AI Shopping Agents Read Your Product Feed, Not Your Product Page

Start with the disappointing part so you can plan around it: onboarding a product feed into ChatGPT "is currently available to approved partners," per OpenAI's merchant documentation. You request access through a form and wait. What you should not do is wait to do the work, because the thing that makes a catalogue legible to a shopping agent is the same thing that has made it legible to Google Shopping for a decade: a clean, complete, correctly typed product feed. That work is available today, it is measurable, and it pays off on a surface you already have.

Here is the hour, in order.

1. Fix identity before anything else

Google's product data specification puts id first for a reason. Maximum 50 characters, unique per product, and, critically, stable across updates. If your platform regenerates product IDs when you edit a listing, or if a variant reuses its parent's ID, every downstream system loses the thread: performance history, review association, price change detection. Check this first because everything else is keyed to it.

Then the identifier trio. brand is required for new products, with exceptions for movies, books and musical recordings. gtin is "strongly recommended if available" and is what actually matches your product to a known catalogue entry. mpn is required only where there is no manufacturer-assigned GTIN. If you genuinely have none of brand, GTIN or MPN, which is normal for handmade or own-design goods, you set identifier_exists to no rather than leaving the fields blank. Blank fields read as an incomplete record; an explicit no reads as a deliberate one.

2. Write the title for a matcher, not a shopper

Title is capped at 150 characters, and Google's guidance is explicit that it should present the product clearly and avoid promotional language and all caps. "SALE!! BEST Handmade Ceramic Mug FREE SHIPPING" fails on two counts and tells a matching system almost nothing. The pattern that survives both human scanning and machine parsing is attribute-ordered: brand, product type, then the distinguishing attributes in the order a buyer would narrow by.

• Weak: "Beautiful Handmade Mug, Perfect Gift!"
• Strong: "Northfield Ceramics Stoneware Mug, 350ml, Speckled Blue, Dishwasher Safe"

The second one answers the three questions an agent has to resolve before it can recommend anything: is this the category asked for, does it match the stated constraint, and is it distinguishable from the other eleven mugs in the result set.

3. Make description do work, and keep links out of it

Up to 5,000 characters, must accurately describe the product, and must not contain links or references to competitors. Most small catalogues waste this field on brand voice. The version that helps a machine answer questions puts the facts that buyers filter on into plain sentences: materials, dimensions, capacity, compatibility, care, what is in the box. If a shopper would ask it before buying, and the answer is not in the feed, the agent cannot answer it either, and an agent that cannot answer tends to recommend something else.

4. Get price and availability mechanically right

price is a number plus an ISO 4217 currency code, as in 15.00 USD. availability takes one of exactly four values: in_stock, out_of_stock, preorder or backorder, and availability_date becomes required if you use either of the last two. This looks trivial and it is where most feeds actually fail, because the failure is not a rejected record, it is a stale one. A price that is right in the feed and wrong at checkout is the single fastest way to be dropped from an automated surface, and it is also a returns problem, which is the expensive kind, as we set out in returns are a margin problem you can design around.

OpenAI's guidance on refresh cadence is worth adopting regardless of platform: send "the entire feed once a day via file upload, and then send updates throughout the day via the API." Daily full truth, intraday deltas for the fields that move.

5. Sort the images out before the deadline that is already published

image_link points to the main image and must use a verified domain over http or https. Google enforces a minimum of 500 by 500 pixels from 31 January 2027, which is far enough away to ignore and close enough that a re-shoot planned now costs nothing extra. Two related judgments belong here: where generated imagery helps and where it costs you, which we worked through in AI product images for small stores, and the disclosure expectations that are already appearing on marketplaces, covered in Amazon's tagging requirement.

6. File the access request, then stop building

Request ChatGPT merchant access at the form OpenAI links from its get-started guide, and note the category exclusions before you do: adult content, age-restricted products, harmful or dangerous materials, weapons and prescription-only medications are all prohibited. If your catalogue is in one of those, this surface is closed and the feed work still pays off in Google Shopping.

Then stop. Do not build a checkout integration on spec. The Agentic Commerce Protocol, the open standard Stripe and OpenAI announced on 29 September 2025, is deliberately designed so you do not have to: the Shared PaymentToken "lets applications like ChatGPT initiate a payment without exposing the buyer's payment credentials" and is "scoped to a specific merchant and cart total," while merchants "accept or decline the order, charge the payment method, calculate and remit sales tax, and handle fulfillment and returns, as they normally would." The order lands in the systems you already run. Build the feed; wait for the door.

What we could not verify, and why it matters to your forecast

Two gaps worth naming. First, no page we opened at OpenAI or Stripe states what commission OpenAI charges merchants on an agentic sale. Anyone modelling a margin on this channel is modelling a number nobody has published. Second, OpenAI's feed specification page describes "a shared flat-file schema" with required and optional fields, but the field list itself did not render for us, so nothing in this article claims that Google's attribute names are OpenAI's. Google's spec is used here because it is published, openable and enforced, and because a catalogue that satisfies it is in good shape for any consumer of structured product data.

The broader reason to do this now has nothing to do with ChatGPT. Product feeds are becoming the canonical representation of your catalogue across every surface that is not your own website, at exactly the moment when traffic to your own website is being measurably reduced by AI summaries. The page you spent money on is increasingly not what gets read. The feed is.

The hour, as a checklist

• Export your current feed and confirm every id is unique, under 50 characters, and unchanged since last month.
• Set identifier_exists to no on every product without brand, GTIN or MPN, rather than leaving those fields empty.
• Rewrite the ten titles on your best-selling products to brand, type, attributes order, with no promotional words and no all caps.
• Confirm price carries a currency code and matches the live checkout price on those ten products.
• Confirm availability uses only the four permitted values, and that anything set to preorder or backorder has an availability_date.
• List every main image under 500 by 500 pixels and schedule the re-shoot before 31 January 2027.
• Set a daily full upload and, if your platform supports it, intraday API updates for price and stock.
• Submit the merchant access request and diary a check in 30 days.