Ads Moved Into the AI Shopping Assistant

A grocery chain turned on an AI assistant last week. Nothing unusual about that. What was unusual sat one layer down: the ads were already in it. Not in a later phase, not in a pilot for select partners, but live on day one, inside the list the assistant builds for a shopper who asks what to make for dinner.

That detail is the story, and it is worth five minutes of any seller's attention even if they have never set foot in the grocery trade. It marks the point where the assistant answer stopped being a neutral summary and became inventory.

What actually happened, and when?

On 28 July 2026, Kroger announced an AI Shopping Assistant across its websites and apps. The described behaviour is ordinary enough: plan a meal, photograph a handwritten list, paste a recipe URL, get a cart. Yael Cosset, the chief digital officer, framed it as another digital convenience.

The advertising part was not in the press release. Modern Retail reported that product listing ads shipped with the assistant, surfaced by Kroger Precision Marketing, the retailer's own media arm. Brands already running product listing campaigns need no extra setup to appear inside the assistant's suggestions. The sponsored items are labelled, and they sit next to the organic ones in the same list the shopper is about to accept.

Compare that with the pace elsewhere. Walmart shipped Sparky, then waited roughly a year before opening sponsored prompts inside it in January 2026. Adweek noted the same rollout followed an earlier test the previous autumn. Amazon's Rufus started carrying ads in 2024. The direction was set; Kroger simply removed the gap between launch and monetisation.

Why does an ad inside an assistant behave differently from an ad on a results page?

Because the shopper never sees the alternatives. A search results page is a list of candidates, and a shopper who dislikes the first three scrolls. An assistant answer is a decision that has already been made on the shopper's behalf: here is your basket for a week of school lunches at your budget. The sponsored line is not competing for attention against nine other options. It is competing against nothing, because the other options were never rendered.

This is why the placement is worth more per impression to the retailer and worth more per exclusion to the seller who did not make the cut. The old failure mode was ranking eleventh. The new failure mode is not existing.

There is a second difference that matters more for margin than for traffic. Retail media inside an assistant is a closed loop. Sparky lets an advertiser sponsor visibility for products it already sells through Walmart. Kroger Precision Marketing does the same for products on Kroger shelves. Nobody is selling an outsider a slot. If your goods are not already in that retailer's catalogue, there is no cheque you can write.

Where does the money actually sit?

Enough to explain the hurry. Kroger Precision Marketing grew profit more than 20 percent year over year in the first quarter while e-commerce sales rose 19 percent, according to the Modern Retail account. Walmart Connect turned over 4.4 billion dollars in 2024 and its global advertising business grew 53 percent in the third quarter of 2025. Amazon's advertising line reached 17.7 billion dollars in a single quarter, up 24 percent.

Those are not comparable businesses, and the comparison is still instructive. Walmart Connect grew 33 percent in the United States while the global figure grew 53 percent, which means the growth is coming from markets and surfaces that did not exist two years ago. An assistant is one of those surfaces, and it arrives with no legacy pricing to defend.

Retail media was already the most profitable thing a grocer did. An assistant is a way to serve the same ads into a moment with far higher intent, and to do it without the shopper feeling advertised to.

Why can a grocer price this so confidently?

Because it can see the receipt. Greg Foran, Kroger's chief executive, told Modern Retail that 95 percent of the chain's transactions are attached to a loyalty card, which lets the retailer measure what was actually bought rather than what was clicked. An assistant recommendation, a sponsored line inside it, and the purchase three days later all land on the same identifier.

That closed loop is the reason a grocer can charge for a placement inside a generated basket without an argument about attribution. It is also the reason the numbers coming out of retail media look better than the numbers coming out of every other channel: they are measured on a population that cannot slip out of the measurement.

An independent shop has a version of this and usually ignores it. Order history plus email address is the same loop at a smaller scale, and it answers the same question. If you cannot say which of last quarter's orders came from customers who had bought before, you are running the channel with the measurement switched off, and no amount of assistant traffic will be legible to you when it arrives.

How far along is any of this, really?

Further than the sceptics think and much less far than the decks claim, which is the usual shape of these things. EMARKETER data published on 21 July 2026 puts 62 percent of global AI users on AI tools for shopping in some form. But only 17 percent of shoppers say they have actually bought a product an assistant recommended, rising to 20 percent among Generation Z. Amazon reports more than 250 million people have touched its assistant and that those who do are 60 percent more likely to buy.

Read those together and the picture is a wide funnel with a narrow spout. Assistants are being used constantly for research and rarely for the final act of buying, and the gap between those two numbers is the whole commercial opportunity that Kroger and Walmart are pricing today.

The click data has the same asymmetry. Retail media clicks in the second quarter of 2026 grew 27.2 percent in beauty and personal care, 27.3 percent in home and garden, 28.9 percent in health, and 150 percent in family and community products. That last figure is not a typo and it is not a trend either. One category moving that fast usually means a surface was switched on inside it, not that shoppers changed.

Assistant ad surfaces compared

The four surfaces below carry the same idea and differ on the thing a seller cares about, which is whether an outsider can get in. Assembled from the reporting linked in this article.

