ChatGPT Shops From Product Feeds Now. Your Website Is the Fallback.
Ask ChatGPT for running leggings and it shows you a short list: a photo, a price, a one-line verdict like "Best for workouts" and a link to buy. It looks like a search result. It is closer to a shelf, and the products on it increasingly come from retailers who handed OpenAI their catalog directly rather than from whatever ChatGPT could find on the open web.
That shift is now measurable, and it is the clearest sign yet that being easy for an agent to read is no longer a nice-to-have. It decides whether you are on the shelf at all.
Feeds went from a tenth of the shelf to most of it overnight
ChatGPT Shopping results lean hard on product feeds
Across 1,757,723 tracked ChatGPT Shopping prompt runs in July 2026, the share of product picks from feed-integrated merchants rose from 8.26% to 61.54% on 10 July. By 3 September, feeds supplied about 65% of recommendations. The ten most-referenced stores went from 22.5% of references to 41.8%, and the number of distinct merchants referenced fell by more than a fifth, from 13,524 to 10,607.
The switch did not happen gradually. It happened in a day, and it moved real businesses. Of 687 retailers Profound tracks, 450 lost at least a third of their ChatGPT shopping visibility between 7–9 July and 10–12 July. Sixty-seven gained by the same margin. Profound's model attributes most of that movement to whether a store's products were being retrieved from a feed or from web search.
Earlier Profound research, reported by WARC, found 87.3% of product cards coming from web crawl. That fits the picture before July rather than contradicting it: until this summer, ChatGPT mostly shopped the open web.
Why an agent prefers a feed
None of this requires believing OpenAI is playing favourites. Its help centre says product results "are not ads, nor influenced by any OpenAI partnerships", and nothing in Profound's data shows otherwise. A feed wins for a duller reason: it is easier to trust.
A product page is written for a person. The price might be rendered by JavaScript after the page loads. The stock status might sit behind a size selector. The sale price and the original price might both appear with nothing in the markup to say which is current. An agent reading that page has to guess, and every guess is a chance to show a shopper a price that is wrong.
A feed answers the same questions without guessing. OpenAI's product feed specification requires nine fields for every item: an ID, title, description, URL, brand, seller name, image, availability and price. It asks merchants to update the feed whenever a price or stock level changes. An agent comparing two leggings, one described by a feed refreshed this morning and one by a page it scraped last week, will reasonably pick the one it can stand behind.
That is the whole lesson, and it reaches well beyond ChatGPT shopping. Agents favour whatever they can read reliably. When a structured, first-party source exists, it beats a page the agent has to interpret. When it does not, the agent falls back on the page, and the quality of that page decides whether you appear at all.
This is the pattern, not the exception
Shopping is simply where the stakes are easiest to count. The same thing is happening wherever agents act for people.
- Answers. Assistants cite pages they can fetch and parse. A page that arrives as an empty shell until JavaScript runs is invisible to most of them. In our State of Agent Readiness scan of 100 well-known home pages, 27 could not be read by an automated agent at all.
- Applications. An agent booking, filing or buying inside a product needs an API or an MCP server it can call. One that has to click through your interface like a person will be slower, break more often, and get replaced by a competitor that offers the direct route.
- Commerce. Shopify stores reach ChatGPT through Shopify Catalog. Other retailers request direct feed access and, for now, join a waitlist. The platform that made structured data the default handed its merchants a head start.
The direction is the same each time. Agents are moving from reading your interface to reading your data, and the businesses that publish clean data first get chosen first.
What to do about it
If you sell products, apply for a direct feed with OpenAI, or make sure your platform's integration is switched on. Then treat the feed as a live system, not an export you set up once. A feed that says "in stock at $98" and a page that says "sold out at $128" is worse than no feed, because the shopper blames you for the mismatch.
Keep the website agent-readable anyway. Feeds supply most of the shelf
in Profound's data, not all of it, and the July switch shows how quickly
the balance can move again. Every product click still lands on your page,
and an agent that checks the price before handing over a shopper reads it
there. Server-render the price and availability, mark them up with
Product and Offer structured data,
and make sure bot protection is not turning away the agents you want. In
our State of Agent Readiness report, bot
challenges were the most common reason a page could not be read, at 18 of
the 27 failures. The Agent Analyzer checks a page for
all of this, and our
ecommerce guide covers the details.
If you run an application, give agents a direct route to the tasks people ask them to do: a documented API, an MCP server, or at minimum pages and forms an agent can operate without guessing. The Agent Visualizer shows the walk an agent takes through one of your pages: where it arrives, what it reads, where it stops, and what it never reaches.
Watch the concentration. The top ten stores nearly doubled their share of ChatGPT's shopping references in a day. When an agent narrows the shelf, the merchants left on it are the ones that made themselves easy to read. That is a window, and it will not stay open as more retailers catch up.
What this does not tell you
Profound's figures come from prompts its own customers track, not from real shoppers' sessions or purchases. They show where ChatGPT retrieved products from, not how it ranked them. OpenAI has not confirmed the 10 July change or published how often it uses feeds rather than the web, and a newer model began rolling out in September, so the split may already have moved. Nor does it show whether a feed outweighs price: whether ChatGPT would pick a feed-synced retailer over a cheaper one it found on the web is untested.
The direction does not depend on any of those numbers holding exactly. An agent shopping for someone will recommend what it can confirm, and the fastest way to be confirmed is to tell it directly.
See the journey on your own site
Being on the shelf is the first step. Watch the walk an agent takes through one of your product pages, and see where it stops before it reaches the price, the stock or the buy button.