OpenAI launched product feeds in Business Hub on November 10, 2026. A store owner claims a business profile, connects a product feed, and ChatGPT can show those products in its answers. I build tools for WooCommerce store owners, so I read the feed spec that week and started testing real catalogs against it.
I built the checker before I built anything to sell. The checker reads a store's public product data, maps each product to the spec, and reports every row that breaks a rule. Then I pointed it at stores I do not run.
The scan results
First batch: 459 products across four large WooCommerce stores. Zero passed. I kept scanning other stores, and the total now stands at 1,541 products with 32 passing every rule.
One failure leads the list by a wide margin: brand. WooCommerce has no brand field in its default setup. A store owner can run a clean store for years and still fail the spec on the first row, because the field the spec demands has no home in the platform. seller_name has the same problem. It describes the store, and the spec wants it repeated on every product row.
The 9 required fields
Every row in the feed needs all nine fields, in the formats the spec sets:
Field
The spec wants
In WooCommerce
item_id
A stable ID for the product
Product ID works
title
The product name
Present
description
A real product description
Present, and empty on many stores
url
A public link to the product page
Present
brand
The product's brand
No default field. Top failure in my scans
seller_name
The store's name, on every row
Store-level data, not a product field
image_url
A public link to the product image
Featured image works
availability
One of five fixed values
Needs translation from stock status
price
Amount plus currency code, e.g. 79.99 USD
Needs reformatting, and zero fails
Two of the nine fields start missing on a stock WooCommerce install. The other seven sit in the catalog in shapes the spec rejects.
The format traps
Price carries the tightest format rule. The spec wants a decimal amount, a space, and a currency code: 79.99 USD. A price string with a currency symbol or a thousands separator breaks the rule. Rows priced at zero failed in every scan I ran. WooCommerce's own store had 21 of them.
Availability takes one of five values: in_stock, out_of_stock, pre_order, backorder, unknown. WooCommerce tracks stock in its own terms, so a mapper has to translate each status into the enum. A feed that passes the raw store value through fails the check.
Two conditional rules catch feed builders out. A sale_price must sit below price, in the same currency. Variants of one product share a group_id, so the feed presents them as one product with options instead of a pile of near-duplicate rows.
The freshness rule
The spec asks for a full snapshot at least once a day. OpenAI retains the last processed record for up to 14 days. Past that window, the product ages out of the results.
Failure carries no signal. OpenAI drops a row that breaks a rule and sends the store no notification. The owner finds out from missing traffic weeks later, or from a customer asking why a product vanished from ChatGPT. I heard versions of that story before I understood the cause, and it pushed the product design more than any spec paragraph did.
Inside the checker
The free check reads the first 300 products from a store's public WooCommerce Store API. A mapper converts each product into a feed row: it builds the price string, translates stock status into the availability enum, and attaches the store name to every row. A validator then runs the rules in order: presence, format, enum membership, then the conditional rules.
The validator groups its output by rule. Grouping matters more than I expected. An owner facing 300 broken rows needs one line per problem, along the lines of "129 rows miss brand, 9 miss a description," and a fix for each line. A raw dump of 300 errors sends that owner back to whatever they were doing before.
From checker to guard
A feed generator writes the file once, and the spec punishes that approach twice. Rows break as prices change and variants sell out, and the snapshot ages out after 14 days even if every row stays valid. The gap sits after generation: validation and watching.
Feed Guardian, the paid tool behind the checker, runs a daily cycle. It pulls the store's live products, rebuilds the feed, validates every row, and compares the result against the previous day. The feed lives at one hosted URL. The owner pastes that URL into Business Hub once and leaves it there. A broken row or a drifted price triggers an email that names the product and the rule it broke. Activation costs $49, one time.
Lessons from the build
Scan real catalogs before designing the product. I expected format errors to lead the failure list. Missing fields led instead, and that finding changed the product: the gap-filling step, where the owner sets brand and seller name once, matters more than the export step.
Silent failure is a product category. OpenAI gives the store owner no error surface, so monitoring carries the value. A prettier exporter would have solved the wrong problem.
Platform defaults decide compliance. Stores failed on fields their platform never asked them to fill. A validator aimed at one platform has to encode that platform's defaults, field by field.
Group errors by rule. Counts by rule turn into a fix list. Raw error lists turn into abandoned browser tabs.
Run your store
The free check lives at usefeedguard.com/pre-check. Paste a store URL and it returns the pass count and the failing rows, with no signup. If you run a WooCommerce store, post your score in the comments. And if a rule in the validator misreads the spec, tell me. Corrections make the checker better, and I would rather fix it than defend it.
Top comments (0)