AI shopping assistants sell. Your feed decides if you're in

Three kinds of evidence that keep getting mixed up

The retail trade press has spent the summer running assistant figures, traffic figures and consumer surveys together in the same paragraph. They measure different things and only one of them tells you anything about your own catalogue.

Start with the loudest. On the Williams-Sonoma Q2 FY26 call, for the quarter ended 2 August 2026, chief technology and digital officer Sameer Hassan said engagement with Olive, the assistant on the Williams Sonoma site, was up 700% since the start of the year, with revenue associated with the assistant up 620%, and that shoppers who use it convert at three times the rate of those who do not. Read that carefully before you quote it. Those are spoken remarks: none of the three figures appears in the company's Form 8-K earnings exhibit, which reports total net revenue of $1.96bn, up 6.7%, and never mentions the assistant. Olive launched in 2025, so 620% is growth off a young base with no denominator attached. And the three-times comparison sets shoppers who chose to open a shopping assistant against everyone else. People who open a shopping assistant were already further down the aisle.

The second kind of number is more useful, because it is about data rather than about one retailer's product. Shopify president Harley Finkelstein told the Q2 2026 call that "AI searches powered by Catalog converted twice the rate of those using scraped data". Same shoppers, same engines, same merchants. The variable is whether the platform handed the machines structured product data or left them to read the page. He added that conversion from AI search runs nearly 80% higher than traditional organic search, and that half of all AI-referred sessions land directly on a product page, 2.5 times the rate for traditional search.

The third kind is market-wide traffic data. Adobe, whose retail analysis covers more than a trillion visits to US retail sites, found that in July 2026 AI-referred visits converted 60% better than non-AI traffic, the eleventh consecutive month AI traffic came out ahead. In March 2026 the gap was 42%, and in March a year earlier the same comparison ran 38% the other way, with AI traffic converting worse than everything else. The reversal is the story, and it matches what we found looking at how much less traffic AI sends and how much more each visit is worth, and again in Swedish e-commerce, where AI referrals now out-convert Google. Two cautions. Adobe's figures move month to month, so quote the month with the number. And the panel is US retail: we have found no equivalent Irish, UK or EU dataset.

Small, unusually valuable, growing fast

Etsy was the flagship launch partner for OpenAI's Instant Checkout. Its chief executive Kruti Patel Goyal told the Q2 2026 call that "traffic from agentic experiences is still less than 1% of our overall traffic", while adding that it is higher-intent and higher-average-order-value traffic. That is the honest denominator for everything above. Shopify said its AI-driven traffic and orders each tripled year on year while the volume stays small relative to its overall GMV. Target said its digital traffic from external AI platforms is growing more than 3.5 times the industry rate, and led with the words "while still small in total today".

The plumbing is churning too. OpenAI retired the original Instant Checkout in March 2026, saying it "did not offer the level of flexibility that we aspire to provide", and sent merchants back to their own checkouts. Emily Pfeiffer, a principal analyst at Forrester, said that as of February 2026 roughly 30 Shopify merchants were available via Instant Checkout, against a launch promise of more than a million coming soon. Walmart said it found conversion three times lower for products sold directly in ChatGPT than for shoppers rerouted to a retailer's own site for checkout. Buying inside the chat window is the part of this that is still being argued about.

Geography narrows it further for anyone reading this from Galway, Manchester or Berlin. Google's UCP-powered checkout applies only to products with eligibility in the United States, Canada and Australia, and only for participating merchants and partners. Microsoft's UCP-ready feeds are generally available in the US. Adobe's conversion premium is measured on US sites. None of this is a switch an Irish or UK merchant can flip today. What does travel is the preparation, because the preparation is identical everywhere and it is cheap.

Every one of these programmes runs on a merchant feed

This is the part that gets skipped. Google's Universal Commerce Protocol was co-developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than 20 others including Adyen, Stripe and Zalando, and it is published as an open standard. It reads your Merchant Center product data. Google says so directly: Merchant Center "will continue to be the central hub to prepare your product data", and the Buy button in the new checkout experience appears only for product listings using the native_commerce(checkout_eligibility) product attribute. The UCP developer docs add that you must update your product feed to signal eligibility and supply compliance data, that a missing eligibility value defaults to ineligible, and that return policies including cost and window must be defined in Merchant Center because they are a Merchant of Record requirement shown on the checkout screen.

