On this page
- What the ChatGPT product feed specification is (and how to read it)
- Required fields: the 11 columns every row needs
- Variant fields: itemid, groupid, offerid and variantdict
- Price, sale price and availability: formats that get rows rejected
- Optional fields worth filling: gtin, mpn, condition, images, category, shipping and returns
- OpenAI flags and checkout fields: search and checkout eligibility
- Common feed errors and a pre-upload checklist
- Where Revnu fits after the feed is live
- Fix the 11 required columns first, then track what the feed earns
- Related Business Hub guides
The ChatGPT product feed specification asks for one row per purchasable item or variant, requiring 11 fields in total: nine basic product fields (including availability) plus target_countries and store_country. Everything else is optional or conditional. This guide walks through each field in OpenAI's official schema: what it means, whether you must send it, the exact format, an example value and the mistake that most often gets a row rejected.
Field names are as of 18 November 2026, per OpenAI's developer docs. The feed is in beta, so recheck the schema before you upload. If you are still deciding whether the feed is worth building, start with ChatGPT Business Hub: What It Is and Who Can Use It. For the upload workflow itself, see ChatGPT Business Hub Product Feeds: CSV Setup Guide.
What the ChatGPT product feed specification is (and how to read it)
The specification is OpenAI's product feed schema, a reference table of attribute names, data types, requirement levels, descriptions and examples (OpenAI product feed schema). It sits alongside the Submit product feeds guide and the Product feeds overview, which also links to best practices. The schema says what to submit. The submit guide says how to deliver it.
Read the requirement column in three tiers. Required means every row needs it. Conditional means it becomes required when a situation applies, such as variants. Optional means omit it when you don't know the value. Don't fill optional columns with guesses.
The core rule is granularity: one row per purchasable item or variant. A blue size-10 shoe and a black size-10 shoe are two rows. Money is written as a decimal amount in major units, a space, then an uppercase three-letter ISO 4217 currency code, for example 79.99 USD. Keep that pattern in mind, because it recurs in price, sale price and shipping.
The schema's sections cover basic product data, OpenAI flags, variants, item information, media, price and promotions, availability, merchant info, fulfilment, returns, reviews, geo tagging, ads and checkout. This guide takes them in the order a feed manager actually needs them.
How does the Agentic Commerce Protocol relate to the product feed?
OpenAI describes the Agentic Commerce Protocol as open infrastructure for AI-native commerce in ChatGPT, where the product feed supports product discovery, though the feed itself does not complete purchases. The feed is a structured file of your catalogue, one row per product or variant, that ChatGPT reads to know what you sell. Checkout is handled separately through the protocol's checkout fields. If you only want products to appear in shopping results, a clean feed is the part you have to get right.
Required fields: the 11 columns every row needs
Every row needs these 11 columns. Miss any one and the row is unusable for discovery.
| Field | What it is | Format and example | Usual mistake |
|---|---|---|---|
item_id |
Stable, unique ID for the item or variant | String, e.g. MUG-350-BLUE |
Reusing an ID, or changing IDs between uploads |
title |
Product name | Concise plain text, about 150 characters or fewer, include the selected variant where relevant | Keyword stuffing or generic parent titles on variant rows |
description |
Factual description of this specific item | Plain text, 5,000 characters or fewer | Pasting HTML or marketing copy |
url |
Product detail page | Full URL to the page for this item | Pointing at a category or homepage |
brand |
Product brand | String, e.g. Trailhead |
Putting the store name here instead |
seller_name |
Who sells it | String | Leaving it blank on single-brand stores |
image_url |
Main product image | Publicly accessible direct image URL, HTTPS preferred | Linking to a page that contains the image, or a login-gated asset |
availability |
Current stock status | in_stock, out_of_stock, pre_order, backorder or unknown |
Free text like In Stock or implying stock when unsure |
price |
Regular price | 79.99 USD |
Missing the currency code, or using a comma decimal |
target_countries |
Where the item is offered | Array of uppercase two-letter codes, e.g. ["US"] |
Lowercase codes or full country names |
store_country |
Country of the store | Uppercase two-letter code, e.g. US |
Confusing it with where shoppers live |
Two cautions. First, standard uploads currently target the US regardless of what you put in the country fields, so confirm multi-country support before relying on other markets. Second, target_countries is an array. How arrays are written depends on your file format: a JSON list in JSONL, a delimited string in a CSV cell. Check the schema's value conventions for your format rather than copying a pattern from another feed.
