The agentic protocol landscape
Updated
ResourcesGuide
The Commerce Agent Index
Every AI shopping agent we've put through a real store, side by side: how it reads a site, whether it found the store on its own, how far it got toward paying, and what it tripped on. Updated each time we test a new one.
Last updated: 10 November 2026 · 3 agents tested · Next up: Instinct
AI agents that shop on someone’s behalf are not a forecast. Meta, OpenAI and Anthropic each have one in people’s hands today, and each of them has already walked through a real store’s catalog, built a real cart and stopped one click short of paying.
We know because we sent them. Every agent on this page was handed a shopping list and the same storefront: an independent lighting showroom with 130,000-plus products, not on Shopify, built and run on our platform. We watched what each one read, where it got stuck and what it said about the store, then fixed what it found, usually the same day. Each row links to the full field test.
The index
| Muse | Dot | Duet | |
|---|---|---|---|
| Made by | Meta | OpenAI | Anthropic |
| We tested it | September 2026 | 1 October 2026 | 4 November 2026 |
| How it reads a store | Checks a product catalog first; sends a browser only if that comes back empty, and reads pages like a reader-mode view | Its own cloud browser for shopping; its research tools read a cached copy of the web about a month old | Its own virtual machine and browser; compares several retailers at once and reads the page body, not the head tags |
| Found the store on its own? | No. It shortlisted big-box retailers until we named the store | No. We handed it the address | Yes, by its own account, from local listings |
| How far it got | Real cart, real checkout, paused at final review ($40.37) | Product page through to the payment step | Built a $109.39 spend request in Stripe’s Link wallet and stopped for approval |
| Said it was an agent? | No. It looked like an ordinary browser on a datacenter IP | Yes. OpenAI signs its requests | Yes. The spend request named Claude |
| What tripped it | Saw 255 characters of the homepage; a search filter we didn’t support returned the whole catalog without an error | Couldn’t read anything we’d published since its last crawl | Our own robots.txt rule refused 169 of the 218 shop links in the menu |
| What we shipped | One plain sentence pointing to our agent guide, plus a search fix, on every storefront that day | A live catalog endpoint on an open standard (UCP), in about two hours | Categories moved into the URL path; refused links fell to 2 of 218 |
| Full field test | What Meta’s Muse sees | What OpenAI’s Dot sees | What Anthropic’s Duet sees |
No agent placed an order. We stopped every run at approval and cancelled.
What all three have in common
They fail quietly. Not one of these agents threw an error when it got the store wrong. Muse presented a $19,998 chandelier as a kitchen pendant because a search returned everything instead of nothing, and the server answered “OK” the whole time. If an agent misreads your store, your logs will look normal.
They read differently from people, and from each other. One reads a stripped-down version of the page. One reads a month-old copy. One reads the body and ignores the tags in the head. A site that looks perfect to a person can be close to empty to an agent. Plain words on the page beat clever markup.
Finding a store is still the hard part. One of three found the store without being told where it was. The other two were capable shoppers once they arrived. Being easy to read matters only after an agent gets to you, and for now most of them have to be pointed there.
The checkout is the store’s own. All three shopped the store’s normal site and reached its normal checkout. None needed a special agent checkout to get there. What they need is a site they can read, products described in clean, consistent fields, and nothing in the way of the pages.
They’re starting to say who they are. Muse looked like an anonymous browser. OpenAI signs its traffic. Duet’s payment request named Claude in plain text. The payment step is becoming where an agent introduces itself, and that’s where stores will first be able to tell an agent sale from a human one.
What this means for your store
You don’t need to understand any of the protocols to act on this. Four things, in order:
- Ask an agent to shop your store, and read what it says back. It will tell you what it sees. It won’t tell you when it’s wrong, so check its answer against your real prices and stock.
- Make the site cheap to read. One plain sentence near the top saying what you sell and where an agent can find your catalog does more than any tag.
- Check nothing is locking agents out. A rule written years ago to keep search crawlers out of filter pages can block every agent that follows the rules. Ours did.
- Clean up the product data. Agents compare on fields like size, finish and price. A wrong dimension or a mismatched finish is a lost comparison you’ll never see.
If you’d rather see where you stand first, run the free agent-readiness check. It reads your site the way these agents do.
How we test
- Our own sessions, our own store. Every run is ours, on a storefront we build and operate, so we can see the server side of what the agent did.
- Same store, same kind of shopping list. That keeps the comparison about the agent, not the store.
- We stop at approval. No order is placed; we cancel at the final step.
- This is not a score. An agent’s opinion of a store is its opinion. These are field notes from one store on one day, and agents change weekly. When we start running repeatable tests with a published score, it will say so here.
- Disclosure: we’re in the Claude Startup Program and build on Claude every day.
Common questions
What is an AI shopping agent?
An AI that shops on a person’s behalf: it searches, compares, builds a cart and, with the person’s approval, pays. The ones on this page run in their own browser or virtual machine, so they visit your site much as a person would, but they read it very differently.
Are AI shopping agents actually buying things today?
On our store, Muse, Dot and Duet each got to the payment step. Whether a purchase completes is up to the person approving it, and the payment step increasingly identifies the agent.
Do I need a special checkout for AI agents?
Not to be shopped. All three agents used the store’s normal checkout. What they need first is a site they can read and product data they can trust.
Which agent is next?
Instinct. We’ll add it to this page once the field test is published.
How current is this page?
We update it every time we test a new agent, and the date at the top says when that was. Older rows describe each agent as it was on the day we tested it; agents change quickly.
Want to know what these agents see on your site?
We build storefronts these agents can read, and we test them with the agents themselves. That work ships as our agentic commerce practice. For the protocols underneath, see the agentic protocol landscape.
