ChatGPT product feed for Shopify

The OpenAI product feed is one row per purchasable item or variant, with nine required fields: item_id, title, description, url, brand, seller_name, image_url, availability and price. Titles stay at or under 150 characters and descriptions are plain text up to 5,000: factual copy that answers the questions a shopper would ask, because that is what the model reads.

OpenAI prefers parquet and also accepts jsonl.gz, csv.gz and tsv.gz, all UTF-8 with absolute, publicly reachable URLs. Feedflock renders the tab-delimited form with OpenAI's canonical column names. Unknown optional values stay empty, since the spec rules out placeholders like n/a or unknown, and an empty cell is never read as zero or false.

An item_id is a promise: stable, unique per item or variant, and never reused for a different product. OpenAI matches products by item_id, not by file name, and keeps it through price, stock, title and image changes. Feedflock uses the variant SKU and falls back to the variant ID when a SKU is missing.

Variants of one product go out as separate rows sharing a group_id that differs from every item_id, with listing_has_variations set and a variant_dict naming the selected options, as serialized JSON in the cell, as the spec requires. Each row carries its own price, availability, URL and image. Where the manufacturer assigns a GTIN, send it: exactly 8, 12, 13 or 14 digits with a valid check digit, no spaces or dashes.

ChatGPT ads build on the same product feed. A row opts into Ads processing with is_ads_eligible: true, independently of search eligibility, so you can advertise a product without changing how it surfaces in organic answers, or keep it discoverable without advertising it. In Feedflock the flag is false by default; set it for the whole feed or per product, for example true only on products tagged chatgpt-ads in Shopify.

Checkout inside ChatGPT needs its own integration with OpenAI; an eligibility flag alone does not complete it. The template therefore sends is_eligible_checkout as false, so buyers land on your store. If OpenAI enables checkout for you, it also needs seller_privacy_policy and seller_tos, which Feedflock can derive from your Shopify policy pages.

Delivery is push, not pull: OpenAI provisions an SFTP endpoint during onboarding and your pipeline uploads a complete catalog export: a full snapshot treated as the source of truth, at least once a day, under a stable file name overwritten in place. There is no completion marker: once uploads stop changing, OpenAI processes every file in the SFTP root. Feedflock's part is the file itself, regenerated hourly at a stable URL, so a cron script that downloads and pushes it always ships current prices and stock.

A product missing from a processed snapshot is kept for up to 14 days, so one broken upload never empties your catalog. And to retire an item immediately, set is_eligible_search to false. Money is written as amount, space, ISO 4217 code (79.99 USD), and a sale_price counts only when it is positive, strictly below price and in the same currency. A standard upload targets the US; a currency or product URL on its own doesn't select another market.