How ChatGPT Shopping Works for Merchants (2026)

On 25 September 2026, ChatGPT shopping is two different projects that vendors keep merging. One is discovery: your products show up when someone asks what to buy. The other is checkout: an approved merchant completes the order inside the conversation. Most stores are in the first. Almost none should assume the second is a setting.

Who this is for. An eCommerce leader or architect deciding what to build for ChatGPT. The protocol comparison is ACP vs UCP vs MCP. This page is the merchant path for the OpenAI side only.

Our take: publish a feed you can defend. Do not staff an ACP checkout build until a partner agreement exists and the catalog survives a comparison.

What OpenAI actually documents

The Agentic Commerce Protocol is maintained by OpenAI and Stripe and the repository still marks it beta. It was released 29 September 2025 under Apache 2.0.

OpenAI’s checkout spec says merchants integrate by implementing the Agentic Checkout Spec, and that Instant Checkout in ChatGPT is currently available to approved partners. The same spec says the point is to keep orders, payments, and compliance on the merchant’s existing stack.

OpenAI’s product discovery note extends ACP so merchants share product feeds and promotions. It names a Shopify Catalog path: Shopify merchants’ product data is already integrated, and OpenAI says no additional feed work is required from those individual merchants. It also says the first Instant Checkout did not give merchants the flexibility OpenAI wanted, so merchants may use their own checkout while OpenAI focuses on discovery. ChatGPT apps remain a separate, deeper option (Walmart is the example they give).

Those two pages can both be true. Discovery is the wide door. In-chat complete is a partner program, not a default.

What the shopper sees versus what you run

ChatGPT is not your storefront theme. It will not apply your onsite merchandising rules unless those rules are in the data you publish.

The five endpoints, if you are approved

The 2026-01-30 checkout OpenAPI says the merchant implements:

• POST /checkout_sessions — create, return 201 and a rich cart
• POST /checkout_sessions/{id} — update (address, fulfillment)
• GET /checkout_sessions/{id} — retrieve
• POST /checkout_sessions/{id}/complete — pay and create the order
• POST /checkout_sessions/{id}/cancel

Every response is supposed to carry authoritative cart state: items, pricing, tax, shipping, discounts, totals, status. Optional headers worth honoring: Idempotency-Key, signature, timestamp, request id. The spec allows affiliate_attribution on the agent side. That field is a hint, not an analytics platform. What you can actually report is covered in agentic commerce attribution.

A later spec tree, 2026-04-17, adds feeds, carts, orders, authentication, and MCP. Pin the version you implement. Do not mix 2026-01-30 payloads with 2026-04-17 assumptions.

Where this fails

What broke — A team mapped “be in ChatGPT” to a custom app with write_orders because a demo needed to place a test order. The feed was never validated. Detection: two prices for one SKU, one in the feed and one in checkout, and a token that could refund. Fix: delete the admin token, publish one feed, and stop the checkout build until partner status and a price match exist. Lesson: discovery does not require order-write scope.

Other misses:

• Variants encoded in the title instead of structured options. The agent cannot tell size from color.
• Inventory without a timestamp. You sell a unit the warehouse already allocated.
• Logging the payment credential beside the order JSON.

The catalog work is the AI-ready product catalog. Readiness checks are shopping-agent readiness.

If you only do one thing

Open the product that would embarrass you if an agent quoted it: missing size, stale price, or a description that contradicts the spec table. Fix that record before you talk about endpoints.

What to do this week

1. Read OpenAI’s checkout page and the product-discovery note. Confirm whether you are an approved Instant Checkout partner. If not, do not schedule the five endpoints.
2. If you are on Shopify, audit Shopify Catalog accuracy instead of building a second feed that can drift.
3. If you are not on Shopify, define one feed and a refresh SLA.
4. Compare Google’s path in Gemini shopping for merchants before you fund both.
5. Prepare the commerce platform with agentic commerce readiness when the catalog is the blocker. The field guide hub is /resources/ecommerce-ai-agents/.

What this post doesn’t cover

• A revenue estimate for ChatGPT referrals. We do not have one.
• Google’s Buy button. That is UCP.
• Building an agent that buys on behalf of your company. That is the outbound agent program, not this page.
• PCI redesign. Delegate tokens do not shrink the cardholder environment on your side; they change what the agent is allowed to hold.