We expected ChatGPT's shopping answers to vary by product category. They barely do. What they vary by is how the question is phrased — and the variation is not cosmetic. Different phrasings produce different buy links, pointing at different kinds of pages, owned by different companies.
Across 28 shopping prompts run in a signed-out Chrome window on 21 October 2026, we counted 11 distinct display formats. Nine of them are below, each paired with the prompt that produced it.
Key takeaways
- 28 prompts, 11 display formats. The prompt, not the category, decides which one you get.
- The default is an inline list with a text buy link — 18 of 28 runs.
- Asking for "product cards with prices and buy buttons" produces black buy buttons, which usually carry
utm_medium=feed— meaning the destination came from a merchant catalogue. - Asking for official links moves the buy link from a retailer to the brand's own site, though not always to a product page.
- Asking to compare merchant prices produces a quote table — and tends to surface category and search-results pages rather than product pages.
- Repeating the same question does not reliably give the same answer: one tracking study found the top-recommended brand changed between adjacent weeks 56.4% of the time.
The nine formats
Reading that grid as a practitioner, four things stand out.
The default is a text link, not a button. Eighteen of 28 runs produced the plain inline list with a "Shop at …" line. If you are modelling what most shoppers see, model this.
Buttons mean feed. When we explicitly asked for product cards and buy buttons, the answers rendered full-width black buttons — and those URLs disproportionately carried utm_medium=feed, meaning the destination came from a submitted product catalogue rather than a crawled page. The display format and the data source are correlated. See the ChatGPT product feed.
Comparison prompts dilute the destination. The merchant price table looks like the most commercially useful format and produces some of the weakest links: in our AirPods Pro 3 comparison, two rows opened retailer category pages and one opened a search results page.
Some formats have no way to buy at all. Asked not to browse and not to give links, ChatGPT still rendered product images and labels — pure brand exposure with zero attributable traffic. That is a real outcome your analytics will never show you, and one more reason clicks alone are the wrong scoreboard.
The prompt-to-outcome map
How the question was asked | What came back | Buy-link shape |
|---|---|---|
Plain recommendation, no qualifiers | Inline list with text links — 18 of 28 runs | Mostly retailer product pages or brand homepages |
"Product cards with prices and buy buttons" | Full-width black button under each product | Black buy button, often |
"Include official brand website purchase links" | Text-only list, one "Official product page" per item | Brand site — product, home or category page |
"Compare prices across merchants" | Merchant quote table, several sellers per product | Category and search-results pages appear |
"Show as a grid or carousel" | Two-column large images with price, merchant, size, rating | No horizontal carousel appeared in this session |
Chinese prompt asking for domestic storefronts | Three buttons side by side: brand site, official store, JD search | Several buy buttons for one product |
Buyer has not chosen a category yet | A few category suggestions, then a multiple-choice questionnaire | None |
"Do not browse, do not give links" | Product images and labels retained | None |
One question, two answers
This is the cleanest demonstration we captured. The question was "Recommend three sleep earbuds for a side sleeper in the US." Soundcore Sleep A30 came first, with a buy link to Best Buy; the Sleep A20's link went to Target.
Appending one sentence — "Include official brand website purchase links for each." — moved both links to soundcore.com and us.anker.com product pages.
Same products, same ranking, same session. The buy link moved from retailer to brand because of a single clause the shopper typed.
Two conclusions follow, and they point in opposite directions:
- Brands have less control than the demo suggests. Almost no real shopper appends that sentence. The default phrasing sends the click to a retailer, and you cannot phrase the shopper's question for them.
- It confirms the data is there. ChatGPT knows the brand's product page exists and can route to it. The default simply does not favour it — which puts the problem back on source presence, covered in the "Shop at" button and how AI search picks sources.
What the model actually searched for
One more thing changes the picture. Signed out, ChatGPT shows only "Searching the web". Signed in, it shows the terms it is searching.
The shopper asked: "My husband has tried several sleep earbuds but they're uncomfortable or fall out during the night. Which ones actually stay in?"
ChatGPT searched: "best sleep earbuds stay in ears overnight side sleepers 2026 reviews".
It added a year, a content type and an audience qualifier that were not in the question. OpenAI's own help documentation notes that ChatGPT typically rewrites a question into one or more more-specific queries. The practical consequence for content teams: optimise for the rewritten query, not for the conversational one. The pattern we observed was year + category + audience or situation + "reviews". Methods for finding those are in how to find grounding queries and query fan-out.
The repeatability problem
A caution before anyone builds a dashboard on single observations. Our prompt-to-format mapping comes from one session. Asking the same question again changes which merchants appear — and the instability is larger than most teams assume. BusySeed tracked 160 purchase-intent questions across three AI assistants between 27 April and 3 August 2026 and found the top-recommended brand changed between adjacent weeks 56.4% of the time.
That number is the argument against one-off audits. A single prompt run tells you what happened once. Only repeated sampling over time tells you your actual position — which is what Share of Model measures, and what GEOly's AI shopping monitoring runs on a schedule across ChatGPT, Google AI Mode, Gemini, Google AI Overviews and Microsoft Copilot.
FAQ
Does the way I phrase a ChatGPT shopping prompt change the results? Yes, substantially. In our test, 28 prompts produced 11 display formats, and adding a single clause about official links moved buy links from retailers to the brands' own sites.
What is the default ChatGPT shopping format? An inline list inside the answer body with a text "Shop at …" link under each product. It accounted for 18 of our 28 runs.
Why do some products show a black button and others a text link? The display format tracks the prompt, and the black button correlates with catalogue-sourced data — those URLs disproportionately carried utm_medium=feed.
Can I make ChatGPT link to my store instead of Amazon? Not by asking on the shopper's behalf. You influence it by being cited among the answer's sources and by submitting accurate catalogue data. Cited brand pages get the buy button 34% of the time versus 6% when uncited.
Will I see the same formats? Possibly not. OpenAI runs different versions for different users simultaneously, and these observations come from one signed-out session on 21 October 2026.
Method and limits
28 shopping prompts submitted in a signed-out Chrome incognito window on 21 October 2026 — 27 for the US market, 1 in Chinese for mainland China — with the shopping side panel opened in 3 answers and every item clicked. All screenshots are unedited captures from that session, composed into plates by theme.
What this does not establish: the frequency of each format across all answers, whether different accounts, regions or experiment groups see different formats, whether the top-of-answer carousel is fully retired, or whether the signed-in search-term display is a function of session state or of the underlying model. The weekly-volatility figure is from BusySeed, 160 purchase questions × 3 assistants, 27 April – 3 August 2026. The full series method is in how ChatGPT shopping results work after GPT-6.
