Short answer: ChatGPT shopping shows a product when the model judges it relevant to what the shopper said they need, and it builds that judgement from four things OpenAI names: the shopper's query and saved context, structured product metadata from merchants and third-party providers, other third-party content its search finds, and the answer the model would have given before searching. Once a shopper opens a product, a second step ranks the merchants selling it on availability, price, quality, and whether the seller is the maker or primary seller. Neither step is paid placement. So the work of getting products recommended by ChatGPT splits cleanly in two: get accurate product data onto a path ChatGPT reads, and make each product the obvious answer to a specific, stated need.

This article is about the mechanism, checked against OpenAI's own help-center article and merchant page and against Shopify's documentation for its ChatGPT channel. Where those sources say nothing, we say so rather than guess. Brand-level advice (being named in ordinary answers, entity consistency, corroboration) is covered in how to get recommended by ChatGPT; this piece stays at the level of the product card and the merchant list.

  • Two inputs reach the carousel: structured product data (Shopify Catalog, Etsy, direct feeds, third-party providers) and web content from the search index. A Shopify or Etsy seller is already on the first path.
  • Selection is intent-first: the same catalogue produces different carousels for different shoppers, because Memory and custom instructions are inputs.
  • Merchant order is a separate ranking with four named factors. Being the maker or primary seller is one of them, which matters for brands that also sell through resellers.
  • Purchases complete on your own site. OpenAI's merchant page says it is moving away from standalone in-chat checkout, and there are no fees on purchases that start in ChatGPT.
  • Shopping is live for U.S. users only as of October 2026, so tests from other countries tell you nothing about the carousel.

What ChatGPT shopping is, as of October 2026

When a question signals shopping intent ("help me find a waterproof trail shoe for wide feet"), ChatGPT can answer with product options rendered as cards: an image, a simplified title, a price, and links to sellers. OpenAI's help center describes this as shopping with ChatGPT search. Clothing and accessories carry a "Try on" button that generates an image of the shopper wearing the item, and shoppers can save products into folders in their ChatGPT Library. The feature is available on web and mobile.

The three surfaces a shopper moves through

A shopper sees three distinct surfaces, and each is decided by a different rule. Treating them as one "ranking" is the first mistake most teams make when they start working on ChatGPT shopping.

SurfaceWhat decides itWhat a merchant can influence
The written answerThe model's reasoning about the question, informed by search resultsThird-party coverage, clear product facts on crawlable pages
The product carouselRelevance to the shopper's intent, using structured metadata, third-party content and the model's prior answerCompleteness and accuracy of product data; how well the product page states who it is for
The merchant list for one productAvailability, price, quality, and maker or primary-seller statusStock accuracy, price freshness, being the authoritative seller of your own products
CheckoutThe merchant's own site or appEverything, because it is your checkout

What changed with checkout

For a period, OpenAI promoted Instant Checkout, a flow in which the shopper paid inside ChatGPT and the merchant processed the order through its own payment provider. OpenAI's merchant page now says the company is moving away from a standalone Instant Checkout experience and prioritising merchant-owned checkout: shoppers discover and evaluate products in ChatGPT, then buy on the merchant's website or app. Shopify's help center describes the ChatGPT channel the same way, as a discovery-focused referrer: the buyer completes the purchase on the store's own checkout, in ChatGPT's in-app browser on mobile or a new tab on the web, so a store's payment methods, discounts and branding all apply.

The help-center article still mentions that some eligible products may show an Instant Checkout option, and the merchant-list labels it describes include one called "Instant checkout". Read that as a surface in transition. For planning, assume the click leaves ChatGPT and lands on your product page, which means the product page has to finish the sale on its own.

What it is not

Product results are not ads. OpenAI states that results are selected independently by ChatGPT, are not influenced by any OpenAI partnership, and that ads are separate from product results. OpenAI also states there are no fees on purchases that start in ChatGPT. There is therefore no budget line that buys a better carousel position; there is only data quality, product-to-need fit, and the reputation the model has already absorbed.

Two ways products reach ChatGPT shopping: feeds and the open web

Every product in a carousel arrived by one of two routes. Knowing which route your catalogue uses tells you where a wrong price or a missing variant came from, and therefore where to fix it.

