What retailers can document about AI shopping agents
IC Guides · AI Agents · Compare
When shoppers ask an AI assistant what to buy, the retailer’s product data decides whether the product appears. OpenAI’s Agentic Commerce Protocol (ACP) is one documented route. This guide covers what OpenAI’s developer docs say a retailer must do, and what they leave unanswered, especially on cost.
What ACP is
OpenAI describes ACP as an open standard that connects merchants and ChatGPT users. ChatGPT ingests structured catalog data, learns what a merchant stocks, and shows relevant products in context. The protocol is also published at agenticcommerce.dev and on GitHub.
Who can join
OpenAI states that onboarding product feeds in ChatGPT is currently available to approved partners only. Merchants apply through a form. Joining is not self serve.
What the retailer supplies
The retailer supplies a structured product feed with titles, descriptions, images, price and availability. It can deliver the feed by file upload, API or both. OpenAI generally recommends a full feed once a day by file upload, with updates through the day by API, and promotions data can only be sent by API.
Each purchasable variant needs its row, with a stable parent product ID and a unique variant ID, and variant specific title, price, availability and media where they differ. The feed also needs an accurate seller name, durable public policy links and valid, encoded URLs. OpenAI suggests adding tracking parameters such as utm_medium=feed for feed specific click tracking.
Rules that can cost you reach
OpenAI allows only legal, safe products suitable for a general audience. Prohibited categories include adult content, age restricted goods such as alcohol, nicotine and gambling, weapons, prescription only medicines and unlicensed financial products. The merchant is responsible for compliance, and OpenAI may remove a product or ban a seller from appearing in ChatGPT. A retailer with age restricted lines should expect part of its range to be excluded.
What the docs do not say
The pages we could read state no fee or commission for joining or for sales. They also leave open how agent checkout, payments and merchant of record work, because the current docs navigation lists feed specifications only. Nothing we found explains how products are ranked or selected, and there are no volume, conversion or revenue results.
Other shopping assistants may use different routes. We have not documented them here.
Questions to ask before you commit
- What does it cost, as a fee per sale, a subscription or nothing?
- Who handles payment and returns, and who is the seller of record?
- How do we measure sales that start in an assistant? A feed tag shows clicks. Revenue needs your analytics.
- How often must we refresh the feed, and what happens when prices or stock drift?
- Which products will be excluded under the policy?
Sources OpenAI developers, Commerce, Get started and Best practices, checked 12 October 2026. To model the return on AI tools in a store, use the eCommerce ROI Dashboard, and see the AI Agent Pricing Tracker for tracked vendors.
IC Lab · AI Agent Pricing Tracker
Vendor prices change. The tracker records when.
Every entry links to the vendor’s pricing page and shows the date we checked it. Open the tracker, or get major pricing changes by email every Friday in TalkHead.
Related guides Which AI agent, and what it costs · AI agent glossary · Permissions and safety · Salesforce Agentforce · OpenAI Agents API cost · AI Agent Pricing Tracker · All AI agent guides
Continue with the evidence
Why AI shopping agents struggle to understand intent The analysis explains the operational and commercial implications behind the tools. How Pinterest tested visual product discovery Read the measured company experiment and what its results can tell operators. Compare the pricing models side by side Compare billing units and linked vendor sources before budgeting.