What does a verified product feed inside ChatGPT actually certify? Not the price a shopper will pay, at least not yet (Relevant Audience). OpenAI opened Businesses in ChatGPT in beta on November 10, letting merchants claim a profile, attach a self-service feed carrying pricing and availability, and watch aggregated counts of mentions and website clicks. The programme is free as of November 12, 2026 and deliberately narrow, with Business Agents limited to eligible e-commerce and online-service businesses in the United States, profile claims open wherever ChatGPT operates, and Shopify catalogues already flowing into answers, which makes the feed behind the profile the part that changes merchant behaviour.
A product feed is now a ranking input
A product feed used to be plumbing between a merchant and a marketplace. In an answer engine it becomes the source a model reads before it says anything about price or stock. That changes the failure mode: a stale feed does not produce a broken page, it produces a confident sentence with the wrong number. Trade coverage of the launch made the point directly, noting that invented pricing has been one of the fastest ways for AI shopping experiences to lose consumer trust. Seen from the category side, the useful market size is not the number of listings a catalogue holds but the number of buying questions it can answer without inventing a number, and that figure is a consumer demand signal merchants have never had to measure before.
What the three numbers measure
The dashboard reports mentions, website clicks and discovery topics. Each answers a narrower question than merchants assume. Mentions count appearances in answers, clicks count recorded link clicks, and discovery topics summarise recurring themes in conversations where the site was cited. None of them measures sales, and none of them is a share of voice against competitors.
| Metric | What it counts | What it can answer | What it cannot answer |
|---|---|---|---|
| Mentions | Answers that reference the business | Absolute visibility in answers | Position relative to competitors |
| Website clicks | Recorded clicks on the business link | Referral volume from answers | Whether the visit converted |
| Discovery topics | Themes in conversations citing the site | Which questions the brand appears in | Which questions it is missing from |
Mentions are narrower than they look
For an online business, a mention counts only when the answer references the website; naming the brand alone does not register. For a local business, a mention counts when the result is displayed, even if nobody clicks. The asymmetry is easy to miss and it makes cross-category comparisons unreliable. A local retailer and an online catalogue comparing mention counts are comparing two different definitions.
Feed hygiene becomes a merchandising job
If the feed decides what the model says about price, then maintaining it belongs to merchandising rather than to the web team. In practice that means a named owner for price and availability accuracy, a refresh cadence matched to how often promotions change, and a rule for what happens when an item goes out of stock mid-promotion. None of that is exotic, but all of it has to be assigned.
The reconciliation problem
Most merchants now maintain product data in at least three places: the storefront, the marketplace listings and the ad platform's catalogue. Adding a fourth destination creates a reconciliation problem that only shows up as a customer complaint. The practical fix is to treat one system as the source of record and push the others from it, rather than editing each destination by hand during a promotion.
What is still missing
Three things are absent from this beta and each limits what can be concluded. There is no share-of-voice metric, so competitive position has to be estimated by sampling the same questions across platforms. Insights appear only above an interaction threshold, which disadvantages small catalogues. And the agent cannot complete a purchase, so the click still has to convert on the merchant's own site. Until those gaps close, the dashboard is a diagnostic tool, not a performance report.
FAQ
Does claiming a profile improve ranking?
OpenAI says claiming does not change how often a business appears. It changes the accuracy of what the model can say.
Who can use Business Agents now?
Eligible e-commerce and online-service businesses in the United States, at launch.
Why do some dashboards show nothing?
Insights are only displayed above an interaction threshold, and a blank panel can mean low volume rather than no activity.
What is the first thing to fix?
Price and stock accuracy. A wrong number in an answer costs more trust than a missing profile.