What can a seller outside those catalogues actually do this month?

Feed the assistants that will read a feed. The Agentic Commerce Protocol product feed specification is public and unglamorous, which is a good sign. It takes CSV, TSV, XML or JSON. It wants seller_name and seller_url on every submission, and it makes seller_privacy_policy and seller_tos mandatory once checkout is involved. Feeds are pushed over HTTPS to an allow listed endpoint and re-read as often as every 15 minutes, so stock and price can stay close to live rather than close to yesterday.

None of that is exotic. Most of it is data a shop already holds and describes badly. Variants get named inconsistently, shipping estimates live in a policy page rather than a field, and the price in the feed lags the price on the site by a day. An assistant reading that will either skip the product or state something wrong about it, and a wrong price quoted by a machine is a refund conversation you have already lost.

The practical sequence for a one person shop, in the order that pays back fastest:

1. Make the product data machine readable before making it persuasive. Field level accuracy on price, stock, variant and delivery beats another round of copywriting.
2. Check what an assistant currently says about your best selling product. Ask three of them. The answers are usually assembled from your own pages plus a marketplace listing you forgot you had.
3. Fix the contradiction that turns up. It is almost always stale stock or an old price on a third party listing.
4. Only then look at the paid surfaces, and only in the retailers where you already sell.

If your catalogue images are inconsistent across those listings, that is worth an afternoon too. We covered the mechanics of that in the piece on generating consistent product images without a studio, and the same discipline applies here: an assistant comparing two listings of the same item treats visual and textual inconsistency as two different products.

Does labelling protect the shopper, or the retailer?

Both, unevenly. The sponsored items in Kroger's assistant are labelled, which satisfies the disclosure question and does almost nothing about the selection question. The label tells a shopper that this line was paid for. It does not tell them which unpaid lines were never considered, and that omission is not visible to anyone outside the retailer.

There is a practical version of this question that a seller can answer today without waiting for a regulator. If an assistant recommends your product and gets the delivery window wrong, the customer holds you responsible, not the assistant. The retailer that owns the surface has a contract with its advertisers; you have a refund request and a review. That review then joins the pile that is worth reading for broken expectations rather than scoring for sentiment, because it records a promise somebody else made on your behalf. Every inaccuracy you leave in a feed becomes someone else's confident sentence, and confident sentences are what shoppers act on.

That asymmetry is new. Search engines quoted you; assistants speak for you. The remedy is the same either way and it is unromantic: keep the authoritative fields authoritative, and make sure the place a machine reads them is the same place your checkout reads them from.

This is the part worth watching over the next year, because it is where regulation eventually lands. A ranked page can be audited by scrolling. A generated basket cannot be audited at all from the outside, and the accountability question of who answers when an assistant recommends the wrong thing is still unsettled. Sellers who have thought about how machine written material gets judged on a shop will recognise the shape of the problem.

The three numbers worth tracking from here

Forget the forecasts. Three measurements tell you when this stops being someone else's news.

Share of sessions arriving from an assistant. Most analytics still bucket these as direct traffic or as referrals from a handful of hosts. Separate them now, while the number is small enough to be boring, so that you can see the slope later. A shop that starts measuring at 4 percent knows something a shop that starts at 20 percent has already lost.

Conversion rate on that traffic against everything else. The Amazon figure of 60 percent higher purchase likelihood is a platform statistic on platform traffic, and it will not transfer to your site. Your own ratio will. If assistant traffic converts worse than search traffic, the assistant is describing your product badly and you can fix that. If it converts better, you have found the cheapest acquisition channel you own.

What a placement costs when the surface opens. Closed retail media does not stay closed forever; it opens once the inventory exceeds the demand from existing advertisers. Watch for the first grocer or marketplace that sells assistant placements to sellers outside its own catalogue. That is the moment the economics of this become a decision rather than a spectator sport.

What this changes for a small shop, honestly

This month, almost nothing. No independent seller is losing revenue today because a grocery assistant in the United States started carrying sponsored listings. The reason to read it is directional. Two of the largest retailers on earth have now decided that the generated answer is the ad unit, and the gap between launching an assistant and monetising it has collapsed from a year to zero.

Whatever comes next will inherit that assumption. When an assistant surface opens to sellers outside a retailer's catalogue, it will arrive already priced, already labelled and already optimised for the retailer that owns it. The preparation that helps is not budget, it is being the kind of shop whose data an assistant can read without guessing: accurate stock, honest delivery windows, one description per product rather than four. That work is dull, it is cheap, and it is the same work that made a shop legible to search engines fifteen years ago.

If you are building that shop from scratch rather than retrofitting one, the structured product data question is easier to solve at the start than in year three. Our AI store builder generates the catalogue schema and the storefront together, and you keep the code, which means the feed you hand an assistant is one you control rather than one a platform exports on your behalf. Sellers weighing that against a hosted platform will find the trade offs laid out on the pricing page.

The uncomfortable summary: discovery is moving into a box you do not own, the box is already selling advertising, and the only lever available to most sellers is the quality of the data they hand it. Start there.