Ask ChatGPT to recommend a pair of running shoes, a standing desk, or a carry-on suitcase, and it does not crawl your storefront the way Google used to. It reads a handful of product signals, compares them against competitors, and writes a recommendation in its own words. Your product page either supplies the material for that answer or it does not.
This is the practical difference between ranking and being cited. Ranking gets you a blue link. Citation gets your product named, described, and recommended inside the answer itself, often with no click required at all. For e-commerce, that shift changes what a product page is actually for.
This guide covers how AI shopping assistants decide what to surface, what schema they lean on to describe your products accurately, and why the product pages that get cited look different from the ones that get ignored.
ChatGPT, Gemini, and Perplexity are not indexing your site the way a search crawler does. They are pulling from a mix of structured product feeds, on-page markup, and general web content, then synthesizing an answer that names specific products with specific reasons attached.
OpenAI has been explicit about this mechanism for ChatGPT's shopping experience. Its Agentic Commerce documentation describes merchants submitting structured product feeds, in CSV or JSON, containing identifiers, pricing, inventory, media, and fulfillment details, refreshed on a regular cadence so the assistant is working from current stock and pricing rather than stale listings. The feed is the primary channel, but it is not the only one: on-page structured data acts as a secondary, always-on signal that confirms and enriches what the feed says.
Gemini and Perplexity work from a different pipeline (retrieval over the open web rather than a dedicated merchant feed) but they lean on the same underlying signal: structured, unambiguous product data that can be lifted into an answer without the model having to guess. A paragraph of marketing copy is hard to extract facts from. A clearly marked price, availability status, and rating are not.
The practical implication is that your product page now has two audiences: the shopper reading it, and the model deciding whether to mention it at all. Writing for only the first one is how pages become invisible to the second.
Three things tend to separate a cited product from an ignored one: verifiable specifics, comparability, and evidence. None of these are exotic. They are the opposite of what most product pages optimize for, which is persuasive copy with soft claims.
Verifiable specifics means numbers, dimensions, materials, and compatibility details stated plainly rather than implied. "Lightweight and durable" gives a model nothing to cite. "1.2kg, aircraft-grade aluminum shell, IP54 rated" gives it three separate facts it can attribute to your product by name.
Comparability means the page answers the question a shopper is actually asking, which is almost always relative: is this better than the alternative, and for whom. A page that only describes itself in isolation forces the model to go find comparison data elsewhere, often from a competitor's page or a third-party review site, and cite that instead.
Evidence means reviews, ratings, and real usage detail that back up the claims on the page. A specification is a claim. A specification plus an aggregate rating of 4.6 from 800 reviews is a claim a model can repeat with confidence, because it has a second source agreeing with the first.
Put together, these three things describe a page that reads less like a sales pitch and more like a spec sheet with a point of view. That combination is what is easy to quote directly.
Structured data is how you make the above machine-readable instead of just human-readable. Schema.org's Product type defines the vocabulary that AI systems and search engines use to parse a product page: name, description, image, brand, and identifiers like sku, gtin, and mpn.
Two nested types do most of the heavy lifting for AI citation specifically. Offer carries price, currency, and availability, which is exactly what a shopping assistant needs to answer "is this in stock and what does it cost" without having to parse prose. AggregateRating and Review carry the social proof: a rating value, a review count, and optionally the review text itself.
Here is a minimal but complete example of what that looks like combined:
{ "@type": "Product", "name": "...", "sku": "...", "brand": {"@type":"Brand","name":"..."}, "offers": {"@type":"Offer","price":"129.00","priceCurrency":"USD","availability":"https://schema.org/InStock"}, "aggregateRating": {"@type":"AggregateRating","ratingValue":"4.6","reviewCount":"812"} }
Consistency matters as much as presence. If the price in your schema does not match the price on the page or in your merchant feed, assistants and search engines alike tend to deprioritize the listing rather than guess which number is correct. One source of truth, reflected everywhere, beats three sources of truth that occasionally disagree.
If you have not yet audited how your broader site exposes structured data to language models, our piece on structured data for LLMs covers the underlying mechanics in more depth, including how schema differs from what a crawler versus a model actually consumes.
Most product description copy is written to sound appealing, not to be extracted. "Our premium blend delivers an exceptional experience" is unusable by a model because it contains no facts. It cannot be compared, verified, or repeated as a reason to choose your product over another.
