As of November 12, 2026, three catalogue routes had opened within a week: Microsoft began routing Shopify and WooCommerce catalogues to Merchant Center, Amazon opened a workspace that lets outside merchants submit a feed, and Google said a simpler integration arrives next year. Read together, the three moves point at the same surface, which is the product listing itself. What used to be a page built to persuade a shopper is now also a record read by an assistant, and the fields that decide whether a product appears are price, stock, identifiers and shipping terms rather than the hero image. Around29
Three routes, one catalogue
Microsoft tied its route to software many stores already run. Accepting the terms for Clarity Brand Agents sends an eligible Shopify or WooCommerce catalogue to Microsoft Merchant Center, so products can surface in Copilot, Bing and MSN when a shopper asks for a recommendation. No new integration project is required, which is precisely why the data quality problem becomes urgent. Around29
Amazon's Shop Direct Merchant Workspace lets eligible merchants submit a feed, choose which products to include and see which items are active on Amazon and Alexa for Shopping. Amazon states there are no fees to participate and that merchants can opt out, while fulfilment, returns and customer service stay with the merchant. Google, meanwhile, extended checkout inside AI Mode to Canadian shoppers on November 4, 2026 and said a simpler Merchant Center integration will arrive next year. Around29
What the assistant actually reads
Structured catalogues give a platform consistent fields for names, prices, availability and shipping, which is easier to process than information scattered across pages. That is the mechanical reason feeds matter. The commercial reason is that a field which is wrong does not degrade gracefully; it removes the item from the answer. Troiana
| Field | What agents do with it | Typical failure |
|---|---|---|
| Price and promotion | Compare against rival listings | Stale price carried over from a past campaign |
| Stock status | Decide whether to offer the item | Available online, empty on the shelf |
| Product identifier | Match the same item across sites | Missing GTIN, item treated as unknown |
| Shipping and returns | Estimate landed cost and risk | Terms buried in policy pages |
The identifier row is the quiet one
Without a real product identifier such as a GTIN, two listings of the same item cannot be reconciled, and an assistant that cannot reconcile them tends to prefer the catalogue it can read cleanly. This is where price monitoring has to start: not with a rival discount, but with the question of whether the two items being compared are the same item. Troiana
Feed and page have to agree
Structured data on the product page should match the feed, otherwise the assistant receives two versions of the same item. When they diverge, the cheaper or the better-identified version wins the comparison, and the merchant discovers the discrepancy only after a campaign underperforms. Troiana
The identifier row is the quiet one. Without a real product identifier such as a GTIN, two listings of the same item cannot be reconciled, and an assistant that cannot reconcile them tends to prefer the catalogue it can read cleanly. Price competitiveness on paper does not help if the item never enters the comparison. Troiana
The discovery pivot changed the target
OpenAI's agentic commerce work moved from in-chat checkout toward product discovery after Instant Checkout was discontinued in March 2026, with large retailers integrating the discovery-side experience instead. The practical consequence is that the transaction still happens on the merchant's own site, but the shortlist is assembled elsewhere. Universal Commerce Protocol
That inversion rewards a different kind of work. Conversion rate optimisation assumes a shopper already arrived. Discovery optimisation assumes a machine decided whether to send anyone at all, and that decision depends on attributes, availability and the consistency between the feed and the page. Universal Commerce Protocol
Some stores block the agent by accident
One agency sent an automated shopping agent to 200 UK and US retailers twice. Security software refused it at 29 of them, including large chains, while every refusing store whose crawler rules could be read permitted the major AI crawlers. The gap sits between robots.txt, which states which automated visitors are welcome, and bot protection, which decides who actually gets through. Around29
Refusal ran at roughly 20 percent among large retailers and 2.5 percent among smaller ones in that test, which is one agency's sample rather than a market measure. The pattern is still worth checking, because the failure is silent: nobody is told that a legitimate agent was turned away at the edge. Around29
What to fix before connecting anything
The sequence matters more than the tooling. Audit the feed first, checking prices, stock, images and shipping terms, and fix every field before adding a new channel. Then review crawler rules and bot protection together, because a permissive robots.txt behind aggressive bot filtering produces the same outcome as a block. Around29
Only after that does eligibility make sense. Reviewing whether a store qualifies for a programme is cheap; discovering three months later that the catalogue it submitted was wrong is not, because the assistant has already learned to skip the item. Around29
FAQ
Do product feeds replace the store website?
No. The feed is the record an assistant reads; the site still carries fuller detail and closes the sale for most journeys.
Which feed fields cause the most damage when wrong?
Price, stock, product identifier and shipping terms. Each one can remove an item from an answer rather than merely misdescribe it.
Why would a retailer block AI shopping agents without meaning to?
Because robots.txt and bot protection are configured separately. The first can allow a crawler while the second rejects it, and nobody is notified.
Does the discovery pivot reduce the need for clean data?
It increases it. When the shortlist is assembled off-site, attributes and availability decide whether the brand is in the comparison at all.
