ChatGPT Product Feeds: Separate Discovery, Ads, and Checkout

OpenAI now documents two different product-feed systems: a commerce product file for ChatGPT product discovery and checkout eligibility, and an Ads Manager feed for paid campaigns. They can share catalog fields, but they don’t produce the same outcome. A merchant should track discovery, merchant presentation, checkout, ad delivery, and referral as separate states.

The most important operational detail is deletion. Omitting a product from a later commerce snapshot does not guarantee immediate removal. OpenAI says an omitted product can remain for up to 14 days, while explicitly setting is_eligible_search to false disables search and checkout eligibility after the next processed snapshot.

Start with four product states, not one feed status

This state model prevents a common reporting error. A valid row is evidence that the system accepted a record. It is not evidence of product selection, merchant ranking, ad delivery, a checkout session, or a sale.

The commerce file is a full snapshot.

OpenAI’s current commerce specification describes a daily SFTP delivery. Each delivery is a full snapshot of the products that the merchant intends to keep current. Parquet is the preferred format. JSONL, CSV, and TSV are also supported, including gzip-compressed files.

At minimum, a usable product record needs a stable item identifier and enough data to explain what the item is, where it is sold, what it costs, and whether it is available. The current required-field set includes item_idtitle, description, product URL, brand, seller name, image URL, availability, and price. Optional fields can improve matching and presentation, but only when they remain accurate.

Deletion is a state transition.

A full snapshot invites a dangerous assumption: if a row is missing tomorrow, the product disappears tomorrow. OpenAI’s specification does not promise that. Omitted products can persist for up to 14 days. A merchant that must stop exposure should send an explicit eligibility change.

The boolean defaults to true when omitted. It is therefore safer to generate it from an explicit merchandising rule than to rely on the default. OpenAI also says that setting it to false disables both search and checkout eligibility. It is not a discovery-only switch.

Google-compatible input trades detail for portability

OpenAI accepts a Google-compatible product feed, which can reduce the first integration step for merchants that already maintain one. The compatibility path has important limitations. The current documentation enables product search but not checkout, and it does not support the per-item search opt-out available in OpenAI’s native schema.

The compatible file is useful for discovery pilots. It is not a drop-in substitute when checkout or granular eligibility control is a requirement.

Ads Manager remains a separate pipeline.

The commerce file does not replace the Ads Manager product feed, and the advertising feed does not place products into organic conversations. A paid product campaign has its own upload history, filtering, review, budget, bid, delivery, and reporting states. OpenAI’s current Ads Manager guidance says the Products tab is a reporting view: products appear after an active campaign generates delivery data and that reporting becomes available.

1. Confirm that the advertising file processed without row errors.
2. Check the product count after applying ad-group filters.
3. Confirm campaign dates, status, review, budget, and bid.
4. Record first delivery separately from product-level reporting.
5. Allow for the documented reporting delay before diagnosing an empty Products tab.
6. Keep organic shopping observations outside the paid-delivery chain.

Crawler checks also differ. OAI-AdsBotSupports advertising landing-page validation, while OAI-SearchBot supports search discovery. A successful request by one does not prove that the other can reach the page or that any downstream surface selected the product.

A daily reconciliation catches the expensive mistakes

Join the source catalog, exported snapshot, processed result, and observed ChatGPT surface on the stable item ID. For every run, preserve counts for source products, exported records, accepted rows, rejected rows, explicitly disabled items, omitted items, discovery observations, merchant links, checkout events, ad impressions, referrals, and sales.

• Completeness check: every source product intended for the channel appears in one shard.
• Freshness check: price and availability match the landing page at processing time.
• Deletion check: emergency removals use explicit false before omission.
• Surface check: discovery, checkout, ads, and referral are separate columns.
• Exception check: rejected rows return to the source catalog owner as well as the export script.

For a broader model of access, retrieval, citation, and outcome evidence, use the OpenAI publisher-controls guide. Merchants connecting more than one commerce surface should also keep the agentic commerce data-trust map beside the feed specification.

Current limits

OpenAI commerce and advertising access varies by merchant, market, account, and rollout. Eligibility does not guarantee selection, placement, checkout, delivery, or referral. SearchEngineAnswer did not upload a merchant snapshot for this revision and reports no processing time, serving rate, or revenue measurement. Reopen the current specification before changing a production export.

Primary documentation: Commerce file upload overview, product schema, and Ads Manager product-feed campaigns.