WHERE PRODUCT DATA STARTSWHAT CHATGPT RECEIVESWHAT THE SHOPPER SEESShopify CatalogEtsyDirect feed (approved)Third-party providersStructuredproduct metadataprice, description,availability, images,seller, identifiersYour product pagesand third-party pagesSearch index(OAI-SearchBot) andlive fetch (ChatGPT-User)Product carousel:items judged relevantto the shopper’s intentMerchant list:sellers ranked for theselected productPurchase completes onthe merchant’s ownsite or appNeither path is paid placement. OpenAI does not publish how the two paths are weighted when they disagree.
Figure 1. The two paths into ChatGPT shopping: structured product data from feeds and providers, and web content from OpenAI’s search crawler. Both inform the carousel; the merchant list and checkout follow. Source: OpenAI Help Center and merchant page, Shopify Help Center (read 2026-10-23).

The feed path: structured product data

OpenAI's help center says ChatGPT considers structured metadata such as price and description from first-party and third-party providers, and that the merchant list is built from merchant and product metadata supplied by third-party providers or by merchants directly. In practice the feed path has four entrances:

  • Shopify Catalog. OpenAI states that product data for merchants on Shopify is already integrated through Shopify Catalog and no additional work is required. Shopify's own requirements for the ChatGPT channel: the store must sell to customers in the United States (it may be based elsewhere), its products must be eligible for Shopify Catalog, and its terms of service, privacy policy and return and refund policy must be filled in.
  • Etsy. OpenAI's merchant page groups Etsy sellers with Shopify sellers: already integrated, nothing to apply for.
  • A direct feed. Other merchants apply on OpenAI's merchant page to share a product feed. OpenAI's developer documentation says feed onboarding is currently open to approved partners, and the merchant page says applicants join a waitlist while OpenAI builds a self-serve merchant platform.
  • Third-party providers. OpenAI repeatedly refers to metadata "from third-party providers" without naming them. Prices in product listings, it says, are received from third-party providers. You do not control this source, which is why the other three matter.

The crawl path: what the search index sees

The other route is the open web. OpenAI runs OAI-SearchBot to build the index behind ChatGPT's search features and ChatGPT-User to fetch a page when a person in ChatGPT asks for it. Product pages, review sites, editorial round-ups and forum threads all enter through this door as "other third-party content". OpenAI's merchant page answers the obvious question directly: a product feed is not required if ChatGPT already crawls your site, but a feed gives you more control over how products appear and how current the information is.

The two paths are independent. Shopify's documentation spells out the consequence: blocking AI crawlers in robots.txt or at the network layer affects only open-web discovery; it does not stop Shopify Catalog from sending product data to the channels you have enabled. The reverse also holds. Turning off Catalog access for ChatGPT does not stop ChatGPT from finding your products through crawling and indexing.

When the two paths disagree

Suppose your feed says a jacket costs one price and a cached review page quotes last season's price. OpenAI does not publish which source wins. What it does say is that prices shown come from third-party providers, that the first price shown is the first-listed merchant's, and that merchant price and shipping changes may take some time to be reflected. The practical rule is to make the sources agree, not to bet on precedence: the feed, the product page, and the structured data on that page should carry the same price, currency, availability and variant name.

QuestionFeed path (Catalog, Etsy, direct feed)Crawl path (OAI-SearchBot, ChatGPT-User)
What it carries bestPrice, availability, variants, identifiers, images, sellerContext: reviews, comparisons, who the product suits, editorial verdicts
Who controls itYou (or your platform), field by fieldYou for your own pages; publishers and shoppers for the rest
FreshnessSnapshot plus updates; OpenAI recommends at least daily for direct feedsWhenever the page is next crawled or fetched
How to switch it offShopify: remove Catalog access per channel; direct feed: mark items ineligiblerobots.txt rules per agent; an edge block that ignores robots.txt
Typical failureVariant collapsed into one row; stale stock; placeholder brandBot blocked by mistake; price rendered by JavaScript; outdated third-party copy

OpenAI's description of carousel selection is short, and every clause in it has a practical consequence. A product appears when ChatGPT perceives it to be relevant to the shopper's intent, considering the query and context such as Memory or custom instructions. To decide which products to surface, it considers structured metadata, other third-party content, the responses the model generated before it considered new search results, and OpenAI's safety standards and product policies.

