ChatGPT Product Feeds: Current Fields and Delivery Paths

CH.01

Choose the product-data path before building a feed

A direct feed is one way to supply product information, not the only way products can be discovered. First check your commerce platform, product eligibility, target market, and available integration. Discovery, paid advertising, and checkout have separate requirements.

OpenAI’s March 24, 2026 announcement focused on product discovery and merchant checkout choices. It did not establish that every Instant Checkout experience had ended. Current shopping help still describes Instant Checkout for some eligible products and merchants. Shopify’s ChatGPT channel sends buyers to the merchant’s online-store checkout.

Eligible Shopify products can be shared through Shopify Catalog without an individual direct-feed integration. Shopify’s current ChatGPT requirements include selling to US customers, Catalog eligibility, required policies, and acceptance of its terms; a merchant can be based elsewhere. B2B-only products are not supported on that storefront path. Check your own eligibility and controls rather than assuming every store is included.

For a direct integration, OpenAI’s current onboarding is for approved partners. Confirm the supported delivery method and schema before building. No verified daily shopping-query count or guaranteed visibility gain is needed to make this decision. Public-web and other listing routes can still expose products without a direct feed.

CH.02

Use the schema for your confirmed integration

The current supported file-upload reference has a Stable schema. Its Draft view is for planning and feedback, not production. The table below covers the OpenAI-format discovery fields, not every API or platform schema.

The file-upload guide currently supports parquet, jsonl.gz, csv.gz, and tsv.gz via SFTP. Do not assume any CSV/TSV/XML/JSON export is interchangeable. Google-compatible imports have their own mappings and validation; API delivery is another path. Confirm your onboarding contract before adapting an existing feed.

The Stable OpenAI format requires nine fields below. Google-compatible files instead require eight source columns and use the registered merchant name; identifiers and variant fields can be conditional. Optional unknown values should be omitted, not invented. Eligibility does not guarantee display, and a malformed row does not necessarily reject the whole feed.

CH.03

Keep data current through the agreed delivery path

Accurate price and availability help shoppers make decisions. Publication, upload, processing, and the displayed answer are different stages; successful transmission does not mean an instant change in every response.

If the product page and feed disagree, a shopper can encounter the wrong price or stock status. Investigate the source data and processing record. This is a real operational problem, not evidence that the model learned distrust or applied a specific ranking penalty.

OpenAI’s file-upload guide recommends full catalog snapshots at least daily. Get Started separately recommends API updates during the day, or API-only delivery for a small feed. It does not establish a universal 15-minute refresh or display deadline. Match updates to the supported integration and the volatility of your data.

CH.04

Describe the product accurately, without selection promises

Complete, factual data can help represent products accurately. It does not provide a measured conversion from attribute completeness to recommendation probability.

Supply the facts a shopper needs for this item and variant. Do not claim to know an internal confidence threshold or that every incomplete product is automatically skipped. Selection also depends on the user’s request and product context.

Use real size, color, material, and other relevant attributes supported by your chosen schema. Omit unknown optional data. Check conditional variant and identifier requirements instead of adding every possible field.

CH.05

Investigate feed errors from observable evidence

Check row validation, processing, source-data accuracy and shopper-facing information. Do not infer silent exclusion or a hidden ranking penalty from a missing recommendation alone.

Use your validation and processing records to diagnose data issues. A valid record may still not be displayed; other rows may continue processing when some rows are rejected. Recommendation absence alone does not identify the cause.

• →Price mismatch: reconcile the same item, currency and promotion across feed and site; do not label it a learned trust penalty.
• →Stock lag: update the actual availability value when stock changes; do not claim instant processing.
• →item_id reuse: keep identifiers stable and never assign one to a different item; do not invent a historical reputation reset.
• →Thin description: add relevant supported facts, without promising that longer text wins a citation.
• →Image issue: check public access and the correct item/variant, not an unsupported universal ban on lifestyle photography.
• →Currency or units missing: follow the chosen schema’s money and measurement conventions.
• →Refresh failure: investigate delivery, processing and data age; alert according to impact.

CH.06

A route-specific launch and operating checklist

Validate the integration you actually use. This checklist is Tenten’s practical worklist, not a timed guarantee or a claim about every client deployment.

Check eligibility and the confirmed schema, then validate data, delivery and processing. Preserve the distinction between successful ingestion and a product appearing in an answer. Use current provider instructions; do not turn a draft schema or a third-party article into a production contract.