A shopper asks an assistant for a waterproof jacket for cycling to work that packs small enough for a bag. The answer names three jackets, gives a reason for each, shows prices and links to buy. Your jacket is lighter, cheaper and genuinely better for that job, and it isn't in the answer. Assistants don't browse your store and form an opinion. They assemble an answer from what the web says about products, so being named is about being findable, legible and corroborated.
An assistant answers by searching, not by remembering
The model behind an assistant doesn't hold your catalogue, and for anything with a price it can't rely on what it learned in training. So a shopping question triggers a search. Google documents the technique for its AI Mode as query fan-out: the question is broken into sub-questions (what matters in a commuter jacket, which ones are waterproof rather than water-resistant, which pack down, what they cost, what reviewers say), each is searched, and the results are read together. Other assistants describe searching the web before answering and cite the sources they read. The details of how they choose those sources aren't documented, so what follows about them is observed from their answers, not taken from a specification.
The pages read for a shopping question are a mix: product pages, retailer listings, roundups and buying guides, forum threads, and reviews. From those the assistant extracts product names, specs, prices and what people say, then writes an answer naming a few. How AI assistants choose what to cite covers the general mechanism. For products, the practical consequence is three conditions a product has to meet.
It has to be retrievable: the assistant's crawler can fetch the page. It has to be legible: the page states what the product is, who it's for, its specs and its price as text the model can read. And it has to be corroborated: other sources agree it exists and say something about it. Miss any one and the product is invisible to the answer, however good it is.
Product cards are built from data, not prose
Alongside the written answer, some assistants show product cards: an image, a price, a retailer, sometimes a rating, and a link to buy. In some products and markets a shopper can complete the purchase inside the chat. Those cards aren't written by the model. They're assembled from structured product data, which is why the previous lesson matters here.
Google says its Shopping Graph, the product database behind shopping results across Search including its AI features, is built from Merchant Center feeds and crawled pages. Other assistants build their product data from Product structured data on pages and from retailer listings, and some accept product feeds from merchants directly. The feed specifications change, so check the assistant's own merchant documentation rather than any summary, including this one.
What you control is the same in every case: a clean feed, complete Product markup with price, availability and identifiers, and a visible page that agrees with both. Product feeds and free listings covers the feed and Product schema and rich results covers the markup.
What makes a product eligible to be named
Across assistants, the products that get named share a pattern, and you can observe it yourself by asking the same question in several assistants and reading the sources they cite.
The product appears in several independent sources that agree on its name and its key specs. A product that exists only on its own store page has nothing corroborating it, and it rarely gets named.
The product has a clear "who it's for" statement that matches a sub-question. "Packs into its own pocket, cut for a riding position" answers the commuter's question. "Premium performance outerwear" answers nothing.
The product has recent third-party opinion attached to it: reviews on the page, a mention in a roundup, a thread where someone recommends it. The next lesson, Reviews, UGC and brand signals, is about earning that.
The product is in stock at a stated price. An assistant checking live data skips what it can't price.
Being named and being linked are different outcomes. An answer can name your jacket and link a marketplace listing or a review rather than your page. You can't fully control which link it picks, but the lever is clear: make your product page the best source of the product's facts, so that when the assistant needs the spec, the price and the availability, yours is the page it read.
Measure before you change anything
Two kinds of AI surface matter to a store, and they need different checks.
Google's AI Overview and AI Mode appear for some shopping searches and not others. Comparison and "best" searches often have one. Many searches for a specific product don't, and a missing AI surface there isn't a problem to fix. Track the searches where the surface exists and note whether the sources it cites include you.
The assistants outside Google offer no tracking data you can pull. Ask them the question yourself, once a month, in the same words, and record which products are named and which sources are linked. Answers vary between sessions, so look for who appears repeatedly rather than reading one answer as a verdict. Measuring your AI visibility covers how to do that without fooling yourself.
What to take away
- Assistants answer shopping questions by searching the web for sub-questions and reading product pages, roundups, reviews and forums, then naming a few products from what they found.
- A product is eligible when it's retrievable, legible as text and data, and corroborated by sources beyond your own store.
- Product cards come from structured product data, so the feed and the Product markup decide whether you appear in them.
- Being named and being linked differ, and the way to earn the link is to be the best source of the product's facts.
Next
The corroboration assistants look for comes from reviews and from what customers and third parties say, in Reviews, UGC and brand signals.