ChatGPT Shopping now pulls most product picks from merchant feeds

Search Engine Journal reported on 9 September 2026 that research from Profound found ChatGPT's shopping recommendations now come overwhelmingly from merchant product feeds rather than from pages crawled on the open web. The sharpest moment in the data is 10 July 2026, when the share of product picks sourced from an integrated feed moved from 8.26% to 61.54% inside Profound's tracked prompts.

What Profound actually measured

Profound based the analysis on 1,757,723 ChatGPT Shopping prompt runs during July 2026, alongside a separate visibility sample of 687 customers used to see how the change landed on individual merchants. The later figure, roughly 65% of tracked recommendations coming from feed retrieval as of 3 September 2026, was drawn from 97,725 prompt runs sampled at 10%.

The unit of measurement matters before any of the percentages do. These are prompt runs issued and tracked by one vendor, not sessions from real shoppers, so the numbers describe what ChatGPT returned to a monitoring system rather than what buyers saw or bought. Read that way the finding still holds value, because the retrieval mechanism it exposes is the same one a shopper meets.

Profound also ran a statistical model across 517 customers whose visibility moved, and reported that changes in web search versus feed retrieval explained 83% of the variation. That is the part of the study carrying the causal weight. The visibility swings were not scattered noise across the sample. They tracked the switch in where ChatGPT was sourcing products from.

The 10 July 2026 shift, in numbers

The figures below are the ones Profound published and Search Engine Journal reported on 9 September 2026, with the before and after states falling either side of 10 July 2026.

The visibility row is worth reading slowly. Profound compared 7 to 9 July 2026 against 10 to 12 July 2026 and found 450 of the 687 tracked customers lost at least a third of their Shopping visibility, while 67 gained by the same margin. That is close to a seven-to-one ratio of losers to winners inside a three-day window, which is not how gradual algorithm drift behaves.

Fewer stores are getting named at all

The concentration figures deserve a second read. References to the top 10 stores rose from 22.5% to 41.8% of tracked mentions, and the count of unique merchants referenced fell from 13,524 to 10,607. Roughly one store in five that used to appear somewhere in the tracked results stopped appearing.

That is a different failure from ranking lower. A page that slips in classic search still exists on the second results page, and a determined shopper can scroll to it. An assistant that names five products and stops has no second page. A merchant is either inside the retrieved set or absent from the answer, and absence looks identical to not existing.

Concentration also compounds. When retrieval draws from a structured catalogue rather than from whatever the model finds while reading the web, the merchants already inside that catalogue get repeated, and repetition is what a top-10 share of 41.8% describes. Nothing in the research says this state is permanent, but it does describe where the tracked results sat during the sampled window.

Roughly a third of feed retrieval traced back to Shopify

Profound estimated that around 35% of July 2026 feed retrieval was Shopify related. That single number explains a lot of the winners in the sample, because Shopify stores are connected to ChatGPT Shopping through the Shopify catalogue without the merchant doing anything, while an independent retailer on a custom platform has to apply.

For a store already on Shopify, the practical consequence is that catalogue quality became the lever rather than page-level content. Product titles, descriptions and structured attributes are what a feed carries, and they are what gets retrieved. That work overlaps heavily with ordinary Shopify SEO, which is a small mercy: the same catalogue discipline that helps a product page rank is the discipline that fills a feed properly.

How a product gets into ChatGPT Shopping today

Shopify and Etsy catalogues are connected already

Shopify stores are integrated automatically through the Shopify catalogue, with no extra steps required from the merchant, and Etsy catalogues are connected as well. A merchant on either platform does not apply for anything, which also means they do not control the timing of what gets pulled.

Direct feed access sits behind a waitlist

Other retailers can apply for direct feed access, and that access is currently waitlisted. Search Engine Journal did not report how long the wait runs, how applications are prioritised, or what share of applicants has been accepted, so any number attached to the waitlist would be invented.

Feed providers connect through the Agentic Commerce Protocol

Feed providers including Salesforce and Stripe connect through the Agentic Commerce Protocol, which was established on 24 March. A store already working with a connected provider therefore has a route in that does not depend on the direct waitlist. OpenAI also plans a self-serve platform later in 2026 that would let stores connect feeds on their own, which would change the access picture again once it exists.

Shopping itself is live in the United States. The research and the reporting on it do not extend the availability statement beyond that.

What OpenAI has said, and what it has not

OpenAI's stated position on the results themselves is direct: "Product results are selected independently by ChatGPT and are not ads, nor influenced by any OpenAI partnerships." That statement addresses commercial influence over placement. It does not address retrieval source, which is the thing Profound measured.

On the 10 July 2026 change, OpenAI has not confirmed that anything happened, and has not attributed the shift to any model release. GPT-5.6 was released on 9 July 2026, one day before the jump, but the ChatGPT release notes carry no shopping-related update for 9 or 10 July 2026, and no public statement ties the two together. The proximity is real and the causal link is not established, which is exactly the distinction a merchant should keep when planning around this.