1The query, plus Memory andcustom instructions2Structured metadata fromfirst- and third-party providers3Other third-party contentfound by search4The model’s own answer beforenew search results, and policyRelevance tothis shopper’sintentCarousel: a subsetof products, notevery matchWHAT CHATGPT REWRITES BEFORE SHOWING ITSimplified titles and descriptions; labels such as“Budget-friendly”; review summaries from public sitesSource: OpenAI Help Center, “Shopping with ChatGPT Search” (read 2026-10-23). Policy covers safety standards and prohibited products.
Figure 2. The inputs OpenAI lists for choosing which products to surface, and what ChatGPT rewrites before the shopper sees a card. Source: OpenAI Help Center, “Shopping with ChatGPT Search” (read 2026-10-23).

Intent comes first, and intent is personal

OpenAI's own example is a shopper who wants goofy dog costumes: ChatGPT may weigh available options, price, reviews and ease of use, and if the shopper has said they dislike clowns, it may leave clown costumes out. If a shopper sets a budget of $30, price dominates; if price is never mentioned, other attributes take over. Two consequences follow. First, there is no single carousel for a query, because Memory and custom instructions change the result for each person. Second, the attribute that wins depends on what the shopper said, so a product page that states only a price and a tagline gives the model nothing to match against "for wide feet", "quiet enough for an apartment" or "safe for a puppy".

This is the product-level version of a rule that runs through all AI visibility work: attributes are what gets matched. Our guide to product-level visibility covers where those attributes get lost between your catalogue and a crawler.

Structured metadata is the floor, not the ceiling

The metadata a feed carries (title, description, price, availability, images, identifiers, seller) is what lets ChatGPT render a card at all and keep it accurate. It is necessary rather than sufficient: a perfectly formed record for a product that does not fit the stated need still will not be chosen. OpenAI's feed specification makes the point in its own words: eligibility does not guarantee display.

Third-party content and the model's prior answer

The two inputs merchants most often overlook are "other third-party content" and the model's own answer generated before it considers new search results. Together they mean that the product's reputation outside your site, and whatever the model already believes about your brand and category, shape the shortlist before your feed is consulted for details. A product that reviewers, comparison articles and forum threads consistently describe as the right choice for a situation has a head start that no feed field can buy. The feed then decides whether the card that appears is accurate.

Policy removes some products entirely

OpenAI's commerce documentation lists prohibited categories, including adult content, age-restricted products such as alcohol, nicotine and gambling, weapons, prescription-only medicines and unlicensed financial products. Shopify notes that products on OpenAI's prohibited list do not display in ChatGPT product discovery results. If you sell in one of these categories, the carousel is not a channel you can optimise for; the written answer and your own site are what remain.

What ChatGPT rewrites before the shopper sees it

The card is not a copy of your listing. OpenAI says ChatGPT may generate simplified product titles and descriptions from third-party information, because merchants use varying titles for the same item. It adds feature labels such as "Budget-friendly" or "Most popular", which it describes as model-generated and not verified statements; a "Budget-friendly" label might reflect reviewers mentioning good value rather than the lowest price. It may also show review summaries generated from reviews on public websites, which OpenAI does not verify. You cannot set these labels. What you can do is make sure the facts they are generated from are correct and consistent wherever they appear.

What our benchmark shows about ChatGPT and product questions

The carousel is a user-interface feature; it does not appear in API responses. What we can measure at scale is the written answer the same model family gives to buying questions, and that is still informative, because the model's prior answer is one of the four inputs OpenAI lists. In our Q3 2026 benchmark we asked ChatGPT (openai/gpt-5-mini through the API, U.S. market) 20 buying questions per category and classified the form of each answer.

ANSWERS OUT OF 20 PER CATEGORY (CHATGPT VIA API, US, MEASURED 2026-09-07)Electric bikes3215Furniture3611Pet supplies3512Skincare4412Ranked list of optionsUnranked list of optionsGuide, no option namedSource: InsightWonder Q3 2026 benchmark, openai/gpt-5-mini, 20 questions per category.Answer form classified by a model. These are API answers, not the shopping carousel.
Figure 3. Form of ChatGPT’s written answers to 20 buying questions per product category. Most were guides that named no option. Source: InsightWonder Q3 2026 benchmark, openai/gpt-5-mini via API, U.S. market, collected 2026-09-07; answer form classified by a model.