OpenAI went the same way, and its March 2026 announcement was explicit about it: "Through ACP, merchants share product feeds and promotions so their catalogs are fully represented in ChatGPT." The current ACP documentation describes the protocol as a layer that "enables ChatGPT to ingest structured catalog data, understand merchant inventory, and surface relevant products in context". Its feed spec is nine required fields per row and a blunt rule for the one that goes stale fastest: "Omitted, empty, or unrecognized values reject the row." Microsoft named the mismatch that catches most retailers out: most product feeds were built for advertising, sharing a subset of the catalogue for campaigns, while an agent needs the whole thing.

Here is what each programme actually asks for, built only from the platforms' own documentation.

Markup is not the entry ticket, and Google says so

Not one of those rows names schema.org, and plenty of agencies are currently selling Product markup as the way into AI Mode and Gemini. Google's own documentation contradicts them in one sentence: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add" (Search Central). The UCP developer docs, ucp.dev and OpenAI's feed spec contain no schema.org requirement either. If someone has told you that adding JSON-LD gets your products into an assistant, they are selling you something.

What markup actually does is quieter and more useful. On that same Google page, among the fundamentals that still help with AI features, is "making sure your structured data matches the visible text on the page". Google's Product structured data guide says that "providing both structured data on web pages and a Merchant Center feed maximizes your eligibility to experiences", and that some experiences combine data from both when both exist. Its ecommerce guide is more specific again: structured data improves the accuracy of Google's understanding of price, discount and shipping, and helps Merchant Center verify your feed against your site. For smaller sites updated less often, Google can build an automated feed from crawled web content, and says structured data improves the accuracy of that extraction. For a small retailer with no feed engineer, your markup can become your feed.

So the audit to run is a consistency audit. Be precise about what Google actually requires, too, because overclaiming here is how technical readers stop trusting you. For a Google merchant listing only name, image, offers, and price and priceCurrency on the Offer are required. Availability, hasMerchantReturnPolicy, priceValidUntil, shippingDetails and gtin are recommended. The concrete failure to go looking for is on the same page: your listing may not display if priceValidUntil carries a past date. That is a one-line fix that nobody notices for a year.

Which brings it back to the page the agents land on. Adobe scored how machine-readable US retail pages are, and product pages came bottom of every page type at 66%, below homepages at 75% and category pages at 74%. That is Adobe's own proprietary metric rather than a schema validity score, so treat it as direction rather than measurement. Put it next to Shopify's finding that half of AI-referred sessions land straight on a product page and the shape of the problem is clear: the worst-prepared page type is the one the machines arrive on. We ran the same test by hand on an Irish brand and found zero schema.org markup across 50 crawled pages of barrystea.ie, which is more common than the industry likes to admit.

What you actually do

Six checks. All of them are about your feed and your page saying the same thing about the same product on the same day.

1. Make the feed complete before you make it clever. If your feed was built for shopping ads, it probably carries the products you bid on rather than everything you sell. Microsoft's point stands: agents need the whole catalogue. Every product you sell should have a row.

2. Diff ten products by hand. Take ten SKUs across your range, and put three things side by side: the feed row, what the product page displays to a human, and the JSON-LD in the page source. Price, availability, currency, title and image. Any disagreement between the three is the actual bug, and you will find one within the first ten.

3. Treat availability as the field most likely to lie. It changes hourly and it is the field OpenAI rejects a row over. If your feed refreshes nightly and your page is live, they will disagree during the day. Decide how often the feed refreshes and know what the gap is.

4. Check priceValidUntil is in the future. Google says a past date may suppress the listing. Grep your templates for a hard-coded year.

5. Fill in returns and support in Merchant Center now. Google requires return policies with cost and window plus a customer support contact method before UCP checkout is available at all. You need a returns policy anyway. Putting it where Google reads it costs an afternoon and removes a blocker before the programme reaches your market.

6. Confirm the price is in the HTML, not painted in afterwards. If price and stock arrive by JavaScript after load, a crawler reading raw HTML may see a page with no price on it. Your visible text and your structured data both need to be there when the machine reads the page, which is Google's matching instruction applied literally.

All of that is catalogue hygiene rather than speculation about a channel that might arrive, and it pays in ordinary shopping results, in Merchant Center approval rates and in every agent that shows up later. The cheapest way to start is to see what your product pages currently say to a machine.

Audit your product markup

Questions people ask about this