ChatGPT Product Feed: How OpenAI Lists Products

What is a ChatGPT product feed?

A ChatGPT product feed, the merchant feed a seller submits to OpenAI, is a structured file of product data that lets ChatGPT show those products in its shopping results with accurate prices, images and stock status. OpenAI publishes the format as a commerce feed specification (opens in a new tab), filed under agentic commerce, and accepts merchants by application.

The feed is not the only way products reach ChatGPT. OpenAI's own merchant FAQ answers the question directly: a feed is not required, because ChatGPT already crawls product pages through OAI-SearchBot. What the feed buys is control. A crawled page is read when it is read, and a merchant who relies on crawling should first confirm that OAI-SearchBot is not blocked. A feed carries price and availability that the merchant states and refreshes.

ChatGPT Shopping is live in the United States only at the time of writing, which OpenAI states on its merchant page. A merchant outside the United States can prepare and submit a feed to OpenAI, but no shopper will see the result yet. The shopper side, how ChatGPT Shopping chooses and shows products, is defined in the glossary.

A ChatGPT product feed governs products, and nothing else. It does not decide whether ChatGPT cites a merchant's guides, comparisons or category pages, which is a separate discipline covered in how to rank in AI search. A shop that sells through ChatGPT Shopping and is never mentioned in a ChatGPT answer has two problems, and one feed solves one of them.

How can I sell my products on ChatGPT?

Check whether a platform has already done it for you, because for two of them it has. OpenAI's merchant page states that a seller on Shopify or Etsy has a catalogue that is "already integrated, and no additional setup or application is required". A merchant on either platform who fills in the application form is doing work OpenAI does not ask for.

Everyone else applies. OpenAI runs a merchant intake form asking for company details, the merchant website, primary product categories and the unique SKU count of the catalogue, and applicants join a waitlist. Feed delivery is then a push: merchants receive a delivery endpoint and credentials during onboarding, submit a sample of representative products for validation, and push the full catalogue once the sample passes. The sequence is:

1. Confirm your platform has not already integrated you. Shopify and Etsy sellers stop here.
2. Apply through OpenAI's merchant form and wait for approval.
3. Receive delivery credentials as part of onboarding.
4. Build a sample feed covering every required attribute.
5. Submit the sample and fix whatever the validation rejects.
6. Push the full catalogue, then automate the refresh cycle.

The step most merchants underestimate is the third. Validation is per row, and a row that fails its required attributes is dropped rather than queued, so a catalogue can pass its sample and still lose products at full scale if the sample was not representative of the messy parts.

Ecommerce platforms are starting to shortcut this. CNBC reported in March 2026 that Shopify launched an Agentic Plan, letting merchants without a Shopify storefront surface products in ChatGPT and Google Gemini through Shopify's own tooling. Check what a platform or feed vendor already offers before building a pipeline, because the integration may exist.

What file format does a ChatGPT product feed use?

OpenAI prefers Parquet, ideally zstd-compressed, and also supports gzipped JSONL, CSV and TSV. The file upload overview (opens in a new tab) states the delivery model in the same table: merchants push feeds to OpenAI over SFTP, and the SFTP root directory represents the entire catalogue.

That last row is worth checking against whatever guide you read next. Several of the top-ranking guides on this topic list XML among the accepted formats. OpenAI's own documentation lists neither XML nor plain JSON on the native path, and its Google-compatible path accepts UTF-8 tab-delimited .txt or .tsv and comma-delimited .csv, with gzip variants.

Parquet being preferred is the detail that catches teams out, because Parquet is a data-warehouse format, closer to what an analytics team works in than to a shop export. A merchandising team that exports a CSV from a shop admin can still comply, since csv.gz is supported. A large catalogue will get better results from the format OpenAI asks for.

Catalogue size is the reason. OpenAI's file upload overview sets out sharding in the same table as the formats: up to 500,000 items per shard is recommended, with shard files targeted under about 500MB. A 100,000-product catalogue fits in one shard. A million-product catalogue does not, and a merchant at that scale is running a pipeline rather than an export, which is the working assumption behind preferring a columnar format.

There are two delivery paths, not one

OpenAI documents file upload and an API separately, and they are different jobs. File upload pushes a full catalogue snapshot over SFTP. The API path handles updates to individual products between snapshots. A merchant with fast-moving stock will end up using both, and a merchant evaluating the work should price both rather than assuming the feed is a nightly file and nothing else.

Which fields does the ChatGPT product feed require?

The ChatGPT product feed requires nine attributes on every row. OpenAI's specification marks each of the following Required, and a row missing any of them is not eligible for display.

The availability field rejects rows, and that is the trap

OpenAI's specification states that omitted, empty or unrecognised values for availability reject the row, and that a merchant should send unknown explicitly when stock status is unavailable. This is the one required field where the safe-looking action, leaving it blank because the data is missing, is the action that removes the product.

