---
title: "OpenAI's plugin guide asks stores to return the facts our browser agents get stuck on at checkout"
canonical: https://www.checkoutbench.ai/blog/openai-plugin-guide-checkout-failures
date: 2026-11-18
---

# OpenAI's plugin guide asks stores to return the facts our browser agents get stuck on at checkout

We matched each field in OpenAI's commerce plugin guide to the step where agents stopped in our runs at real store checkouts.

Video: [OpenAI's plugin guide asks stores to return the facts our browser agents get stuck on at checkout](https://www.checkoutbench.ai/img/blog/openai-plugin-guide-checkout-failures.mp4)

OpenAI has published a [guide to building a commerce plugin](https://developers.openai.com/commerce/guides/build-a-plugin) for ChatGPT. A plugin is an MCP server plus skills that lets ChatGPT call a store's systems directly. The guide lists what each tool should return. Checkout Bench sends browser agents through real store checkouts, and our run records show where they stop. This post follows the guide's list in order and puts our runs next to each item.

## Product lookup

The guide says the lookup tool should answer one question: can this customer buy this exact product, in this variant, at this price, in their ZIP code? Its sample response carries the answer in a few fields:

```json
"selected_options": { "color": "green", "size": "standard" },
"availability": {
  "status": "in_stock",
  "basis": "requested_zip_code",
  "message": "Shipping is available to 10004."
},
"checkoutable": true
```

**The exact variant.** The guide asks for stable IDs for the product, the selected variant and the seller's offer, plus each variant's options, price, stock and images. In our runs, the most common reason a quote was blocked was that the run could not confirm a listing matched what the shopper asked for, down to size or pack. That reason accounts for 194 of 470 blocked quotes. In one Instacart run, the shopper asked for regular cola and search returned a 24-pack. Nothing on the listing confirmed it was the requested variant, so the agent stopped and asked.

**An ID that lasts until checkout.** The guide's sample keeps the item ID and SKU on every variant. In one Amazon run, DeepSeek V4.1 Flash asked for quotes on two wireless microphone kits, and all five were blocked as product mismatches. The checkout page showed a longer title than the product page and no product ID, so the item at checkout could not be tied to the one on the product page.

**Availability for the shopper's location.** The guide asks for stock at the requested ZIP code and separates in stock, out of stock, unavailable in this location and unknown. On our checkout walk, LCSC let us add an out-of-stock item to the cart. In an Uber Eats run, the agent searched from the account's default address, found no matching restaurant, and told the shopper that was the address they named, with no evidence the two matched.

**Checkout eligibility.** The guide says an item in stock is not necessarily one that can be checked out, and asks for a reason when it can't. Our Instacart runs hit this when a cart was under the store's minimum order, which blocked the quote.

**Fulfillment and cost.** The guide asks for shipping, delivery and pickup options at the shopper's location, with costs given only when known and estimates labeled. We walked 25 store checkouts to the payment page, and at 10 of them the total on the last screen was higher than the first price shown, mostly because of shipping. LCSC showed a lower total when no address was on file.

## Cart

**Guest carts that survive sign-in.** The guide asks stores to allow product lookup and carts without sign-in, and to keep that cart when the customer signs in at checkout. In our production records, fresh sign-ins, where the agent logs in as a returning member, worked 49% of the time. Ulta asked for a password and then an SMS code, and Ticketmaster held the agent at sign-in for most of a day.

**A stable cart ID.** The guide asks for a persistent cart ID, the current line items after every change, and rejected changes reported as errors. In our runs, the agent works in the same account cart the shopper uses. At Chewy, the shopper's existing cart was set aside before each attempt and restored after it. Our Uber Eats driver will not add items when the shopper already has a cart at that store. A cart the agent could address by its own ID would remove the need for both.

**Cart tools, alongside lookup.** In one run we called a store's product tool directly, with no browser. The Best Buy tool returned the SKU, the price and free shipping for a 6 ft USB-C cable, and the run ended there because the tool had no cart step.

## Checkout

**Required steps as messages.** The guide's sample checkout returns the status `requires_escalation` with the error code `shipping_address_required`, the field path, and a sentence telling the buyer what to choose. On our 25 checkout walks, the address was the most common reason a store stopped for the person, ahead of payment, password and name.

**A handoff that keeps the session.** When the plugin can't finish a step, the guide asks for a `continue_url`, ideally a short-lived, one-time authenticated link that reopens the same cart and checkout. The guide notes that a regular checkout URL carries neither that authentication nor that continuity.

**Wallet and direct checkout.** The guide says direct plugin checkout should assume the customer signs in, or enters all their details as a guest. It also says OpenAI is still working on how third-party plugins and ChatGPT Wallet interact.

## Reordering

The guide lets a signed-in customer search their past orders and add an item at its current price and availability. In one Amazon run, the shopper asked for the toothpaste they always buy, the white tube and not the kids' one. The agent quoted an adult gel at the right quantity and never established that it was the usual tube. The guide's `find_past_items` tool returns matching items from the customer's own orders.

## Speed

The guide says a plugin is not automatically faster and tells stores to measure the full flow. We record the time to each stage of a browser run, which gives a store a browser baseline to measure its plugin against.

The items above match steps where our agents stop at real stores today. We can run agents through your checkout and show which of these steps stops them. [Request a private test of your store](/#submit-site).

## Sources

- [OpenAI: Build a plugin](https://developers.openai.com/commerce/guides/build-a-plugin), the commerce guide quoted throughout.
- The blocked-quote reasons, the sign-in rate, the 25 checkout walks and the run examples are on the [Research page](/research).
