Product data that AI shopping features and agents can use

When a shopper asks an assistant for a waterproof jacket under £150 that arrives by Friday, the answer is assembled from product data. The shop rarely knows which copy of its data was read. Google's AI Mode works from its Shopping Graph, ChatGPT from the feeds merchants send it, and a browsing agent from the page itself. Each copy can say something different about price, stock, delivery or returns, and an agent that finds a contradiction has every reason to pick another shop. Most of the work is making the copies agree, and making each one complete enough to answer a specific question.

This goes further than the product data section of the ecommerce SEO guide. Variants have their own post, on URLs, canonicals, ProductGroup markup and Merchant Center, so they get only a short summary here.

Google says shopping in AI Mode is powered by its Shopping Graph, which it describes as more than 50 billion product listings, 2 billion of them updated every hour.¹ There are two ways in: structured data on product pages and a feed uploaded to Merchant Center. Merchant Center is required for the Shopping tab, and Google can update its copy of your product data from the website when it notices a mismatch.² For AI features, two housekeeping points follow: keep Merchant Center information up to date, and make sure structured data matches the visible text on the page.³

ChatGPT has its own route. In March 2026 OpenAI said it was extending the Agentic Commerce Protocol (ACP) to product discovery, with merchants sharing product feeds and promotions so their catalogues are represented in ChatGPT, through several delivery paths including third-party providers such as Salesforce and Stripe. It said merchants on Shopify are already integrated through Shopify Catalog with no extra work.⁴ Shopify's help centre adds that agentic storefronts are active by default for eligible stores, that products reach AI channels through Shopify Catalog, and that ChatGPT users complete their purchase on the store's own checkout.⁵

Agents use a feed in a specific way. The ACP proposal for product feeds says an agent uses one to build a searchable index of products and variants, match a buyer's request to a purchasable variant, show choices with price, stock, seller and policy links, and pass the right item identifier into checkout. It also says feed prices are discovery-time signals that may be stale, and that the checkout response is authoritative for price, tax and fulfilment.⁶ That split matters. A wrong feed price does not cost you money at checkout. It costs you the recommendation, or the sale when checkout disagrees with what the buyer was shown.

The Merchant Center product data specification is still the best checklist for what a product record needs, because the other systems ask for much the same. Four requirements do most of the work:

• GTIN when the manufacturer assigned one, and only if you are sure it is correct. MPN is required when there is no GTIN.⁷
• Brand for all new products, except movies, books and music recordings. Set identifier_exists to "no" only when a product lacks the identifiers its category expects.⁷
• Price and currency that match the landing page, the structured data and the checkout.⁷
• Availability that matches the landing page, the checkout pages and the structured data.⁷

A correct GTIN is what lets any system recognise that your jacket is the same product another retailer sells, which is the comparison an assistant is often asked to make. Own-brand products have no manufacturer GTIN, and the rule there is to leave it blank. Never invent one. OpenAI's feed specification takes the same line: a GTIN must be 8, 12, 13 or 14 digits with a valid check digit, and should be omitted if unassigned.⁸

Conversational attributes

Since May 2026, retailers globally can use Merchant Center's conversational attributes.⁹ Merchant Center Help lists six of them: question and answer, document link, related product, item group title, variant option and popularity rank. They are optional, they complement the main specification, and you do not need to repeat details you already send in the description, product highlights or product details.¹⁰

Google's merchant listing markup applies only to pages where a shopper can buy the product. The required properties are the product's name, image and offers, and the offer needs a price above zero with its currency. Include every applicable global identifier, such as a GTIN or MPN.¹¹ These are standard schema.org properties of Product, along with sku, brand and offers,¹² so the same markup is not written for Google alone.

Shipping and returns are the fields most shops leave out, and shoppers weigh them heavily. In Baymard Institute's 2026 data on why US online shoppers abandon a cart or checkout, setting aside those who were just browsing, 40% gave extra costs such as shipping, tax and fees as a reason and 13% an unsatisfactory returns policy.¹³ Baymard does not state the sample size on that page. In markup, schema.org allows a MerchantReturnPolicy on an Organization, a Product or an Offer.¹⁴ Google recommends a single shipping policy and a single return policy for the business in Organization markup, and keeps product-level markup for products that differ from the standard policy.¹¹ When several sources disagree, Google applies a fixed order: settings in Merchant Center or Search Console outrank product-level markup, which outranks Organization markup.¹⁵¹⁶ So if your returns window lives in Merchant Center settings, that is the version Google uses, whatever the page says. Keep the three in step, or pick one place and remove the others.

For variants, Google's markup uses a ProductGroup with a name, a productGroupID (the parent SKU), variesBy for the dimensions that change and hasVariant for the variants, and each variant Product needs its own sku or gtin and its own offer.¹⁷ The same group and variant shape appears in the ACP feed proposal, where a product groups one or more purchasable variants with stable IDs.⁶ OpenAI's feed specification asks for a separate row for each variant, with its own item_id and a shared group_id, and says to keep IDs stable when price, stock, title or images change.⁸ If the parent SKU in your markup, the item_group_id in Merchant Center and the product ID in your ChatGPT feed are three different numbers, fix that first.

Check that the markup is in the HTML your server sends, before any JavaScript runs. Many AI crawlers do not render JavaScript, as the AI search guide explains, so markup injected by a tag manager may be invisible to them. In bash, curl -s https://www.example.com/products/storm-jacket | grep -o '"@type": *"[A-Za-z]*"' | sort | uniq -c counts the schema types in the raw HTML. In PowerShell, (Invoke-WebRequest 'https://www.example.com/products/storm-jacket' -UseBasicParsing).Content | Select-String -Pattern '"gtin[0-9]*"' -AllMatches confirms a GTIN is there. The Understand skill in my open-source pack runs this kind of check and validates the markup against the visible page.

OpenAI's approach has changed once already, so dates matter here. These are the published milestones up to September 2026.

OpenAI has not published how ChatGPT ranks products, and the ACP feed proposal says plainly that search and ranking are outside its scope.⁶ Treat anyone offering a ranking formula for ChatGPT shopping with suspicion. What is documented is the data a merchant can supply, and that is where the effort belongs.

Data sent in a feed only helps if the page it links to loads for the systems that check it. A product page that your CDN serves to shoppers but blocks for assistant fetchers undoes the feed, which is covered in when your CDN blocks AI crawlers.

Fellside Supply is an invented outdoor retailer selling an own-brand waterproof jacket in three colours and five sizes. It runs a weekend sale. The page price changes on Friday evening, but the Merchant Center feed is generated nightly and the ChatGPT feed weekly. Its returns policy was extended from 28 to 60 days in the spring, and only the website was updated.

Three problems show up. Price drifts for up to a week, so any surface reading a feed can show £149 while the page says £119. The returns window is 28 days in the place Google ranks highest, so Google will use 28 days whatever the markup says.¹⁵ And the ChatGPT feed uses internal IDs that match nothing else, so nobody can trace a variant across systems. The fixes are ordinary: update both feeds when prices change, change the returns setting in Merchant Center, and use the SKU as the variant ID everywhere. The SKU without a GTIN is correct for an own-brand product.

None of this is new work invented for AI. It is the product data discipline good shops already had, applied to more destinations and checked more often. I would fix drift between the feed and the page before writing a single conversational attribute, because an agent that finds the wrong price will not stay to read the answers.