In all four product categories, most answers were buying guides that named no option at all: 15 of 20 for electric bikes, 11 for furniture, 12 for pet supplies and 12 for skincare. Ranked lists were rare, three or four per category. Read carefully, this says something useful about where products enter. When the model answers in prose without a product interface, it tends to teach the shopper how to choose. The product interface, which runs on structured product data, is where specific items appear. A brand that only works on being named in prose is optimising the surface that names products least often.

The caveats are real. These were API calls without Memory or a logged-in profile, the answer form was classified by a model, and 20 questions per category is a sample, not a census. The data was collected on 2026-09-07. For the full category results, including which brands were named, see the electric bikes benchmark page and its siblings.

Where the cited evidence comes from

The same benchmark recorded every source the five models cited. Across ChatGPT, Perplexity, Gemini, Claude and Grok, the share of citations that pointed at a named brand's own website was small in every product category we measured.

SHARE OF ALL CITATIONS POINTING AT A NAMED BRAND’S OWN SITE (FIVE MODELS, US, 2026-09-07)Electric bikes18.6%Furniture15.2%Pet supplies9.3%Skincare3.6%Source: InsightWonder Q3 2026 benchmark; ChatGPT, Perplexity, Gemini, Claude, Grok; 20 questions per category.The rest went to publishers, retailers, communities and other domains.
Figure 4. Share of all citations, across five models, that pointed at a named brand’s own website, by product category. Source: InsightWonder Q3 2026 benchmark, 20 questions per category, collected 2026-09-07.

Electric bikes had the highest share at 18.6%, furniture 15.2%, pet supplies 9.3% and skincare 3.6%. Everything else went to publishers, retailers, communities and other sites. For ChatGPT shopping this is a reminder of what "other third-party content" means in practice: most of the evidence a model sees about a product category was written by someone other than the brand. Your product page has to be correct, but it is outnumbered.

How ChatGPT ranks merchants once a product is chosen

When a shopper taps a product, ChatGPT may show a list of merchants that sell it. This is a second, separate ranking, and it is the one that decides whether the click goes to you, to a reseller, or to a marketplace listing of your own product.

A product has been chosenAvailabilityCan this sellership it now?In stock beatsout of stock orunknownPriceWhat does it costhere vs elsewhere?Lowest known pricemay get “Best price”QualityHow good is thisseller as a source?Not defined furtherby OpenAIMaker / primaryIs this the brandor its main seller?Can favour thebrand’s own storeover a resellerLABELS“Best price” can give way to another label such as “Instant checkout”.The first price shown is the first merchant’s, which is not always the lowest.Source: OpenAI Help Center (read 2026-10-23). OpenAI lists the factors; it does not publish their weights or order.
Figure 5. The four factors OpenAI names for ranking merchants once a product is chosen, and how price labels behave. OpenAI publishes the factors, not their weights. Source: OpenAI Help Center (read 2026-10-23).

Availability

The first factor OpenAI names is availability. A seller that cannot ship the item is a poor answer, so stock status needs to be true at the variant level: "in stock" for the black size 10 when only the white size 9 is left will send a shopper to a dead end and, over time, teach the system that your data is unreliable. Availability is also the field most likely to go stale, because it changes faster than anything else in a catalogue.

Price, and the "Best price" label

Price is the second factor, and OpenAI describes two behaviours worth knowing. The price shown in ChatGPT's first response typically reflects the first-listed merchant, which may not be the lowest. On the product detail view, ChatGPT works to highlight the lowest price it is aware of, and when more than one merchant offers the product, that option may be labelled "Best price". Another label such as "Instant checkout" can take precedence, in which case the lowest-priced option may not show "Best price" at all. If your price is genuinely lower but ChatGPT shows a reseller's, the usual cause is data freshness rather than ranking: OpenAI says price and shipping changes can take time to be reflected.

Quality

OpenAI lists quality without defining it. It could reasonably cover the seller's reputation, the completeness of the listing, or return and shipping terms, but none of that is documented, and we will not pretend otherwise. The defensible response is to remove obvious negatives: a complete return policy, honest shipping terms, and a listing whose images and descriptions match the product.

Maker or primary seller

The fourth factor is whether the merchant is the maker or primary seller of the item. For a brand that sells direct and through resellers, this is the most important line in the document, because it is the factor that can put your own store above a cheaper or faster reseller. To benefit from it, the system has to be able to tell that you are the maker: the brand field in your data should match the brand on the product page exactly, identifiers such as GTINs should be the manufacturer's, and your store should be recognisably the brand's own.

