What AI assistants actually read on your product page

Ask an AI assistant for “trail running shoes that handle mud, under $150” and it answers with two or three specific products. To be one of them, your product has to be legible: the assistant needs to know what it is, who it fits, and how it compares, without inferring any of it from marketing copy.

Both Google and OpenAI publish exactly what they read. It's worth working from those lists rather than from guesses.

Two ways in

A product reaches an AI answer through one of two doors.

• Your pages, as crawled. Crawlers read the page and its structured data. This is what Google AI Overviews, Perplexity, and ChatGPT's web search work from.
• A product feed. A file you send that lists your catalog: prices, availability, identifiers. ChatGPT's shopping results draw on feeds like this.

The two overlap heavily. A catalog that's complete enough for one is usually close for the other.

What Google requires

For a product page to qualify as a merchant listing, Google's documentation lists five required properties:

• name — the product name
• image — a product photo
• offers — the offer block itself
• offers.price — the active price
• offers.priceCurrency — the currency, in ISO 4217 form

Most Shopify themes emit those five on their own. They get you into the listing; they don't say anything about the product.

The recommended list is where the detail lives, and it maps almost one to one onto the questions shoppers ask:

• material, color, size — what it's made of, what it looks like, whether it fits
• audience — who it's for
• gtin, mpn, sku — identifiers that let an assistant match your product to the same product elsewhere
• availability, itemCondition, shippingDetails, hasMerchantReturnPolicy — whether it can actually be bought, and on what terms
• aggregateRating, review — what other people said
• inProductGroupWithID — which variants belong to the same product

The full list, with the exact formats, is in Google's merchant listing documentation.

What ChatGPT's feed requires

OpenAI's product feed spec asks for nine required fields per item:

• item_id, title (up to 150 characters), description (up to 5,000)
• url, image_url, brand, seller_name
• availability — one of in_stock, out_of_stock, pre_order, backorder, unknown
• price — amount and currency together, like 79.99 USD

Variants are grouped with group_id and described with variant_dict. Identifiers like gtin and mpn are optional but are what let a feed be checked against reality. The spec is published in full by OpenAI.

Required gets you listed. Comparable gets you recommended.

This is the part that catches people out. A listing can satisfy every required field and still never be recommended, because nothing in it answers the question that was asked.

“Handles mud” is not a field. An assistant arrives at it from an outsole description, a material, a use case in the product description, maybe a review that mentions wet ground. “True to size” comes from sizing notes or review text. “Under $150” comes from the price, which is the one part most stores already get right.

So the useful question isn't “does my markup validate?” It's “if someone described this product back to me from my own data alone, would they describe it accurately, and would they be able to compare it to the next one?”

Checking your own store in ten minutes

1. Open a product page, view source, and search for ld+json. Confirm you have a Product block with name, image, price, and currency.
2. Paste the URL into Google's Rich Results Test and read the warnings, not just the errors. Warnings are the recommended fields you're missing.
3. Read the product description as if you'd never seen the product. Does it say what it's for, who it fits, and what it's made of? Or does it say “engineered for performance”?
4. Check two variants of the same product. Do they share a group identifier, or do they look like two unrelated products?
5. Pick your three best sellers and ask ChatGPT and Perplexity a question those products should win. Note what gets recommended instead, and why that listing is easier to read than yours.

The last step is the one that tends to change minds. It's also the first thing we do on a call.

What this doesn't do

Complete data makes a product eligible and legible. It doesn't guarantee a recommendation: assistants weigh reviews, brand familiarity, price, and their own retrieval quirks, and none of that is under your control. Google AI Overviews, in particular, follow normal indexing, so changes there move at crawl speed.

What you can control is whether an assistant has to guess. Most catalogs make it guess.