On 7 and 8 October 2026, OpenAI shipped GPT-6 and its Intelligent UI to paid, Free and Go tiers. Within days, ChatGPT shopping results stopped looking like a row of product cards pinned above the answer and started looking like the answer itself — a numbered list of products written into the body copy, each one carrying its own way to buy.
On 21 October we sat down and measured it. In a signed-out Chrome window we ran 28 shopping prompts, opened 3 shopping side panels, and clicked and logged every one of the 97 buy links those answers produced. This article is what we found.
Key takeaways
- Products moved into the answer body. Clicking a product name now opens a description panel, not a list of merchant quotes.
- The format is set by the prompt, not the category. Our 28 prompts produced 11 distinct display formats.
- Of 97 buy links, 51 opened a retailer's product page and only 20 opened a brand's own product page.
- 96 of 97 links carried
utm_source=chatgpt.com; 17 also carriedutm_medium=feed, meaning the link came from a merchant catalogue rather than a web page. - Sources collapsed to a single site name at the end of each paragraph, with the rest folded behind a tap.
- Three separate things changed in October — the model, the attribution plumbing, and the ad business. Treating them as one "GPT-6 revamp" will lead you to the wrong conclusion.
What actually changed in the answer
Before the revamp, a shopping answer put a horizontal carousel of product cards in the middle of the response. Each card showed an image, a price and a merchant count — "De'Longhi and 9 others" — and clicking it opened a side panel listing every merchant's price next to a Buy button.
When we asked "Recommend three sleep earbuds for a side sleeper in the US." on 21 October, ChatGPT instead wrote a heading — "My top 3 sleep earbuds" — and listed three products inside the answer. Each entry had an image, a label such as "Best overall", a price range, two or three sentences of reasoning, a source marker, and one line reading "Shop at Best Buy". Clicking the product name opened a panel containing that product's image and description. There was no quote list.
That is a bigger change than it looks. The carousel was a shopping widget bolted onto a text answer; the list is the text answer. Products are no longer a separate surface you can win or lose independently of the copy. They are the copy.
Three timelines, one month
The most common analytical mistake we have seen since October is attributing everything to GPT-6. Three independent things landed inside three weeks.
Date | What happened | Did OpenAI say so publicly? |
|---|---|---|
1 Oct | Virtual "Try on" and product saves for apparel and accessories | Yes |
5 Oct | Visual ads beside image-generation results, tested in the US first | Yes |
6 Oct | Product labels such as "best all-rounder" appear in testing | No — found by third parties |
7–8 Oct | GPT-6 and Intelligent UI ship: paid tiers on the 7th, Free and Go on the 8th. Paid runs GPT-6 Sol, Free and Go run GPT-6 Luna. Answers may now combine text, images, buttons, forms and charts. | Yes |
8–15 Oct | Outbound links drop | Confirmed as a bug, no public statement |
10 Oct | Black "Shop at" buttons start appearing in shopping answers | No — third-party tracking |
21 Oct | ChatGPT Ads adds 26 markets, reaching 89 countries, on Free and Go only | Yes |
OpenAI publicly documented GPT-6, Try on and the ads expansion. It did not publicly document the change to shopping recommendations or to how sources are displayed. Everything in this article about those two areas comes from hands-on testing and third-party tracking, not from an official spec. OpenAI also runs different versions for different users at the same time, so your interface may not match our screenshots.
Where the 97 buy links actually went
We clicked every buy link in all 28 answers and recorded the destination page type.
Retailer product pages dominate. Brand product pages are a distant second — and that is with several prompts explicitly asking for official-site links, which inflates the brand share relative to a plain prompt.
The weakest outcomes are the ones that look fine in a report. When we asked about laptops for a college student, "View at Apple" opened apple.com's MacBook Air family page, not a specific configuration. In an AirPods Pro 3 price comparison, the Walmart and Target rows opened category pages and the Amazon row opened a search results page. A buy link that lands on a family page, a category page or a search page hands the selection work straight back to the shopper. It counts as a click in your analytics and converts like a bounce.
