Checkout in chat comes after the shortlist

Someone asks ChatGPT which running watch to buy under €300. You are not recommended.

A blog post still tells you that you are ready for agentic commerce, checkout or stock lookup inside a chat, because that checkout is coming, or your feed is clean, or Merchant Center looks fine.

Shops that treat checkout and a recommendation as one milestone end up chasing the wrong one. How recommendations get chosen: How AI chatbots build a shortlist.

Four different jobs

People use one marketing phrase for several different products. ACP is Agentic Commerce Protocol, UCP is Universal Commerce Protocol, and both are checkout protocols inside a chat.

In plain words:

Getting recommended still means the chatbot names you, with the right facts. That still depends on reviews, fit to the question, a clear identity, and what the tool retrieved. How AI chatbots build a shortlist.

Checkout answers a later question: once someone has already chosen you, can they pay without leaving the chat? Useful, but later.

Northloop’s mix-up

Illustrative example, not a logged test. Northloop is a fictional example brand on this site.

Northloop’s Trail Watch has matching JSON-LD and a feed row, which helps whenever a shopping tool needs a correct price.

Being “ready for checkout in ChatGPT” means an assistant might be able to complete a purchase after the customer has already picked that watch.

Neither of those makes an assistant name Northloop first when someone asks for the best GPS watch under €300. That still comes down to reviews, roundups, how well known you are, and whether the tool found you at all.

How to read the sales pitch

When a platform says “agentic commerce,” work out which of these five they mean:

1. Sending them a product feed
2. Letting an assistant query your catalog live
3. Checkout inside the chat
4. Showing up in answers (named or cited)
5. On-site agent-native search — the retailer’s own search resolving agent queries, not getting named somewhere external

Only the fourth is what most shops mean by “AI recommends us.” The rest is setup, worth paying for after you are actually recommended, and only if your buyers will actually use those paths. Training crawls in your logs are not shoppers either — your catalog can be perfect and still never show up splits those jobs.

You can take payment inside chat and still never get recommended.

A clear buy or book URL on your own site is a different job again. It tells the assistant where someone actually completes the purchase or reservation, which reduces invented “available at Shop X” lines. That is not the same as wiring up ACP or UCP checkout inside the chat.

For anyone who asks, the reverse of a shop that can take payment and still never get named is an assistant that names a product and can pay before you compare sources. Trust gets harder in three steps:

1. You still compare sources yourself — you can notice when only a brand is talking about itself.
2. You ask the assistant — the paraphrase reads like a recommendation, so that cue disappears.
3. An agent that recommends and pays — you never reach either earlier step to apply judgment at all.

ChatGPT recommended a deodorant that didn’t exist is the stress test for step two; checkout in chat is step three.

WebMCP

WebMCP is a proposed browser standard that lets a page expose actions as tools, so an agent already on the page does not have to click around. Merj reports a Chrome origin trial in June 2026 and support in the ChatGPT desktop app in August 2026, with Mozilla neutral and WebKit opposed. Support remains Chromium-heavy. WebMCP covers acting on a page the agent has already reached. It says nothing about whether an assistant names you. Reviews and product copy can still instruct an agent that is about to pay — see The assistant said the table was booked.

Read next if: Why you got the same five landmarks · AI place answers come from different directories