Personalisation is coming

OpenAI says it expects merchant rankings to evolve, including by personalising them further to user preferences. A shopper who prefers a particular retailer may see it ranked higher. That makes single screenshots even less reliable as evidence, a point we return to in the measurement section.

Your situationFactor most in your favourFactor most at riskFirst thing to fix
Brand selling only on its own Shopify storeMaker or primary sellerAvailability, if variant stock is not syncedVariant-level stock and price in Shopify Catalog
Brand selling direct and through resellersMaker or primary sellerPrice, where resellers discountBrand and GTIN consistency so ChatGPT can identify you as the maker
Reseller of other brands' productsPrice and availabilityMaker or primary seller (you are neither)Fresh prices and stock; complete shipping and returns information
Marketplace-only sellerDepends on the marketplace's dataControl: the marketplace owns the recordMake sure the marketplace listing carries full attributes

The procedure below follows the mechanism in order: first make sure a path exists, then make the data on that path trustworthy, then make the product the right answer for a need, and only then measure. Skipping ahead is the most common way to waste a quarter.

Step 1: Confirm which path your catalogue is on

If you sell on Shopify, open Sales channels, then Agentic, in the Shopify admin. Shopify shows Shopify Catalog under Sources with either the number of products in Catalog or an "Ineligible" warning; clicking it shows the reason. Check that ChatGPT access is on, and that your store policies are complete, since terms of service, privacy policy and refund policy are listed requirements. If you sell on Etsy, OpenAI says you are already integrated. If you run your own stack, apply for a direct feed on OpenAI's merchant page and, while you wait, make sure the crawl path works.

Step 2: Make sure OAI-SearchBot can read your product pages

Even feed-integrated merchants depend on the crawl path for context. Confirm that robots.txt does not block OAI-SearchBot, which is the agent sites most often block by accident while meaning to opt out of training via GPTBot. A minimal, explicit group looks like this:

# OpenAI: allow the search index and user-triggered fetches,
# decide separately about training (GPTBot).
User-agent: OAI-SearchBot
Allow: /
Disallow: /cart
Disallow: /checkout

User-agent: ChatGPT-User
Allow: /

User-agent: GPTBot
Disallow: /

Then test with a real request as the agent, because a CDN or bot-management rule can block an agent the file allows. Our working robots.txt for AI crawlers walks through the full file and the one-minute test. Check, too, that price, stock and variant names are in the server-rendered HTML rather than injected by JavaScript after load.

Step 3: Fix the product record at its source

Whatever the path, the record has to be correct where it originates, because every downstream copy inherits its errors. For each product, check:

  • One row per purchasable variant, with the variant in the title ("Trail runner, black, size 10"), not a single parent row with a price range.
  • A real brand name that matches the product page, not a placeholder and not the store name when the store resells other brands.
  • Manufacturer identifiers where they exist. Do not invent a part number to fill a missing GTIN; OpenAI's specification says explicitly not to.
  • Price and availability that update when they change, including when a sale starts or ends. OpenAI notes that dates in a feed do not schedule changes.
  • Return and shipping terms stated plainly. An omitted shipping price means "unknown" to the system, not "free".

On Shopify, much of this lives in the product admin and is sent by Catalog automatically. Shopify's Catalog Mapping can pull titles, descriptions and categories from metafields and can group separate products as variants of one listing, which helps stores whose data is spread across custom fields. Shopify also advises putting any legal disclosures in the first 6,000 characters of a product description. The field-by-field detail of a direct feed is a topic of its own and deserves a separate reference.

Step 4: Write the product page for the need, not the category

Because selection is intent-first, the product page has to say which needs the product meets, in sentences a model can lift: who it is for, the conditions it handles, the measurements that decide fit, and the trade-offs. "Waterproof trail shoe with a wide toe box, tested on wet rock; runs half a size small; not designed for snow" gives the model three needs to match and one to exclude. A tagline and a specification table give it almost nothing. State the trade-offs honestly. Products that are described accurately get matched to the right shoppers, which is the only kind of match that turns into a sale.