OpenAI has also not published specifics on how often ChatGPT relies on feeds versus the open web. Everything in the 65% figure is one vendor's observation, not a disclosed system behaviour.

The limits worth carrying into any decision

Four limits sit on top of this research, and none of them are cosmetic.

• The data is observational and comes from one vendor's tracked prompts. It is not ChatGPT Shopping traffic, and it is not a sample of real purchase journeys.
• It shows where recommendations were retrieved from, not how products were ranked once the feeds were pulled. A merchant inside the feed can still be the product that never gets named.
• OpenAI has neither confirmed the 10 July 2026 change nor described the retrieval split, so there is no primary source to check the observation against.
• A GPT-6 Astra rollout began on 3 September 2026, which can itself affect how consistently the model responds, and that rollout overlaps the window the 65% figure was measured in.

None of this makes the finding useless. It makes it a strong signal with a known ceiling on certainty, and the honest response to a strong signal is to prepare for it rather than to bet the quarter on it.

What this means for Thai marketers

Start with what the source did not say. Neither Profound's research nor the Search Engine Journal report mentions Thailand, Thai merchants, or availability outside the United States beyond the statement that Shopping is live in the U.S. Any claim that this has landed in the Thai market, or that it is about to, would not be coming from this reporting.

What does apply is the exposure a Thai business already has. A brand selling into the United States through Shopify is inside the catalogue that produced roughly 35% of tracked feed retrieval in July 2026, whether or not anyone in the business decided to be. A Thai exporter on a custom platform selling to U.S. buyers has the waitlist as its route. A purely domestic Thai retailer has no immediate exposure from this particular change, and treating it as urgent would be a misreading.

The wider point is that catalogue data is turning into a discovery surface in its own right, which sits alongside the rest of an ecommerce marketing programme rather than replacing any of it. For teams already investing in generative engine optimisation, this research says the ecommerce version of that work runs through the feed rather than through the blog.

What to check in your own setup

Nothing below requires a decision about ChatGPT specifically. The work behind ChatGPT product feeds is ordinary catalogue hygiene, and it pays off across every shopping surface a store already sells on.

• Confirm which catalogue integration your store actually has. A Shopify or Etsy store is connected without any action having been taken, so the question is what is in the catalogue, not whether it is connected.
• Check product titles for the words a buyer would use rather than internal SKU language. A feed carries the title as written.
• Fill in structured attributes rather than leaving them blank and describing them in prose. Retrieval reads fields.
• Look at whether descriptions answer the question a shopper asks an assistant, which is usually about fit, compatibility or material, and not usually about brand story.
• If the store is not on a connected platform, find out whether an existing payments or commerce provider already connects through the Agentic Commerce Protocol before joining a waitlist.
• Track branded and category prompts over time rather than checking once. A single spot check on one day tells you nothing about a retrieval change.

Frequently asked questions

Is ChatGPT Shopping live in Thailand?

The source did not say. Search Engine Journal's report on 9 September 2026 states that Shopping is live in the U.S. and makes no statement about Thailand or about any other market, so there is no reported answer either way.

Do I have to do anything if I run a Shopify store?

No action is needed to be connected, because Shopify stores are integrated automatically through the Shopify catalogue. What is worth doing is checking what that catalogue contains, since the feed data is what gets retrieved and the merchant controls the fields.

Does this mean web content no longer affects ChatGPT shopping results?

The research does not support that conclusion. It reports that feed retrieval accounted for roughly 65% of tracked recommendations as of 3 September 2026, which leaves a meaningful remainder, and it does not describe how the two sources are combined inside a single answer.

Did OpenAI change the model on 10 July 2026?

OpenAI has not confirmed any change on that date. GPT-5.6 was released on 9 July 2026, but the ChatGPT release notes contain no shopping-related update for 9 or 10 July 2026 and no public statement connects the release to what Profound observed.

Are the products ChatGPT names paid placements?

OpenAI says they are not, stating that "Product results are selected independently by ChatGPT and are not ads, nor influenced by any OpenAI partnerships." Profound's research is about which data source the products were retrieved from, which is a separate question from whether placement is sold.

Where this leaves merchants

The finding to act on is narrower than the headline. One vendor's tracked prompts show ChatGPT Shopping leaning on structured catalogue data, with a measurable narrowing of which merchants get named, and a plausible but unconfirmed link to a model release. A store cannot control retrieval. It can control what the feed says about its products, and on this evidence that file has started doing work it was never built for.

The reporting behind these figures is Search Engine Journal's write-up of Profound's research, published on 9 September 2026 and available on Search Engine Journal. If your catalogue already sells into the United States, an audit of the feed fields is the cheapest move available while the picture is still settling, and Relevant Audience takes on that kind of review.