ChatGPT shopping turns product feeds into a new channel

What happened

On 29 September 2025 OpenAI announced Instant Checkout in ChatGPT, together with the Agentic Commerce Protocol, an open standard built with Stripe and a set of merchant partners. The protocol was open sourced and documented for merchants and developers. The first version let US users buy from US Etsy sellers without leaving the chat, with Shopify merchants named as coming next, and single item purchases only to begin with.

The commercial mechanics are worth noting. Merchants pay a fee on completed purchases, prices do not change for the buyer, and OpenAI is not the merchant of record. The merchant's own payment provider handles the transaction. OpenAI has also stated that items available through Instant Checkout get no ranking preference in organic product results.

Why the feed is the interesting part

Behind the checkout sits something more consequential for most retailers: a product feed. OpenAI's documentation describes merchants sending structured product data as a secure feed that is refreshed regularly, starting with a sample feed for validation and then daily snapshots. That feed is what determines whether a product can be surfaced in ChatGPT at all. Checkout and advertising are optional layers on top of it.

The required fields for discovery are not exotic. Each row needs a stable item_id, a title, a description, a product URL, a brand, a seller name, an image URL, an availability value from a fixed list, and a price with a currency code. Variants are handled with group_id, listing_has_variations and variant_dict. Separate flags mark whether a row is eligible for ads or for checkout. There is a Google compatible path that accepts tab or comma delimited files, which means many retailers can start from a feed they already produce.

Why this is not simply another channel

A conventional shopping channel shows a grid and lets the shopper filter it. An assistant does something different. It reads the product data, reasons over it, and answers a question in sentences. If a description does not say what a product is made of, the assistant cannot say it either. If an attribute is blank, the product quietly fails to come up when somebody asks for that property. There is no equivalent of a shopper scrolling past a sparse listing and clicking it anyway.

That shifts where the risk sits. On a traditional feed, thin data costs you ranking. On an assistant surface, thin data can make a product effectively invisible for a whole class of question, and you will not see it in any report because there was never an impression to count.

What a retailer should do

The first step is unglamorous: check how complete your product data really is, attribute by attribute and category by category. Most catalogues are far patchier than the team believes, because the gaps sit in the long tail that nobody looks at.

Then make descriptions carry facts rather than atmosphere. Copy written to sound appealing on a product page often contains very little an assistant can use. Material, dimensions, compatibility, care instructions and intended use are what get a product into an answer.

Finally, keep the data in one place and generate each channel's feed from it. Feeds will keep multiplying, and the field lists will keep differing. Fixing each one separately is how a catalogue ends up saying different things about the same product in different places.