBay have run fair tests with random groups. Then they tried other ways to measure the same ads. The other ways were often far off, and most often too high. In one set of 15 tests, they were off by about three times in half of them.
We looked for studies from 2024 to 2026 that measure what a brand gains from a US store ad network. We wanted ones with a stated design, reviewed by experts for a journal. In our searches, we found none. The store ad studies we found were written by the networks, or with them. In one, by Albertsons' ad unit and a partner firm, the groups were matched in different ways. For the same campaign, the highest and lowest results were 6.5 times apart on average.
We also looked at 13 results that stores and firms show to buyers. Just one says the groups were picked at random. None gives a range for how sure the result is.
A US group of advertisers ran two surveys of its members, in 2022 and 2023. They asked if the stores sway the choice to buy these ads. Many said yes: 88% in one and 63% in the other. The questions were worded in different ways, so the two do not show a trend. Chapter 8 goes through the studies.
9. The tools under the stores show the same
Many stores rent their ad tools from tech firms. Those tools let a store set how many days after a click a sale still counts. Just one tool, from Criteo, says what the standard setting is. Just one, from Moloco, says it can run a random test for a store's brands. It says so only in a blog post with no date. Seven firms sell tests of store ads, and none says its standard test picks the groups at random. Chapter 9 has the details.
10. Privacy law lets shoppers say no, and the watchdog looks at prices
US state laws let shoppers say no to having their data sold or shared for ads. For data that shows health, like a medicine you bought, many states need a yes first. Maryland goes further: such data can be used only when it is truly needed for what the shopper asked for.
The FTC is the US trade watchdog. Its look at store data has been about prices, not ads. On 28 October 2026 it told nine firms to report if their prices change from one person to the next. Instacart, Target and Walmart are among them.
In the searches we ran, we found no order, fine or court ruling about a store using shopper data for its ads. We found no breach notice tied to it either. The privacy notices of Amazon, Walmart, Target and Instacart do not name their ad services. Chapter 10 has the rules.
11. Ads inside AI helpers report back to the seller
AI tools like ChatGPT, Google's AI Mode and Amazon's own shopping chat now carry ads. We found 11 kinds of ads in these tools in the US. Just two are said to be fully launched: two kinds of "prompt" ads in Amazon's store, which can open its shopping chat. Amazon reports their sales itself.
None of the 11 says an outside check is fully launched for it: not of its sales, not of what it caused, and not of its quality. But most say nothing about this either way. So we can't say that no one does it.
AI tools talk to each other through shared rules. We read five sets of these rules. One, AdCP, can describe an ad like this and its result on paper. One can do it in part. None has yet been shown to work for such an ad between two real firms. The author helps lead a work group at the group that runs AdCP. Chapter 11 has the details.
How the services scored
We scored 91 services on 12 things a buyer needs to know. Each got 0 to 5, based only on what the firm shows in public. We put them in five groups, so like is only matched with like. The total is called the Receipt Quotient.
2 services show the most in public. One is DoorDash's ads for brands of things like food and soap. The other is Moloco, a firm that runs ad tools for stores. Each owes its top spot to one thing: an outside body's sign-off. For DoorDash it is the check of how it measures lift. For Moloco it is a European trade body's check of what its tool can do. If the gaps in their records were filled with low scores, both would drop a step. Moloco's sign-off does not list the numbers it covers. Scored one step lower, Moloco would drop a band. Without the sign-off, neither would have enough to get a band.
Most services show little on who checks their numbers, how they fix them, or how they treat shopper data. 48 services had records too thin to place in a band. Of those, 44 showed nothing we could count on any of these three things.
A high score means a firm shows more of its work. It does not mean its numbers are right. Chapter 12 has the scores.
What to ask before you buy
Ask what the report counts, and for how many days after a click or a view. Ask if lift was tested, and how the groups were picked. Ask who made the number, when it stops changing, and how you can challenge it. Ask what you can check yourself, and what can leave the data room. Ask what has been approved, by whom and for which number. Ask how an ad seen away from the store was linked to a sale. Ask what an ad inside an AI tool reports, and who else can check it. Chapter 13 turns this into seven questions.
How this was made
AI agents did the research under one set of rules. The rules were fixed before the work began. Some coding rules were added later, as the work showed gaps. Each one is dated and listed in the paper. A second AI checked each file and tried to prove it wrong. Other AI agents scored the services, and a separate AI checked every score. Only scores that passed that check count. 737 of the facts in the paper were checked this way. Facts no checker reached are marked as not settled.
AI helped make this work, and the method is public. Before it came out, a second AI, from another firm, checked some facts at their source. It also tested the scoring rules and the print copy. We checked each point it raised. No outside expert or group has audited this work. No journal has reviewed it.
The author works at Samba TV, a firm that measures TV ads. Amazon and Walmart list Samba TV as a partner that can measure their ads. He also helps lead a work group at AgenticAdvertising.org. That group runs AdCP, one of the sets of rules for AI tools. He used to work at Verve Group, Hearts & Science and Lotame. Hearts & Science was part of Omnicom, which owns two firms in this paper, Flywheel and Perpetua. Services tied to these links carry a mark. The paper also shows the results without them. Three services then move up a band, and one test can no longer give an answer. The marks never change a score. No firm paid for, read or saw the paper first. Every service is scored by the same public rules. Chapter 14 has the method.