Step 5: Earn the third-party evidence

The benchmark figures above show how little of the cited evidence in product categories comes from brand sites. Reviews on independent sites, inclusion in round-ups that state their testing method, and genuine discussion in communities feed the "other third-party content" input and the model's prior beliefs. This is slow work and there is no shortcut that survives scrutiny; paid placements disguised as editorial tend to be discounted by readers and, increasingly, by the systems that summarise them.

Step 6: Keep one identity for the brand and each product

Name the product the same way in the feed, on the page, in structured data, on marketplaces and in press material. A product called "Trail Runner 2" in the feed, "TR2" on Amazon and "Trail Runner II" in reviews is three products to a system trying to match evidence to a card. The same applies to the brand: a legal entity name in the feed and a trading name on the site splits the maker signal.

Step 7: Measure before and after, then repeat

Define the needs you want to win, write them as the questions shoppers actually ask, and measure presence in the carousel and position in the merchant list before you change anything. The next two sections cover diagnosis and measurement in detail.

A worked example: one product traced through both steps

To make the procedure concrete, follow a single illustrative product through it. The brand, product and wording here are invented for the walk-through; the mechanism is the documented one. Imagine a small brand that makes a trail running shoe in a wide fit, sells it on its own Shopify store, and also sells it through two outdoor retailers.

The starting state. The store is eligible for Shopify Catalog and ChatGPT access is on, so a data path exists. But every colour and size is one product with a price range, the description is a brand story with no measurements, and the brand field reads as the company's legal name rather than the brand printed on the shoe. One retailer lists the same shoe with the manufacturer's identifiers and a detailed "wide fit, half-size small" note copied from a review.

What a shopper sees. A U.S. shopper asks for "a waterproof trail shoe for wide feet that handles wet rock". The carousel has the shoe, because reviewers on independent sites describe it that way. But the card shows the retailer's title and price, and when the shopper opens it, the retailer is listed first. On the evidence available to ChatGPT, the retailer looks like the authoritative seller: its record has identifiers, its description matches the need, and the brand's own record does not obviously belong to the same product.

The fixes, in order. First, the brand makes each purchasable colour and size its own variant with the variant in the title, and sets the brand field to the name on the shoe. Second, it adds the manufacturer's GTINs, the same ones the retailer uses, so that the two records resolve to one product with the brand as its maker. Third, it rewrites the product page around the needs it actually meets: the toe-box width, the waterproof membrane, the outsole compound on wet rock, the half-size-small fit, and the fact that it is not built for snow. Fourth, it checks that the page's structured data, the Catalog record and the visible price agree.

What to expect. None of this guarantees the top merchant slot; OpenAI publishes factors, not outcomes. What it changes is the evidence. The brand's record can now be matched to the product, its availability is accurate per variant, and the maker-or-primary-seller factor has something to work with. The measurement then tells you, over repeated clean runs, whether the brand's store appears in the merchant list more often than before. If it does not, the remaining gap is usually price freshness or third-party evidence, and the diagnosis table below points to the next check.

Diagnosing a product that never appears in ChatGPT shopping

When a product is missing, work through the gates in order. Each one is a precondition for the next, and a fix further down the list cannot help while an earlier gate is closed.

1Is the shopper in a market whereChatGPT shopping is live (the U.S.)?noNothing to fix yet: shopping isnot live thereyes2Does a data path exist (Catalog,Etsy, feed, or crawlable pages)?noOpen a path: check Catalog eligibility,apply for a feed, unblock OAI-SearchBotyes3Are price, availability, variantand seller fields complete and current?noFix the record at its source; pricechanges reach ChatGPT with a delayyes4Is the product a fit for aspecific need the shopper states?noWrite the need, the attributes andthe trade-offs on the product pageyes5Does it appear across repeated,logged-out runs of real prompts?noOne run proves little; sample theprompt set and record positionsWork top to bottom: a later fix cannot help while an earlier gate is closed.
Figure 6. Diagnosis order for a product missing from ChatGPT results. Each gate is a precondition for the next. Source: built from the OpenAI and Shopify documentation cited in this article.
SymptomLikely causeHow to checkFix
No products from your store appear for any queryStore not eligible for Shopify Catalog, ChatGPT access turned off, or the tester is outside the U.S.Sales channels, then Agentic, in Shopify admin; confirm test locationResolve the ineligibility reason; complete store policies; test from a U.S. session
Product appears with the wrong priceStale feed or provider data; first-listed merchant's price shownOpen the product card and compare merchantsUpdate the source record; make page, feed and structured data agree
Product appears but your store is not in the merchant listBrand or identifier mismatch hides that you are the makerCompare brand field and GTIN with what resellers publishUse the manufacturer's identifiers and the exact brand name everywhere
Only one variant ever appearsVariants collapsed into a single row or grouped incorrectlyInspect the feed or Catalog Mapping for that productOne row per purchasable variant with a shared group identifier
Product appears for generic queries but not for specific needsPage states no attributes for the needSearch the page text for the need's key termsRewrite the page around the needs, measurements and trade-offs
Results vary every time you lookPersonalisation, Memory, and normal model varianceRepeat in a logged-out or clean sessionSample repeatedly and report rates, not single screenshots

