ChatGPT Shopping ranking factors: what to fix so your products get picked

OpenAI's help centre describes product selection and merchant selection separately, and the factors differ.

• A product appears in the carousel when ChatGPT perceives it to be relevant to the shopper's intent, judged from the question and its context, which includes Memory and custom instructions. OpenAI's example is a request for dog costumes, where ChatGPT may consider available options, price, reviews and ease of use, and leave out clown costumes if the shopper has said they dislike clowns.
• When the shopper selects a product, ChatGPT may show the merchants offering it, ranked, in OpenAI's words, on availability, price, quality and whether they are the maker or primary seller.

A brand selling its own product therefore holds an advantage on the second decision that a reseller does not. OpenAI adds that it expects merchant ranking to evolve, including more personalisation to the user.

OpenAI lists three inputs:

• Structured metadata from first-party and third-party providers, its examples being price and product description, together with other third-party content.
• The model's own response, formed before it considers new search results.
• OpenAI's safety standards and product policies.

The first input is the one a retailer can change. Provider data reaches ChatGPT through Shopify Catalog for Shopify stores, through direct feeds for merchants OpenAI has granted feed access, and through OAI-SearchBot's crawl of product pages. Sites that opt out of that crawler are not shown in search answers, only as navigational links.

Profound's analysis of June 2026 cards put the split for the cards it observed at 87.3 percent web crawl and 12.7 percent direct feeds, so for most stores the product page is the record ChatGPT reads. Weighting shifts with the request: OpenAI's own example is that a stated budget of 30 dollars makes price count for more.

Product results are selected independently by ChatGPT, are not ads and are not influenced by OpenAI partnerships; advertising is a separate surface. Shopify says there are no fees for selling in ChatGPT beyond payment processing. Reviews and ratings are not verified by OpenAI, and the review summaries and labels on cards, such as Budget-friendly or Most popular, are generated by the model from the information available to it rather than awarded.

Two further points matter for anyone optimising copy. ChatGPT may generate simplified product titles and descriptions from provider information, because merchants describe the same item differently, so a title tuned as a slogan may never be shown as written. And OpenAI says not every available product is shown, so an eligible product that does not appear has lost on the factors above rather than failed a rule.

Two studies looked at the output rather than the documentation.

• Profound (Allen Wu, 23 July 2026) analysed 201,137 prompt runs and 812,190 product cards from 18 to 25 June 2026 and compared cards at rank one with cards at rank four and lower. The largest differences were a ChatGPT-generated tag on the card (2.39 percent of rank-one cards against 0.98 percent) and the median review count (787 against 352). Price promotions showed only a modest lift, and the presence of offers, URL length and title length barely moved. Reddit supplied about a third of the citations behind shopping answers.
• Semrush (January 2026) ran 100 prompts five times each, extracted the shopping query ChatGPT issued behind the scenes and ran it in Google Shopping: the top ChatGPT product was in Google Shopping's first three results 75 percent of the time, and the authors called their sample small.

Both are observations of correlation in one period. Neither shows that adding reviews or ranking on Google Shopping causes a product to be picked, only that picked products tend to have both.

1. Access. OAI-SearchBot allowed in robots.txt and through the firewall; nothing else matters while it is blocked.
2. Availability and price that are right today. Both are named merchant factors, prices come from provider data, and OpenAI says updates can lag, so a stale feed is judged on last week's price and stock.
3. A complete record. In a feed, the nine required fields with valid values and a GTIN that passes its check digit. On a page, the same facts stated in text a crawler can read; OpenAI names structured metadata from providers as an input but does not list which page markup it parses.
4. Reviews ChatGPT can read. It summarises reviews from public websites and takes ratings from providers, and the feed format has optional rating and review-count fields. It follows, though OpenAI does not say so, that reviews on a page the crawler cannot reach do not help.
5. Maker or primary seller, stated. OpenAI does not say how it tells a maker from a reseller; the fields it has are brand, seller name and identifiers, so fill them accurately and consistently.
6. Descriptions that answer constraints. The request carries use, fit, material, size and occasion; a description that states those facts gives the model something to match.

The sixth signal is the one Sighted's own experience bears on. SO-ME, the consumer shopping platform Sighted powers, takes requests in the shopper's words (a waterproof jacket that packs small for a weekend away, say) rather than the two keywords a search box would get, and a standing difficulty in running it is matching those requests to retailer feeds.

The feeds from ASOS, REVOLVE, Vinted, M&S, New Look, Debenhams and EGO differ in which attributes they carry and how they categorise the same product, and a request that asks for a fact no feed states cannot be matched by any ranking logic.

Conversational discovery needs deeper product information than keyword search ever did, which is the gap large-scale enrichment exists to close. In Sighted's pilot with The Game Collection, Sighted enriched 200 products and created 17 collections. Over the pilot, AI citation coverage went from 0% to 32% of the priority products tested, and the store's organic conversion rate rose 64%.

Relevance to the shopper's intent is the first product factor OpenAI lists, and relevance is judged on what the record states. A product whose page and feed name the edition, the platform, what the bundle includes and who it suits has more to be matched on than one that names the title and a price. How the enrichment itself works, with a worked example and the review step, is its own guide.

Do reviews decide whether ChatGPT recommends a product?

They are one of the factors OpenAI names, and the Profound sample found higher-ranked cards carried more of them, but OpenAI says the weighting depends on the request and that the ratings are not verified. A product with few reviews can still be shown when it fits the constraints better.

Do Shopify stores rank higher?

OpenAI says the Shopify Catalog integration helps products appear more accurately and completely. It does not say Shopify products rank higher, and the merchant factors it lists apply to every store.

Can you pay for a place in the carousel?

No. Product results are not ads, and ads are separate from them.

Should you optimise for Google Shopping instead?

Semrush's test found most top ChatGPT picks among Google Shopping's first three results, which is a reason to keep the Merchant Center feed accurate. OpenAI documents no such dependency, so treat it as a correlation to use, with no guarantee behind it.

None of the factors applies to a product ChatGPT never received. A Shopify store that does not sell to United States customers, a product Shopify classifies as Uncategorized, a feed row with an unrecognised availability value, or a robots.txt that disallows OAI-SearchBot each stop the product before any factor is weighed. So the order is:

1. Run the eligibility checks first, in the order the troubleshooting guide gives them, and do ranking work only on the products that pass.
2. Work through the six signals in order, starting with access, and treat price and availability as something to keep right every day rather than fix once.
3. Complete the record: the nine feed fields with valid values and a GTIN that passes its check digit, and the same facts in crawlable text on the page.
4. Write descriptions that state use, fit, material, size and occasion, so the request a shopper types has facts to match.
5. Keep the Merchant Center feed accurate as well, since the top ChatGPT picks were mostly among Google Shopping's first three results in Semrush's test.

The readiness score, out of 100 per product, weighs six things that map onto this list: structured data an assistant can read, real detail in the title and description, data in a valid format (a currency on the price, a recognised availability value, a GTIN that passes its check digit), crawler access, feeds ready for the assistants' channels, and whether reviews and store policies are stated. The free AI visibility audit reads up to 50 product pages and scores them on the same signals.

Inside the product, Sighted writes the missing attributes from the merchant's own descriptions, options and photos, in the store's voice and only from facts it can find, and exports the Merchant Center feed, and the ChatGPT Shopping feed for approved OpenAI partners, from the same corrected records. It is the technology behind SO-ME, built to enrich retailer catalogues at scale for exactly this kind of discovery.