ChatGPT Product Feed: Format and How to Submit

The ChatGPT product feed format is the technical specification OpenAI published for merchants who want their products discoverable inside ChatGPT. It defines what fields to send, how to format them, and how to deliver updates so ChatGPT can accurately recommend products when users ask shopping questions.

The spec supports three file formats: JSONL (one product per line as JSON), CSV and Parquet. JSONL is the reference format and most of the spec's examples assume it. CSV works for simpler catalogs. Parquet is for very large catalogs with millions of SKUs.

The current stable spec version is 2026-01-30 (announced by OpenAI on January 30, 2026). That version added support for product extensions, discount fields, and payment handler metadata. The spec lives on GitHub at developers.openai.com/commerce and evolves roughly once per quarter.

Once your feed is submitted, ChatGPT ingests it and the products become eligible for surfacing when users ask relevant shopping questions. Ingestion is typically daily; changes to price, inventory, or availability should be reflected in your feed within 24 hours to stay accurate.

Every product record must include a small set of required fields:

• id: a unique identifier stable across feed refreshes
• title: the product name as a person would say it
• description: a plain-language description, ideally 50 to 300 words
• price and price_currency: current price with an ISO 4217 currency code
• availability: in_stock, out_of_stock, preorder or backorder
• link: the canonical product URL on your site
• image_link: the primary product image (additional_image_link accepts an array)

Beyond the required fields, the spec recommends a set of structured attributes that improve discovery. These include brand, GTIN/MPN, category (Google product taxonomy), size, color, material, gender, age_group, condition, pattern, item_group_id (for variants), shipping, tax, product_highlight, and custom_label_0 through custom_label_4.

For specific verticals, the spec supports category-specific attributes. Apparel feeds should include size_system, size_type, and fit details. Consumer electronics need compatibility fields and spec sheets. Food and beverage items require nutrition_facts and ingredient data.

Merchants who ship only the required minimum often see their products appear as text mentions rather than product cards. A full card, with image, price and availability inline, needs complete attributes, and cards get clicked far more than mentions.

There are three main paths to get products into the ChatGPT feed today:

1. Direct feed submission (via OpenAI). Merchants work directly with OpenAI's commerce team to submit feeds. This path is invite-based for large retailers and is how Walmart, Target, Best Buy, and other launch partners originally onboarded.

2. Platform syndication. If you run on Shopify, commercetools, or Salesforce Commerce Cloud, you can enable ChatGPT syndication through your platform. The platform generates a compliant feed from your catalog and maintains it. Shopify enabled this broadly in Q1 2026. Salesforce launched its pilot in April 2026 with Crocs and Pacsun as early partners.

3. A handled service. Paz prepares the catalog and the feed for ecommerce teams. A Paz expert completes the attributes and descriptions that decide whether a product appears as a card or a mention, our experts check them against agreed guidelines, and the feed ships from data that is already in shape. This path suits brands that want the catalog reviewed before the feed goes out. See how the product feeds service works.

Whichever path you pick, the feed has to refresh on a regular cadence. Stale pricing or inventory gets products de-prioritized or suppressed. Most merchants target a daily refresh. Fast-moving SKUs may need hourly updates through the API.

Most feed quality problems fall into a handful of patterns:

Shallow titles. "Black Shirt" loses to "Men's Merino Wool Short-Sleeve Henley, Slim Fit, Charcoal Black". The assistant matches specific question language to specific product language, so a descriptive title matches more questions.

Missing Google product category. The Google product taxonomy (e.g. "Apparel & Accessories > Clothing > Shirts & Tops") is how ChatGPT maps free-text queries to product sets. Products without a category get fewer matches.

Vague descriptions. Marketing copy that tells a story but leaves out material, fit, care and use gives the assistant nothing to match on. A good description carries both the story and the spec sheet.

Incomplete variant grouping. The item_group_id field tells ChatGPT which products are variants of the same parent. Without it, your T-shirt in 5 colors might appear as 5 separate products competing against each other in rankings.

Stale pricing or inventory. Price mismatches between the feed and the merchant site are a hard de-ranking signal. Same for products shown as in-stock in the feed that 404 or show out-of-stock on the merchant site.

Missing shipping data. Shipping cost and delivery time are increasingly shown inline with recommendations. Feeds without them lose to feeds that have them.