Use unknown for availability when you genuinely don't know. It beats a false in_stock, and you should update the field whenever inventory changes.
Which attributes matter most: title, description, gtin, availability and price?
Title, description, price and availability belong in the required columns on every row. Gtin sits in the optional set, but fill it wherever a barcode exists. Write the title as a plain product name with the attributes a shopper would type, such as brand, model, size and colour. Keep the description factual and free of promotional claims. Use the exact price format and availability values the spec lists, since mismatches are a common reason rows get rejected. Add gtin so ChatGPT can match your listing to the right product.
Variant fields: item_id, group_id, offer_id and variant_dict
Variants are where most feeds break. The schema's rule: use item_id for the specific item, group_id for its parent listing, and offer_id for a seller's offer.
For a product with variations, create a separate row per variant. Each row has its own item_id, shares one group_id, sets listing_has_variations to true and describes the selected options in variant_dict.
group_id: string identifying the parent listing. It must differ from each variant'sitem_id. Non-variant products can omit it.listing_has_variations: boolean flag marking the row as part of a variant set.variant_dict: a JSON object mapping option names to selected values, such as{"color":"Black","size":"10"}. For variants, use it together withlisting_has_variations=trueand agroup_idthat differs fromitem_id, or it is ignored.offer_id: optional, identifies the seller's offer for the item.
The usual mistakes: setting group_id equal to item_id, which makes the row look like a standalone product; writing variant_dict as a loose string rather than a JSON object; and contradicting yourself, for example a title that says "Black" and a variant_dict that says "Navy". Keep top-level attributes such as color and size consistent with the same options inside variant_dict.
Older Custom variant 1 category and Custom variant 1 option pairs (and the 2 and 3 pairs) remain accepted, along with lowercase aliases, but new feeds should use variant_dict.
Price, sale price and availability: formats that get rows rejected
Price rows get rejected or ignored for format reasons far more often than for business reasons. The format is always amount CURRENCY: 79.99 USD, never $79.99, 79,99 or 79.99 alone.
price is required and holds the regular price.
sale_price is optional, same format, for example 59.99 USD. It must be positive, lower than price and in the same currency, or it isn't used. If you omit it, the regular price applies. Don't leave a sale price in the feed after the promotion ends.
availability takes exactly one of five values: in_stock, out_of_stock, pre_order, backorder, unknown. Anything else is a format error. This is the field most likely to drift, since stock changes daily while product copy doesn't. If your feed is a weekly export, availability is stale by Wednesday. Match your refresh cadence to how fast your inventory moves.
Run a quick price audit before upload. Filter for any price cell without a space and a three-letter code, any sale price greater than or equal to its price, and any row where the currency differs between price and sale price.
Optional fields worth filling: gtin, mpn, condition, images, category, shipping and returns
Optional doesn't mean worthless. These are the ones worth the effort, with the schema's formats.
gtin: one valid GTIN of 8, 12, 13 or 14 digits. Preserve leading zeros, which spreadsheets love to strip. One GTIN per row, not a list.mpn: manufacturer part number. Provide it together withbrand.condition:new,refurbishedorused.additional_image_urls: extra images. An array of URLs in JSONL, or comma-separated URLs in a CSV cell. Show the specific variant the row represents.product_category: a category path from broad to specific, separated by>, for exampleApparel > Shoes > Running.shipping: optional, and it needs feed-specific setup. A tuple stringcountry:region:service_class:price, for exampleUS::Standard:5.00 USD. Keep the empty region slot if you have no region. Don't send both this andshipping_pricefor the same charge.accepts_returns: booleantrueorfalse. Omit if unspecified.return_deadline_in_days: positive integer, supplied only whenaccepts_returnsistrue.return_policy: a public HTTP(S) URL. It links the policy but doesn't say whether returns are accepted. That's the job ofaccepts_returns, so send both.