A catalogue with partial stock data therefore needs a deliberate decision before it is exported. Mapping every unclear row to unknown keeps those products eligible. Leaving them empty drops them, and nothing in the shopping results tells you which products went missing or why.

And that safe value is not portable. OpenAI's Google-compatible upload path uses a different vocabulary for the same field. Availability there must be in_stock, out_of_stock, preorder or backorder, and the specification says so in terms: use preorder rather than the OpenAI spelling pre_order, and unknown is not accepted at all. So the mapping that rescues a thin catalogue on the native path has no equivalent on the Google-compatible one.

Is the ChatGPT ads feed the same as the organic shopping feed?

No. OpenAI operates two feed paths with two different eligibility flags, and products in one do not appear in the other. The organic path is gated by is_eligible_search. The advertising path, which powers ads generated from a product catalogue, is gated by a separate is_ads_eligible flag.

OpenAI states the consequence plainly in its help centre article on creating campaigns from product feeds (opens in a new tab): products from that feed are eligible for ads during the beta and "will not appear in organic ChatGPT conversations", though OpenAI says it may offer that capability later.

This is worth stating because the coverage reads the other way. Search Engine Land's report on the ads launch describes the same product data as powering organic answers and advertising, "effectively linking organic presence with paid campaigns". A merchant reading that, and then loading a catalogue into the ads product, would reasonably expect organic placement and would not get it.

is_eligible_search defaults to true, so opting in is not the question

The specification says is_eligible_search is optional, and that an omitted or empty value is treated as true. A merchant who submits an organic feed is therefore in by default and has to set the flag to false to withdraw a product. OpenAI adds a caveat in the same line that is easy to skim past: eligibility does not guarantee display.

Can customers buy inside ChatGPT?

Mostly not, and this is the fact on which the ranking guides have gone stale. OpenAI's merchant FAQ now states that it is "moving away from a standalone Instant Checkout experience in ChatGPT and prioritizing better shopping discovery and merchant-owned checkout experiences". CNBC reported the pivot on 24 March 2026, describing a revamped shopping experience after Instant Checkout failed to take off.

The is_eligible_checkout attribute still exists in the specification, which is why setup guides keep listing it as a step. Read its conditions and the picture changes: OpenAI says the flag opts a product in only when search eligibility is also true and checkout is "enabled for the integration", and setting search eligibility to false overrides it entirely.

For most merchants the practical answer is that shoppers discover products in ChatGPT and complete the purchase on the merchant's own site. Plan the feed around discovery, and treat in-chat checkout as something that arrives through a specific integration rather than through a flag you set.

Why are my products not showing in ChatGPT Shopping?

Because a valid feed buys eligibility rather than placement. OpenAI states the limit inside the specification itself, in the same line that defines the search flag: eligibility does not guarantee display. A merchant whose feed validated cleanly and whose products never appear has not necessarily done anything wrong.

The lever OpenAI names is data quality. Its merchant page frames the benefit of a feed as richer product detail, more accurate information when shoppers compare options and better placement in results. Nothing there promises a ranking mechanism, and no engine publishes one, so treat completeness as the work rather than as a guarantee.

Read the upload report before assuming it is a ranking problem

A feed that uploaded is not a feed whose rows all survived. OpenAI reports the outcome per upload: its specification states that Upload History reports missing required columns and invalid supported values, and that a malformed row can be rejected while valid rows continue processing. A catalogue can therefore upload successfully and still be missing products, with the loss visible only in that report.

The specification is also careful not to promise that validation catches everything. It notes that the stated rules are not a guarantee that every invalid value will be rejected at upload, so a bad value can pass and cause trouble later. Reconcile the row count you pushed against the row count OpenAI accepted, rather than treating a clean upload as proof the catalogue arrived intact.

Optional attributes are where a thin feed loses

The nine required attributes make a product eligible. The optional ones are what let ChatGPT match a described need, and OpenAI documents a long list of them, including material, color, size, gender, dimensions with a unit, and weight with a unit. A shopper asking for a waterproof jacket in a men's large is describing attributes, not a product name.

This is where product discovery in ChatGPT differs from a keyword-matched shopping feed. The shopper writes a sentence, so the structured product data and the description are read together as a description of the thing. A feed carrying only the required nine gives ChatGPT less to match a conversational request against than a competitor's fuller row.

What does a ChatGPT product card actually show?

A ChatGPT Shopping product card names a merchant without linking one, which changes what a merchant's identity in that surface is made of. RankX AI observed 21 product cards across 6 purchase-intent prompts in the 28 days to 12 September 2026, and none of the 21 carried a resolvable merchant domain. ChatGPT named 9 distinct merchants across those 6 prompts, as display strings only.

