An AI assistant recommending a product is not reading your brand story. It is reading fields. Here is what those fields are, which ones are required, and what happens to a product whose facts a machine cannot confirm.
Some brands we cover are clients of our publisher. What we recommend is decided by our editors, and prices are read on the date noted in each article. Our standards.
The quick take
- When a shopper asks an assistant for a recommendation, the deciding layer is the part of your page a machine can parse without guessing: a product name, an image it can crawl, and a nested offer with a price, a currency and an availability value.
- Those three are not a wish list. Under the documented requirements for merchant listings,
name,imageand a nestedoffersobject are required properties, and a merchant listing specifically requires anOfferrather than an aggregate range. - Google states plainly that there is no special markup that makes a page eligible for AI Overviews or AI Mode. That is not permission to skip structured data; it is a reminder that the win comes from stating facts clearly, not from a secret tag.
- The most common failure we see is not a missing tag. It is a page where the visible text and the marked-up data disagree, or where the price only exists after JavaScript runs.
Legibility is now a sales channel
For most of the last two decades, a product page had one job: convince a person. Everything else was decoration. Photography carried the feel of the thing, the size chart lived in a tab, the material was in a bullet somewhere, and if the weight was only printed on an image of the spec sheet, nobody minded, because a human reader fills gaps automatically.
Increasingly, the first reader is not a person. It is a system assembling an answer, and it fills gaps by dropping whatever it cannot confirm. The consequence is blunt: a product with three unstated attributes is not a product with three small gaps. It is a product that quietly fails to qualify for comparisons it would otherwise win.
The behavior shift behind this is measurable rather than speculative. Traffic arriving at United States retail sites from AI sources grew 393% year over year in the first quarter of 2026, according to Adobe Analytics, and by May 2026 Adobe found those visitors converting at a rate 54% higher than visitors from other sources. Gartner has projected a 25% decline in traditional search volume as assistants absorb queries. Whatever your view on the pace, the direction settles the argument about whether machine-readable product data is a technical nicety.
Editors’ note: this piece is about documentation, not tactics. Every requirement described below is quoted from public specifications and public developer documentation, linked in place. We have not tested any tool, and nothing here is a promise about where a brand will appear.
What the specification actually asks for
Start with the vocabulary itself. Schema.org’s Product type is a shared, open definition of what a product is — name, brand, identifiers, material, colour, size, dimensions, audience — and it exists so that a description written for humans can also be read as data. Nested inside it, the Offer type handles the commercial facts: what this thing costs, in what currency, from whom, and whether you can buy it right now.
That nesting is the part brands most often get wrong conceptually. Price is not a property of a product. It is a property of an offer to sell that product, which is why two sellers can list the same item honestly at different prices, and why availability belongs in the offer rather than in the product description. A model reading your page is looking for that structure, not for a sentence.
The public introduction to Product structured data then splits the world into two classes, and choosing the wrong one is a common own goal. Product snippets are for pages where the item cannot be bought directly — editorial reviews, for instance, which is why that class carries the pros-and-cons properties. Merchant listings are for pages where a customer can actually purchase from you, which is why that class carries apparel sizing, shipping detail and return policy.
For a merchant listing, the documented requirements are short enough to audit in an afternoon: name, at least one crawlable and indexable image, and a nested offers object. Recommended additions read like a list of the questions a shopper asks before buying: brand.name, category, aggregateRating, audience. The documentation also recommends providing multiple high-resolution images, a minimum of 50,000 pixels when width is multiplied by height, at 16:9, 4:3 and 1:1.
Three failure modes that never show up in a dashboard
The price that only exists after JavaScript. The documentation recommends putting Product markup in the initial HTML, and warns that dynamically generated markup makes shopping crawls less frequent and less reliable — a real problem for fast-changing fields like price and availability. If your only source of truth is a client-side render, you have made your most volatile data your least reliable data.
Markup that disagrees with the page. Among the practices listed in Google’s own guidance on AI features and your website is a deceptively boring one: make sure your structured data matches the visible text on the page. Mismatches are common after a sale ends, after a supplier change, or when a plugin caches a number the template no longer shows. A machine that finds two answers to “what does this cost” has learned that neither can be trusted.
Variants that cannot be addressed. If you sell one product in nine colourways, the variant documentation requires that each variant be preselectable at a distinct URL, typically with query parameters, showing the right image, price and availability, and able to be added to the cart. Where variant state lives only in a dropdown that changes nothing about the URL, eight of your nine colourways are effectively invisible as individually comparable items.
The sentence everyone misreads
The same developer documentation says there are no additional requirements to appear in AI Overviews or AI Mode, no special optimizations necessary, and no need to create new machine-readable files or AI-specific markup. People read that as either “nothing to do” or as proof that structured data is pointless. Both readings are wrong.
What it actually means is that the underlying qualification is ordinary: a page must be indexed and eligible to be shown with a snippet. Crawling must be allowed in robots.txt and by your CDN or hosting layer — a genuinely frequent silent blocker — and important content must be available in textual form rather than trapped in an image. The documentation also describes a “query fan-out” technique, where a single question triggers multiple related searches across subtopics, which is a useful mental model: you are not being matched against one query but against a spray of them, and breadth of stated fact is what catches them.
So the honest summary is that nobody can sell you a private entrance to an AI answer, and the work that remains is the work of being unambiguous. Sites that pass that bar tend to be sites that already invested in helpful, people-first content and then wrote their facts down as data instead of as layout.
A field-level audit you can run this week
Pick your best-selling product and answer these questions from the raw HTML the server sends, not from what your browser paints:
- Is there a
Productobject at all, and is it in the initial HTML? - Does it carry a
namethat matches the visible product title exactly? - Is there at least one
imageat a URL a crawler can fetch and index? - Is there a nested
Offer— not an aggregate range — withprice,priceCurrencyand anavailabilityvalue? - Does the marked-up price equal the price a customer pays today, including after a promotion ended?
- Is
brand.namepresent, spelled one way, everywhere? - Can each variant be reached at its own URL with the correct image, price and availability preselected?
- Are the attributes a shopper compares on — material, dimensions, weight, compatibility, care — present as text rather than only inside a photograph of a spec panel?
- Does robots.txt, plus any protection at the CDN layer, actually allow the crawl you think it allows?
- Have you added an FAQPage built from questions your inbox already receives, rather than invented ones?
Ten questions, one page. If the answer to any of them is “I assume so,” that is the finding. Fixing one page completely is also more useful than fixing every page partially, because a page that passes becomes the template for the catalogue.
Where this fits as a discipline
The practice of making a brand readable, resolvable and recommendable to AI systems has a name — generative engine optimization — and it sits on top of ordinary search fundamentals rather than replacing them. SAMA Labs publishes its approach to AI search optimization in sequence, beginning with a visibility baseline and machine-readability review before any content work, and its GEO for ecommerce brands material is a reasonable place to see what an ordered programme looks like. We link it because the sequencing argument matches what the documentation implies: audit, then fix what cannot be read, then enrich, then keep watching.
The reason that order holds is that every step above removes ambiguity rather than adding persuasion. A synthesized answer is a compression of what a system could verify about a category, and compression discards the unclear. Brands that lose in AI answers are not usually worse products. They are usually less legible ones — and legibility, unlike taste, is a specification you can check line by line.
How we choose what to cover, and what we do when we get something wrong, is set out in how we choose, our editorial standards and our corrections policy. Specifications and documentation cited here were read on 22 September 2026.
