OpenAI Product Feeds for ChatGPT Shopping: A Retailer Workflow

OpenAI now documents a product-feed path for merchants that want ChatGPT to receive structured catalog data for shopping discovery. It is useful AEO infrastructure for retailers, but it is not a generic content tactic and it is not a promise that a product, merchant or brand will be shown.

The practical objective is narrower: give an approved integration a current, internally consistent catalog with public product pages, accurate prices and availability. Then keep three decisions separate: product discovery in ChatGPT, paid ads, and checkout. OpenAI documents different requirements and controls for each.

What a product feed is for

OpenAI's Product feeds documentation describes structured catalog data for discovery and checkout in ChatGPT. Its getting-started guide says onboarding is currently available to approved partners, so a merchant should not assume that publishing a CSV at a URL alone enrolls it.

For a supported direct file-upload feed, OpenAI documents one row per purchasable item or variant. The required discovery fields cover a stable item identifier, title, factual description, public product URL, brand, seller name, direct image URL, availability and price. The product reference also says product and image URLs must be publicly accessible, and that the item URL should select the relevant variant where possible.

This is why the work belongs with product operations as much as SEO. An answer-writing team cannot correct an old sale price, an out-of-stock variant or a product page that the merchant's own catalog no longer serves. The feed is a representation of the offer, not a substitute for the live offer.

Do not merge discovery, ads and checkout into one claim

OpenAI's product reference explicitly separates discovery data from Ads and checkout requirements. A search-eligible product is not automatically an ad, and a checkout-enabled product requires a separately enabled integration. In the OpenAI feed format, is_ads_eligible is independent of search eligibility, while checkout also depends on search eligibility and an enabled checkout integration.

That separation matters because product ads are paid media. It is the same reporting discipline behind why ChatGPT Sponsored Agents are not AEO: paid delivery, organic product discovery and a user-facing answer are different surfaces with different evidence. Do not combine their impressions, clicks or outcomes into one visibility number.

OpenAI's shopping guidance says product results are selected independently and are not ads. It also says the merchant list can consider availability, price, quality, and whether the merchant is the maker or primary seller. Those stated factors are not a recipe for a guaranteed placement, and OpenAI says the experience may evolve.

A small-team implementation workflow

1. Confirm which integration path actually applies

Start with the approved-partner status and the feed format OpenAI has confirmed for the merchant. The getting-started guide documents file upload and API paths. It generally recommends a full daily file snapshot with API updates throughout the day; a small feed can use the API for both the full catalog and regular changes.

Shopify merchants are a separate case. OpenAI says Shopify Catalog already integrates their product data and that individual merchants need no additional action for that catalog integration. Verify the current state with the merchant's platform and OpenAI rather than sending a duplicate feed by default.

2. Treat a variant as a sellable record

OpenAI's file-upload reference instructs merchants to submit a separate row for each variant, with its own item identifier, selected options, price, availability, URL and images. That is more exacting than a broad category page, but it matches the question a shopper is asking: can I buy this size, color or configuration now, and at what price?

Make the product page, feed record and checkout catalog agree on the same variant. When they disagree, correct the merchant data first. A more persuasive product description cannot make an unavailable item available.

3. Keep the feed current and validate failures at the source

OpenAI says to keep snapshots current for file upload or upsert products through the API. Its reference says Upload History reports missing required columns and invalid supported values, while a malformed row can be rejected without rejecting every valid row.

Use that processing evidence before interpreting a missing product in ChatGPT. Check a small set of high-value records after every price, inventory or merchandising release: the public URL resolves, the selected variant is correct, the image is public, the price is current and availability uses the documented value.

4. Be careful with Google-compatible imports

OpenAI supports a Google-compatible product-data path, but its documentation says to use it only after OpenAI confirms it for the registered feed. The reference also documents different behavior: accepted Google-compatible products have search enabled and checkout disabled, and uploaded fields such as is_eligible_search=false do not opt an item out on that path.

That is a material implementation distinction. An existing Google feed can be a starting asset, not proof that OpenAI-format search controls, checkout rules or Ads fields mean the same thing. Confirm the format selected during onboarding and the controls that apply to that format.

5. Separate prompt observations from feed health

After validating the catalog and first-party ingestion feedback, run a fixed set of real shopping questions that reflect the products and constraints a buyer uses. Teams can preserve that question panel in CiteCue. AEO Landscape and CiteCue have common ownership; see our disclosure record. CiteCue can then record source observations across its supported engines.

CiteCue is not OpenAI feed processing, inventory data, merchant ranking data or checkout reporting. Its observations cannot prove that a product was ingested, eligible, selected, purchased or seen by a shopper. The fit is repeatable prompt sampling after catalog changes, while the feed's validation and the merchant's own analytics remain the evidence for their respective layers.

For the underlying fetch and crawl checks on public product pages, use the technical AEO crawler guide. The prior OpenAI crawler decision about OAI-SearchBot, GPTBot and noindex is a separate control plane, covered in our publisher-control workflow.

What a good result does and does not mean

OpenAI's Shopping Research help article says the experience may use ACP merchant data, public product information and other relevant retail sources. It also says prices, stock and discounts can change frequently and may not always match what is shown, and that a shopper should verify final terms on the retailer's site.

For a merchant, that means a clean feed is a prerequisite for accurate catalog representation, not a conclusion about demand or visibility. A correct product record can still be absent from a particular response. A product surfaced in one sample does not establish a stable ranking. And a paid product campaign is neither proof nor a substitute for organic product discovery.

The credible workflow is therefore unglamorous: maintain the catalog, validate the ingestion path, preserve public product-page access, separate paid from organic reporting, and sample buyer prompts without treating a sample as a guarantee.

Sources and verification

• OpenAI: Product feeds
• OpenAI: Get Started with Agentic Commerce product feeds
• OpenAI: Product file-upload reference
• OpenAI: Shopping with ChatGPT Search
• OpenAI: Using shopping research in ChatGPT

Sources checked 4 October 2026. Product-feed access, schemas, field requirements, format behavior and ChatGPT shopping descriptions are attributed to the OpenAI documentation above. The implementation workflow, evidence separation and AEO interpretation are AEO Landscape editorial guidance. No setting, feed or observation is presented as a guarantee of product display, ranking, citation, traffic, conversions or revenue.