When a shopper asks Google's AI Mode or ChatGPT to compare products, the answer is not assembled by reading your product pages the way a person would. It is assembled largely from structured product data: your Merchant Center feed, the feeds you share with AI platforms, and the schema markup on your pages. Your product page still matters, but increasingly as a source of verification rather than the primary input.
That shifts where enterprise ecommerce teams win or lose. The product data layer used to be an operational concern owned by whoever ran Shopping ads. In 2026 it is the shelf your products sit on in AI answers. Here are the tactical issues we see most often, and how to fix them.
What changed this year
Three developments reshaped the landscape in 2026.
Google formalized agentic shopping. In January, Google introduced the Universal Commerce Protocol, an open standard for how AI agents and merchant systems communicate across discovery, checkout, and post-purchase. It powers checkout inside AI Mode and Gemini for eligible merchants, and Google added dozens of new Merchant Center attributes aimed at conversational discovery. Crucially, AI Mode draws on the same Merchant Center feed as Shopping, so there is no separate AI feed to build. There is just a feed that is either rich enough or is not.
OpenAI moved from checkout to discovery. OpenAI retired its in-chat Instant Checkout in March and extended its Agentic Commerce Protocol to product discovery, with merchants sharing product feeds and promotions so their catalogs are represented in ChatGPT shopping results. Large retailers integrated quickly, and Shopify merchants are connected through Shopify Catalog. Most enterprise brands on other platforms have to build and maintain the feed themselves.
Agents now compare on attributes. Shoppers describe requirements in full sentences, and AI systems match those requirements against structured attributes. A title, description, and price were enough for keyword search. They are not enough to answer "which of these is rated for outdoor use below freezing."
We covered what agents can and cannot do in Agentic Commerce in B2B. This piece is about the data plumbing underneath.
Issue 1: Three versions of the truth
On most enterprise sites, product facts are published three separate ways: the visible product page, the JSON-LD schema in the page source, and one or more channel feeds. Each is generated by a different system, on a different schedule, often owned by a different team.
They drift. The page shows a sale price the feed has not picked up. The schema says in stock while the feed says backorder. The feed title was rewritten for ads and no longer matches the page. Google treats price and availability mismatches as grounds for disapproving items, and every AI system that cross-checks sources has reason to trust you less.
The fix: generate every output from one source. Page content, schema, and feeds should all render from the same product record, whether that lives in a PIM or the commerce platform, rather than being assembled independently. Then add an automated parity check that samples SKUs daily and flags any difference in price, availability, GTIN, or title across the three. If you are unsure where that product record should live, our PIM decision guide walks through it.
Issue 2: Schema that validates but says almost nothing
Passing a validator is a low bar. A great deal of enterprise Product markup contains a name, an image, a price, and little else. It is technically correct and nearly useless to a system trying to compare products.
The fix: mark up what a buyer would compare. At minimum, for each product:
- Identifiers: GTIN where one exists, plus MPN and brand. Identifiers are how systems match your product to the same item elsewhere.
- Offer detail: price, currency, availability, and condition that match the page exactly.
- Shipping and returns: OfferShippingDetails and MerchantReturnPolicy, which answer two of the most common pre-purchase questions.
- Variants: a ProductGroup with hasVariant and variesBy for size, color, or configuration, instead of a single Product that hides the options.
- Specifications: additionalProperty entries for the technical attributes your buyers filter on, in consistent units.
- Reviews, where you have genuine ones: AggregateRating tied to real review data.
Then confirm it is present in the server-rendered HTML. Structured data injected by client-side JavaScript is invisible to most AI crawlers, which do not execute scripts. We cover that rendering gap in detail in Faceted Navigation, Crawl Waste, and AI Crawlers.
Issue 3: B2B pricing that disappears behind the login
B2B sites have a structural problem consumer sites do not. The real price is contract pricing, shown only after login. The public page shows nothing, or "request a quote." AI systems then describe the product as price on request, and buyers comparing options gravitate to a supplier who shows a number.