Edge cases that look like bugs

Several situations produce confusing results without anything being broken. Business-to-business products are excluded from Shopify's agentic channels when Shopify can identify them, for example products published only to B2B catalogues or behind a customer login; if you sell the same item to both audiences, the consumer price is the one sent. Products set to Unlisted in Shopify disappear from AI channels, but also from sitemaps, search engines and store search, so use that setting deliberately. After you turn off Catalog access, Shopify says it can take up to 7 days before data stops being shared. And for direct feeds, OpenAI retains a product's last processed record for up to 14 days after it disappears from a snapshot, so a product you removed can linger.

What you cannot control in ChatGPT shopping, and claims to ignore

A mechanism-level view is also a good filter for advice. Several claims circulate that the documentation does not support.

  • "You can pay for a top carousel spot." OpenAI states product results are not ads and not influenced by partnerships. Ads exist in ChatGPT, but they are separate from product results.
  • "Submitting a feed guarantees you will appear." Eligibility does not guarantee display, in OpenAI's own words, and not all available products will be shown for a query.
  • "Labels like 'Most popular' reflect sales data." OpenAI describes them as model-generated from available information, possibly including third-party data, and not verified.
  • "There is a ranking formula to reverse-engineer." OpenAI names factors, not weights, and says the merchant ranking will become more personalised. Anyone selling the formula is selling a guess.
  • "Results from my country show what U.S. shoppers see." Shopping is live in the U.S.; elsewhere, you are looking at a different product.

There are also real limits on what a merchant can change. You cannot edit the simplified title ChatGPT generates, set a label, remove a review summary, or choose which merchant is listed first. You can make the underlying facts correct, consistent and current, and you can report errors: OpenAI's help center invites feedback through the thumbs icons and a menu on each product for reporting incorrect information.

Measuring ChatGPT shopping visibility without fooling yourself

Measurement goes wrong in predictable ways. A team runs a query once from a logged-in account with years of Memory, sees its product, and declares victory; or runs it once, does not see it, and rewrites the catalogue. Neither run is evidence. A useful measurement for ChatGPT shopping has five properties.

  1. A fixed set of real buying questions, written the way shoppers phrase needs and grouped by need, not by product. Our guide to building a prompt set covers how many to use and how to avoid questions that name your own brand.
  2. Clean sessions in the right market. Use logged-out or fresh accounts without Memory, in the U.S., so personalisation does not flatter or punish you.
  3. Repeated runs. Record whether each product appears across several runs and report a rate, not a yes or no.
  4. Separate outcomes. Track carousel presence, position in the merchant list, the price shown, and any label, as four distinct measures. A product can win the carousel and lose the merchant list to a reseller.
  5. Attribution on the click. OpenAI's developer guidance suggests adding attribution parameters such as utm_medium=feed to feed URLs when you need feed-specific click tracking, and Shopify shows orders from AI channels with channel or referrer attribution in the admin.

Context helps set expectations. Shopify reported that in Q1 2026, AI-driven traffic to Shopify stores grew 8 times year over year while orders from AI-powered searches grew nearly 13 times. Adobe has reported that AI-referred retail visitors browse more pages per visit and bounce less, while converting somewhat less often than other traffic, which it reads as AI being used for research. Both point the same way: shoppers arriving from ChatGPT have usually narrowed their choice already, so the product page they land on needs to confirm the facts that brought them there.

If you want this measured continuously rather than by hand, InsightWonder's prompt monitoring runs a fixed question set across ChatGPT and other engines on a schedule and keeps the results comparable over time.

Sources