The schema also has sections for reviews and Q&A, merchant info, geo tagging and ads. Some of these need an integration configured to carry them. If you're unsure, leave them out and add them once the basics are accepted.
The pattern across all of these: send a clean value or send nothing.
OpenAI flags and checkout fields: search and checkout eligibility
Two boolean flags control where a row can appear.
is_eligible_search:trueorfalse. It defaults totruewhen omitted or empty. Setting it tofalsealso disables checkout eligibility.is_eligible_checkout: settrueto opt a row in to checkout, but only if search eligibility is alsotrueand checkout is enabled for your integration. Omitted, empty orfalsemeans disabled.
Checkout isn't a column you can switch on by itself. The schema states that checkout requires a separately enabled integration, so setting is_eligible_checkout to true in a feed without that integration does nothing useful. Discovery is the default path. Checkout is an extra layer on top.
In delimited files, lowercase true and false are accepted as text. The usual mistake is setting is_eligible_search to false on a row you still want purchasable in checkout. Because false there turns checkout off as well, the two flags aren't independent.
Common feed errors and a pre-upload checklist
Most rejected rows fall into a handful of patterns. Run this checklist before every upload.
One row per purchasable item or variant. No parent rows, no duplicated
item_idvalues.All 11 required fields present on every row, including
target_countriesandstore_country.IDs stable.
item_iddoesn't change between uploads.Variants wired correctly. Shared
group_id, different fromitem_id;listing_has_variations=true;variant_dictis a JSON object.Money formatted as
amount CURRENCYeverywhere, withsale_pricebelowpricein the same currency.Availability is one of the five allowed values and is current.
GTINs keep leading zeros and are 8, 12, 13 or 14 digits.
Arrays serialised for your format. JSON lists in JSONL, delimited strings in CSV. Check this against the schema rather than a template from elsewhere.
Images public. No login walls, direct image URLs, HTTPS where possible.
Unknown optional values omitted, not guessed.
Once the file passes, follow the steps in Submit product feeds to deliver it, and read the best practices linked from the Product feeds overview. If you haven't yet claimed your business, do that first with How to Claim and Verify Your Business in ChatGPT.
What are the best practices for a ChatGPT product feed?
Keep the feed accurate, current and consistent with your product pages. Fill every required column on every row, then add gtin, mpn, condition, images, category, shipping and returns wherever you have the data. Use unique ids that never change, group variants correctly, and refresh price and availability often so shoppers don't land on stale offers. Write plain, specific titles and descriptions without keyword stuffing. Validate the file against the spec before each upload, and recheck it whenever the spec changes, because requirements can be updated.
Where Revnu fits after the feed is live
A feed puts your products within ChatGPT's reach. It doesn't tell you whether buyers find you or what to do next. Revnu is an AI agent that finds customers and grows revenue across channels. It tracks appearances in Google, ChatGPT, Claude, and Perplexity (with Gemini mentioned in a specific use-case description), publishes search pages, and runs outreach and ads subject to approval or budget controls, reporting back via messaging options including Slack or iMessage.
Where it helps is the loop around the feed: seeing where competitors are named in answers instead of you, writing the pages that fill those gaps, and measuring the outcome. For how to read Hub reporting alongside your own analytics, see ChatGPT Business Hub Metrics and Weekly Reporting. Pro is $99 a month, self-serve, starting from your website.
Fix the 11 required columns first, then track what the feed earns
A clean feed is 11 correct columns on every row, variants wired with a distinct group_id, and money in amount CURRENCY form. Get those right before touching optional fields, because that's where rows fail. Then measure whether it changes how you show up in AI answers. Start with your website at revnu.com and let Revnu track your visibility and run the follow-up work.
Related Business Hub guides
Written by
Co-founder and CEO, Revnu