This is not a call to abandon persuasive writing. It is a call to put the extractable facts somewhere on the page too, ideally in a scannable format: a spec table, a bulleted list of materials and dimensions, a clearly labeled compatibility section. The narrative copy can stay for the human reader. The facts need their own home for the model reading alongside them.
A useful test: read your product page and ask whether a stranger could write an accurate one-sentence comparison against a named competitor using only what is on the page. If the answer is no, an AI assistant summarizing your category is going to reach for a page where the answer is yes, and that page will not be yours.
The product pages that consistently get cited share a few structural habits worth copying directly.
FAQPage schema where it applies.None of this requires rewriting your entire catalog at once. Start with your highest-traffic or highest-margin SKUs, add the spec table and comparison section, and confirm the schema on those pages is complete and matches your live pricing before expanding further.
Review content does double duty in an AI-driven shopping flow. It gives your product page the social proof a model looks for before recommending something, and it often supplies the specific, concrete language that a generic product description lacks: "runs a half size small," "battery lasts about two days of moderate use," "assembly took 20 minutes with two people."
That specificity is exactly the kind of detail a model will quote or paraphrase, because it answers a real question a shopper would otherwise have to ask a human. Surfacing a handful of the most detailed, specific reviews near the top of the page, rather than only showing a star average, gives an assistant more usable material.
Marking that content up with Review and aggregateRating properties, and keeping the review count current, is what turns "customers seem to like it" into a citable fact: a rating, a count, and a source.
Schema and content changes are only useful if you can tell whether they are working. That means checking, on a recurring basis, whether ChatGPT, Perplexity, Gemini, Claude, and Copilot are actually naming your products when someone asks a shopping question in your category, and how accurately they describe them when they do.
That is a different exercise than traditional rank tracking, because there is no single results page to screenshot. It means running the actual comparison and recommendation prompts a shopper would type, across each assistant, and recording whether your brand shows up, how it is described, and which competitors show up instead. Our guide on how to check your AI brand visibility walks through how to do this manually before automating it.
At Bold GEO, this is the daily-refresh version of that same check: we run your brand and product terms against ChatGPT, Perplexity, Gemini, Claude, and Copilot on a schedule, so you can see when a schema fix or a rewritten spec table actually moved the needle on whether your products get mentioned, rather than guessing.
A few recurring issues show up across the product pages that never get cited, and most of them are fixable in an afternoon rather than a redesign.
Schema that describes a different product than the page. Templates get copied, and a price or SKU from a previous product lingers in the JSON-LD after the visible content changes. Assistants that cross-check schema against visible page text will quietly drop a listing that contradicts itself rather than guess which version is correct.
Missing or stale availability. An Offer block that still says InStock for a product that sold out weeks ago is worse than having no schema at all, because it gets surfaced and then fails the shopper at checkout. Availability should update on the same cadence as your actual inventory, not on a quarterly content refresh schedule.
Reviews locked behind JavaScript rendering. If your review widget loads client-side and never appears in the initial HTML response, a lot of retrieval pipelines never see it, schema or no schema. Server-rendering at least the review count and rating value, even if the interactive widget loads separately, keeps that signal visible.
Category pages standing in for product pages. A comparison chart buried on a category or collection page, rather than repeated in some form on the individual product page, means the model has to already be on the right page to find it. Bring the comparison down to the product level, even briefly, rather than assuming a shopper (or an assistant) will hop back up a level to find context.
One-size-fits-all descriptions across variants. A product with five colorways or three sizes that all share identical description text gives an assistant nothing to distinguish them by. Where a variant has a genuine difference in weight, fit, or material, that difference belongs in the text and the schema for that specific variant, not just in a dropdown label.
None of these require a platform migration. They require treating your product data pipeline, the thing that feeds your Merchant Center feed, your on-page schema, and your review widget, as one system that needs to agree with itself, rather than three separate integrations that were each set up correctly once and never reconciled since.
AI shopping assistants reward the same qualities a careful human shopper would: clear facts, honest comparisons, and evidence that other people have used and liked the product. The mechanism for surfacing those qualities to a model is structured data. The substance underneath it is a product page that actually tells you something.
Start with your feed and schema accuracy, because an assistant working from wrong prices or missing availability will not recommend you regardless of how good your copy is. Then invest in the spec density, comparisons, and review evidence that give the model something worth quoting. The pages that do both are the ones that show up when someone asks an AI assistant what to buy.
Bold GEO monitors how your brand is cited across ChatGPT, Perplexity, Gemini, Claude, and Copilot on a daily refresh. 7-day free trial, no credit card.