The fix: make a deliberate public pricing decision. Options include a published list price with a note that account pricing may be lower, a price range, or quantity-break pricing expressed with UnitPriceSpecification. None of these require exposing contract terms. What matters is that the decision is made on purpose rather than defaulting to silence. Products that genuinely require configuration can say so clearly, with a starting configuration priced. The quoting side of this is covered in The B2B Quote-to-Order Problem.
Issue 4: Attributes built for filters, not questions
Product attributes on most catalogs were designed to power faceted navigation: a manageable list of values that fit in a sidebar. Conversational shopping asks different questions. Is it compatible with my existing system? What certifications does it carry? Will it work in this environment? What is the replacement part?
The fix: run an attribute gap analysis against real questions. Collect the questions buyers actually ask from sales calls, support tickets, on-site search logs, and AI prompt audits. Map each one to an attribute. Where no attribute exists, add it to the product model and enrich top sellers first. Prioritize compatibility, use case, certifications, materials, dimensions, and operating ranges, and keep units and value formats consistent across the catalog. Then map the enriched attributes into Merchant Center's conversational fields and your other feeds.
Issue 5: Nobody owns product syndication
Ask an enterprise team who owns product data for AI channels and you often get three answers. Marketing owns Merchant Center because it started as an ads tool. IT owns the catalog and the PIM. Ecommerce owns the website. The ChatGPT feed, if it exists, belongs to whoever set it up.
The fix: name a product syndication owner and build a channel matrix. One role should be accountable for every outbound product data channel, with a simple matrix covering each channel, the fields it requires, its refresh frequency, its error reporting, and who fixes what. Feed errors should route to the team that owns the source data, not to whoever happens to see the dashboard. This is integration work as much as marketing work, and the patterns in our ERP to eCommerce integration playbook apply directly.
Issue 6: No measurement of AI shopping visibility
Most analytics setups lump AI referrals into general referral or direct traffic, and feed health gets checked only when something breaks.
The fix: measure three layers. First, feed health: disapproval rates, missing attributes, and mismatch warnings, tracked weekly. Second, traffic: a dedicated channel grouping for AI referral sources such as ChatGPT, Perplexity, Gemini, and Copilot, so you can see landing pages and conversion. Third, presence: a recurring prompt audit of how AI tools describe and recommend your products, using the method in What Does AI Say About Your Brand?
A 90-day plan
- Weeks 1 to 2: Run a parity check across page, schema, and feed for your top 500 SKUs. Fix price and availability drift first.
- Weeks 3 to 4: Audit schema depth and server-side rendering on product and category templates.
- Weeks 5 to 8: Run the attribute gap analysis and enrich top sellers. Map new attributes to Merchant Center's conversational fields.
- Weeks 9 to 10: Assign syndication ownership and build the channel matrix, including an OpenAI product feed if you are not on Shopify.
- Weeks 11 to 12: Stand up AI referral reporting and run your baseline prompt audit.
FAQ
Q: Do we need a separate feed for Google AI Mode? No. AI Mode uses the same Merchant Center data as Google Shopping. The difference is that richer attributes now matter more, because AI Mode matches products to detailed, open-ended questions rather than short keyword queries. Checkout inside AI Mode through the Universal Commerce Protocol requires additional eligibility and integration.
Q: How do products get into ChatGPT shopping results? Through product data OpenAI can access: merchant product feeds shared under its Agentic Commerce Protocol, platform integrations such as Shopify Catalog, and crawled web content. In 2026 OpenAI shifted its focus from in-chat checkout to discovery, so for most brands the priority is a complete, accurate feed rather than a checkout integration.
Q: Is schema markup still worth the effort if feeds matter more? Yes. Schema on your pages is how AI systems and search engines verify what your feeds claim, and it is what crawlers read directly when no feed exists. Feeds and schema reinforce each other when they agree and undermine each other when they do not.
The bottom line
In AI shopping, your product data is the shelf. Brands that publish one consistent version of the truth, describe products in the attributes buyers actually ask about, and treat syndication as an owned discipline get recommended. Brands with drifting feeds and thin schema get described inaccurately or skipped.
Our eCommerce Audit includes a product data and syndication review covering parity, schema depth, attribute gaps, and feed ownership, so you know exactly which fixes will move AI visibility first.