ChatGPT Product Feed Specification Mapped to Your Google Feed

OpenAI's product feed reference asks for nine fields on every row. Seven are columns a Google Merchant Center feed already requires under the same or a similar name. The eighth is brand, which Google requires for most new products, and for the ninth, seller_name, we found no matching column in Google's specification. A merchant with a clean Google feed therefore holds most of a ChatGPT feed already. The map below reflects both specifications as we read them on October 7, 2026, when OpenAI's getting-started page still opened with the notice, "Onboarding product feeds in ChatGPT is currently available to approved partners."

What does OpenAI require on every row?

The ChatGPT shopping product feed specification requirements begin with nine fields that must hold a value on every row, listed under basic product data in the feed reference from OpenAI. The table sets each one beside its attribute in the product data specification from Google.

Check the availability column before any export. OpenAI's reference says omitted, empty or unrecognized availability values reject the row, and it cautions elsewhere, "Format guidance is not a guarantee that every invalid value will be rejected at upload."

Which optional fields carry over, and which need new data?

Everything past those nine is optional or conditional, and it sorts into three piles.

Same name, same meaning: gtin, mpn, condition, color, size, material, gender, age_group and sale_price. OpenAI specifies a GTIN of exactly 8, 12, 13 or 14 digits with a valid check digit.

Renames: item_group_id becomes group_id (the old name stays an accepted alias), product_type becomes product_category, and additional_image_link becomes additional_image_urls.

New data: every variant row needs listing_has_variations set to true and a variant_dict that maps option names to the selected values. Returns get three fields (accepts_returns, return_deadline_in_days and return_policy), where Google steers most catalogs toward account-level return policies. review_count and star_rating have to come from your reviews platform, because we found no review count or star rating attribute in Google's product data specification. shipping_price, where your integration supports it, is a single charge in the price currency, and OpenAI reads an omitted value as unknown rather than free.

A hypothetical example shows the size of the job: a footwear merchant's Google row for a black trail shoe in size 10 carries id, link, image_link, item_group_id, color, size and an availability of preorder. For an OpenAI-format file that row takes four renames to OpenAI's current names (item_id, url, image_url, group_id), one changed value (pre_order) and three additions (seller_name, listing_has_variations and a variant_dict holding color and size). Title, description, brand, price and GTIN go through untouched.

Where is the schema published, and how is the file delivered?

The ChatGPT shopping product feed requirements schema is published in two views, Stable and Draft, and only Stable is supported for production uploads. OpenAI says Draft "is published for planning and feedback only."

Delivery is a push. The file upload guide from OpenAI describes a full snapshot feed sent over SFTP and treated as the source of truth, with a recommended cadence of at least daily. Parquet is the preferred format, and jsonl.gz, csv.gz and tsv.gz are also supported. Keep one stable file name and overwrite it on each run. For large catalogs the guide recommends up to 500k items per shard and files under about 500MB. A product missing from a snapshot is retained for up to 14 days, so omission is a slow way to pull a product.

For changes during the day, the getting-started guide from OpenAI generally recommends the entire feed once a day by file upload with updates sent through the API, and it notes that promotions data can only be provided via the API. The Merchant Center habit of syncing with the site carries over, since Google tells merchants to submit product data at the same time they update their website and says it usually updates existing products within minutes.

How does the feed control what appears in ChatGPT search?

In the ChatGPT search product feed specification, search eligibility is a row-level flag that defaults to on. Leave is_eligible_search out and the product is eligible, or set it to false and both search and checkout are disabled for that row the next time a snapshot is processed. The reference adds, "Eligibility does not guarantee display." Checkout defaults the other way. A row opts in only when is_eligible_checkout is true, search is also true and checkout is enabled for the integration, a separate step that also calls for seller privacy policy and terms of service URLs. Ads have their own flag, is_ads_eligible, and their own guide.

Since eligibility and display differ, a validated feed still has to be checked inside ChatGPT, and our notes on where prompt-based tracking breaks explain why a handful of test prompts can mislead. The wider metric is AI share of voice, and the AI Visibility Checker is a further resource for a first read.

Does a product listing need structured data on the page too?

For ChatGPT shopping feed product listing requirements, structured data means the feed file itself: we found no Schema.org or JSON-LD requirement in OpenAI's feed reference or file upload guide. The reference does lean on the page, though: product and image URLs must be publicly accessible, brand should read as shown on the product page, and variants should be grouped as presented on your site.

Google's specification ties feed and page together more tightly. It marks which attributes have a Schema.org property, and it lets Merchant Center update product information automatically from structured data on your site once you turn that feature on. How an assistant describes or cites a page is a separate question from the feed, which we cover in how to get cited by ChatGPT. The brand and seller names in the feed should match the page either way, for the reasons in our piece on entity consistency.

What must a merchant have in place before the first upload?

The ChatGPT shopping feed product requirements a merchant meets before any row is read concern access, policy and market. Onboarding is limited to approved partners, who apply through a form linked from the getting-started page. That page also carries a prohibited products policy that covers adult content, weapons, prescription-only medications and age-restricted products such as alcohol and nicotine, among other categories. Country columns on a row do not widen the market until OpenAI confirms the integration, because the reference states, "The standard OpenAI-format upload currently targets the US." OpenAI suggests validating in phases, starting with a sample of around 100 items.

What does Google's feed ask for that ChatGPT's does not?

A merchant catalog built to ChatGPT shopping product feed requirements is leaner than a Google AI Mode product feed, because Google makes conditional several things OpenAI leaves optional. Google requires an MPN when a product has no manufacturer-assigned GTIN, shipping costs for a list of countries that includes the US, age_group, gender, color and size for apparel aimed at six named countries, and condition for used or refurbished goods. The OpenAI format marks the matching fields optional, though it says to always specify used or refurbished condition.

We found no separate AI Mode feed in Google's documentation. The help page on conversational attributes from Google describes six optional additions to the Merchant Center specification, among them question_and_answer, related_product and popularity_rank, meant to help shoppers find product information on AI-driven surfaces such as AI Mode in Search. OpenAI draws the line elsewhere: "Raw review entries and question-and-answer lists are not part of this discovery contract." We also found no counterpart in OpenAI's Stable reference for Google's related_product, popularity_rank, adult, unit_pricing_measure, energy_efficiency_class or loyalty_program attributes.

There is a shortcut. OpenAI accepts a Google-compatible file "only after OpenAI confirms it for your registered feed": a tab-delimited .txt or .tsv or a comma-delimited .csv, gzip allowed, with XML, RSS and Atom unsupported. Eight Google columns are required on that path, and Google's identifier rule returns with them, so each row needs a valid GTIN or MPN unless identifier_exists is no, and preorder or backorder rows need an availability_date. The registered merchant display name becomes the seller on every row, accepted products have search enabled and checkout disabled, and a false value in is_eligible_search does not opt a product out.

The Google-compatible path is the quicker start for a discovery-only catalog, and the OpenAI format is the one to build when you need per-item search control, your own seller name on each offer, returns and review fields, or checkout. For help with the mapping and the pages behind it, our AI search optimization service is the place to go next. Whichever file you send, get the required fields right first, because nothing optional helps a row that was rejected.

Sources

• OpenAI, Products feed reference (Stable view)
• OpenAI, File upload overview
• OpenAI, Get started
• Google, Product data specification
• Google, How to use conversational attributes
• Google, Provide high-quality data