For what the buttons specifically do — and why the "Shop at" button is the single most consequential element in the new layout — see our companion piece on where the "Shop at" button actually sends shoppers.
A buy link is not a citation
The new layout contains two completely different link systems, and conflating them will wreck your reporting.
Buy links | Source markers | |
|---|---|---|
Where they appear | Inside the product entry: "Shop at Best Buy", "Official site", "View at Apple" | At the end of a paragraph: a site name with "+N" |
What they are for | Sending the shopper to a place that sells the product | Attributing the claim in the paragraph |
Typical destination | Retailer or brand commerce page | Review, comparison, listicle or editorial page |
What they tell you | Whether you are on the shelf | Whether you are in the evidence |
Since May 2026, brand names mentioned in an answer also link directly to the brand homepage. That is a buy link, not a citation.
The two systems are connected in one direction only, and it is the direction that matters: being in the sources makes it far more likely that the buy link points at you. That relationship is quantified in the Shop at button article, and the mechanics of becoming a source are covered in how AI search picks sources and how to get cited by ChatGPT.
What we could not answer
We would rather state the limits than let the numbers travel further than they should.
- We cannot say how common each display format is across all answers, or whether different accounts, regions and experiment groups see different formats.
- We cannot confirm the top-of-answer product carousel is fully retired. It did not appear once in 28 prompts, but 28 prompts is a small sample.
- We did not isolate whether signed-in sessions reveal search terms because of the session state itself or because of the model those sessions run.
- We did not inspect whether a black button's destination URL is present in the page data before the click.
- OpenAI has not published a specification for any of this.
What to do about it
The short version, in dependency order: make sure your product data is readable in server-rendered HTML; make sure your catalogue feed is actually submitted and accurate; get your pages into the answer's sources; split your analytics by utm_source and utm_medium; and plan paid and organic on the same board. Each of those layers has its own article in this series — the full sequence is laid out in how to get listed in ChatGPT shopping.
Underneath all of it is a measurement problem. Clicks and conversions only record the bottom of the funnel. If you want to know whether a change to your feed or your content actually worked, you have to track whether you are being recommended at all — your Share of Model across ChatGPT, Google AI Mode, Gemini, Google AI Overviews and Microsoft Copilot. That is what GEOly's AI shopping monitoring is built to measure.
FAQ
Did ChatGPT remove product cards entirely? We did not see the top-of-answer carousel in any of our 28 prompts on 21 October, and products appeared inline instead. We cannot confirm it has been removed everywhere — OpenAI tests different versions with different users simultaneously.
Why do my ChatGPT referrals look like they collapsed in mid-October? Between 8 and 15 October, ChatGPT's outbound links dropped the utm_source parameter, so analytics tools recorded those visits as direct traffic. This was confirmed as a bug and fixed on 15 October. Exclude that week from any period-over-period comparison — see utm_source=chatgpt.com explained.
Does ChatGPT get product data from my website or from a feed? Increasingly from a feed. Links carrying utm_medium=feed originate from a merchant catalogue submitted to OpenAI rather than from a crawled page. We cover the shift and how to get into it in the ChatGPT product feed.
Can I pay to be recommended? No. ChatGPT Ads and product recommendations are different systems on the same screen. Advertising buys reach on the Free and Go tiers; it does not buy a recommendation slot. See ChatGPT Ads now reach 89 countries.
Method
On 21 October 2026 we submitted 28 shopping prompts in a signed-out Chrome incognito window — 27 targeting the US market and 1 in Chinese targeting mainland China — and opened the shopping side panel in 3 of the resulting answers, clicking through each item. For every buy link we recorded the visible text, the URL, the channel parameters and the destination page type. Element types were classified by how the browser presents them to a user: a link can be hovered to reveal its URL and right-clicked to copy it; a button cannot. The signed-out versus signed-in comparison was run separately across several accounts. All 20 forensic screenshots in this series are unedited captures from that session.
External datasets referenced across this series — WISLR, Profound, Evertune, Search Engine Roundtable / Adobe, BusySeed — are attributed to their original authors with sample size and date at each point of use.
