Can an AI agent do business with you?
Your next customer is asking an assistant, not a search box. This checks whether that assistant can find you, understand what you do, and actually book or buy — the way it would for a real person, right now.
- FindCan an agent discover you?
- ReadCan it understand what you do?
- ActCan it book or buy?
Discovery
—
agents can read you
Completion
—
agents can act on you
Scroll it. It stops where you need to look.
In the order an agent hit them. What is working stays out of your way — the page catches on what is not, one at a time.
02Allowed in
Assistants are told to stay out, so when someone asks one to recommend a business like yours, you are not in the running.
Assistants are allowed in. You are in the running when someone asks one to recommend a business like yours.
robots_ai
This is not the same as being blocked from Google — search engines are still fine. It is the assistant your customer is actually talking to that has been turned away, usually by a rule nobody meant to apply to it.
03Read you
Your visitors see your site. An assistant sees a nearly empty page, because your content needs a browser to appear and it is not using one.
An assistant sees the same page your visitors do — your content does not need a browser to appear.
cloaking
It reports what it actually received, not what you meant to publish. A page that renders beautifully for a person and returns a shell to a machine is worse than a plain one, because you cannot tell from looking.
Your page structure is ambiguous to a machine — it cannot reliably tell a button from a heading from a price.
Your page structure is unambiguous — a machine can tell a button from a heading from a price.
a11y_tree
The same thing that makes a page hard for a screen reader makes it hard for an agent. Fixing it helps both, and it is the same fix.
Content shifts around while the page loads. An agent never notices. A customer on a phone does.
The page settles as it loads rather than jumping around under a reader.
layout_stability · doesn’t affect score
04Understood you
An assistant can read your words but cannot tell what you are or where, so it has to guess.
An assistant can tell what you are and where, without guessing.
structured_data
It guesses conservatively, and then recommends the business that did not make it guess. This is the single highest-value thing on the page to fix.
When your link is shared or summarised there is no title or description to use, so the machine writes its own.
When your link is shared or summarised, it uses your words rather than its own.
meta_og
Nothing tells language models how to describe you, so they infer your pitch instead of being handed it.
You tell language models how to describe you instead of letting them infer it.
llms_txt
One small text file at a known address. Cheapest fix in the report.
An assistant has to guess which of your pages matter by following links, so anything not linked from the homepage may never be found.
There is a map of your pages, so nothing important is missed.
sitemap
05Act
Someone using an AI assistant cannot submit your contact form. They wanted to reach you and could not.
An assistant can fill in and submit your forms on someone’s behalf.
forms
Your inputs have no name or autocomplete attributes, so an assistant filling the form for someone cannot tell which box is the email. It gives up and tells them to visit the site themselves, which is the moment most people do not.
Your primary actions only work if a human clicks them in a browser, so an agent finds no link to follow.
Your key actions are real links an agent can follow.
actionable_links
Agents follow links. They cannot fire JavaScript click handlers. A button that should have been a link is the most common version of this, and usually the easiest to fix.
Something on your key pages refuses non-human visitors — a captcha or challenge page. To an assistant that is a locked door, and it does not knock twice.
Nothing on your key pages refuses a non-human visitor.
human_gates
There is no machine-readable description of how to work with you, so an assistant that wants to hand a job to you has nowhere to hand it.
You publish a machine-readable description of how to work with you.
agent_card
Almost nobody has this yet, which is exactly why it is worth having. It is one small file at a known address.
There is no published way for an assistant to browse or buy from you, so if your customer’s assistant is doing the shopping it cannot shop here.
An assistant can search your live catalog programmatically, using the standard shopping agents are adopting.
ucp
This is the one that decides whether you are part of agentic commerce at all. It matters the moment you sell online, and not much before.
An assistant acting for a customer — checking an order, managing a booking — has no standard way to sign in, so it does not try.
An assistant acting for a customer has a standard way to sign in.
mcp
Your prices and availability are on the screen but not in a form a machine can read, so an assistant comparing options leaves you out of the comparison.
Your prices and availability are machine-readable, so you appear in comparisons.
product_schema
Your site exposes no in-page tools an assistant could call directly, so anything it does for a customer it has to do the long way.
Your site exposes in-page tools an assistant can call directly — rare air.
webmcp · doesn’t affect score
Almost nobody has this. It is what "agent-ready" will mean next rather than what it means today.
There is no DNS record naming an agent or tool server you run. No major assistant looks for one yet, so today this costs you nothing.
A DNS record names the agent or tool server you run, for any assistant that looks for one.
aid · doesn’t affect score
An emerging standard (Agent Identity & Discovery) that no major assistant reads yet, which is why it is reported here and not scored.
Take it away, make sense of it, or hand it to us.
The report is yours either way.
01
Download the report
All seventeen checks, what each costs you, and the technical name for whoever fixes it. Yours to keep and forward.
02
Make sense of it
Ask questions of it in plain language, using the assistant you already have.
Copies your results and a question, ready to paste into ChatGPT, Claude or Gemini.
03
Have us fix it
Some of these are a few hours of work. The rest are the frontier, and we build those for a living.
Fix this with usWe re-run the scan when it is done, so the number moves where you can see it.