Those strings arrive as the feed wrote them. The nine included "CAPiTA Snowboards | NA", carrying a pipe and a region code, and "BS The Board Shop". Whatever a merchant puts in seller_name is what a shopper reads, and in this sample it was the only merchant identity the card offered.

The sample is small and is reported as an instance rather than a study: 6 prompts on one monitored test catalogue, not a survey of the surface. Scope is the sharper limit, and it is worth stating plainly. All 6 prompts are snowboarding, so this is one vertical. The Peec study cited below sampled 10 industry categories precisely because category matters.

What survives those limits is the part that concerns what a ChatGPT card contains rather than how often something appears in one. The 21 cards came from 9 real merchants answering real purchase questions, and the absent domain is a property of the card format. Treat seller_name and brand as customer-facing copy, not as plumbing.

Google AI Mode returned no carousel at all in the same window

The same 6 prompts ran on both surfaces over the same 28 days, giving 12 checks in total. ChatGPT Shopping returned a product carousel on 5 of 6 checks. Google AI Mode returned one on 0 of 6 checks, and every one of the 21 cards therefore came from ChatGPT.

That null is about these 6 prompts, and not about the surface. Google AI Mode does return product cards. It returns far fewer of them: Productrise tracked over 2 million product listings across more than 100,000 results pages and AI Mode responses between 9 and 31 August 2026 in the US and UK. It found AI Mode showing an average of 3.9 products against 27.8 in traditional search.

A much sparser shelf is the context this 0 of 6 sits in. What survives is that the two surfaces are not one channel, and that a merchant who watches only the one that answers will conclude more than the data supports.

How does the ChatGPT product feed differ from a Google Shopping feed?

The ChatGPT product feed uses different attribute names from Google's, so a Google Shopping export cannot be repointed without mapping. OpenAI's native schema asks for item_id, url and image_url where Google asks for id, link and image_link, and it adds merchant fields Google does not require, including seller_name on every row.

OpenAI also documents a Google-compatible upload path, which shortens the mapping work for merchants already running a Google Merchant Center feed. Confirm that with OpenAI during onboarding; the required fields differ between the two paths. Perplexity runs its own merchant programme with its own application, so a shop working several assistants is managing more than one relationship.

On the claim that ChatGPT sources its carousels from Google Shopping

The most repeated claim on this topic is that 83% of ChatGPT carousel products come from Google Shopping's top 40 organic results, from a Peec AI dataset of roughly 5,000 carousels and 43,000 products written up on Search Engine Land on 5 March 2026. Precis cites the same vendor for a stronger figure, 100%, attributed to more than 1 million fanout queries and carrying no collection date.

The two are not in conflict. They are one vendor's data at two different strictness settings. The Search Engine Land analysis reports 45.8% of carousel products as exact title matches and over 83% as strong matches at a 0.8 similarity threshold, from the same 43,000 products, so the figure moves by nearly forty points on the matching rule alone. Precis quotes a looser test again, products that could be explained by the top 40.

Read together they say the carousel overlaps heavily with Google Shopping's top 40, and that the exact percentage is a property of the measurement rather than of ChatGPT. The practical reading is that a merchant with a healthy Google Shopping presence is doing work that appears to carry over.

Both measurements predate September 2026, and the question has since become contested. Trade coverage in early September reported ChatGPT shopping results as fed by merchant-submitted feeds, with no Google Shopping data mixed in. The analyst credited with confirming that shift then published a rebuttal, arguing that feeds do not decide which products the model selects in the first place. Treat the overlap as well evidenced up to March 2026 and unsettled now.

How do I check whether my products appear in ChatGPT Shopping?

Checking whether products appear in ChatGPT Shopping means running the same purchase-intent prompts repeatedly and counting, because a single check is one draw from a variable surface. A product absent from one answer may be present in the next. A merchant who runs 20 prompts once has 20 answers and no rate, which is the shape of a number nobody can act on. The same 20 prompts run weekly give a trend.

Build the prompt set from how shoppers describe the problem rather than from product names. The 6 prompts behind the measurement above read like "best beginner snowboard to buy" and "good snowboard for intermediate riders", which is the shape of a shopping question. Nobody types a SKU at an assistant. This is the same discipline that ranking in ChatGPT asks for on the text side, applied to a catalogue.

Record which surface answered, whether a carousel appeared at all, and which merchants took the slots. The carousel-absent case matters as much as the rest: an answer with no products in it is a fact about the surface on that day, and counting it as a loss for the merchant makes the number worse than useless.

RankX AI tracks this as a product. AI Shopping runs a merchant's shopping prompts against ChatGPT Shopping and Google AI Mode on a schedule, matches the cards that come back against a monitored catalogue, and records which merchants took the slots the products did not. The measurements in this article came from it.