The agent-readiness score

The score answers one question and nothing else: if a customer’s AI assistant arrives at this site and is told to book, quote or enquire, how much of what it needs is actually there?

It is five categories, 20 points each, 100 in total. Each category is capped at 20, so no single check can carry a site, and a category can only earn points for something the scanner positively found. There is no "nothing looked broken" route to a good score.

The three bands #

The five categories #

1. Crawlability and discovery: 20 points #

Before an agent reads a page it checks whether the site has published anything about itself at the addresses agents look at. This category is that check.

2. Structured data: 20 points #

Whether the page tells a machine what this business is, in the format machines already read. With no structured data at all, an agent has to guess the business type, the hours and the address from prose.

3. Action affordances: 20 points #

Is there anything on the page for an agent to do? A booking or quote form, a known booking widget, a clickable phone number, real buttons rather than clickable divs.

Two filters do most of the work here, and both exist because an earlier version of the scanner counted things that were not actions.

A form matching an action keyword is necessary but not sufficient. It is excluded, from the detected actions and from the form-quality score, when it carries a positive signature of being something else: a search box (role="search", a search input, a search/typeahead token in its attributes, or a field named s, q, query, search or keyword), a newsletter signup (a newsletter/subscribe/mailchimp token, or a lone email field, or newsletter wording when every field is an email or a checkbox), or a form with nothing to submit. A genuine enquiry form carrying a newsletter opt-in tick-box is not excluded, and a genuine one-field-plus-submit enquiry form still counts: the floor is structural, never field count alone.

Widget signatures are matched structurally, against URL-bearing attributes only, and each vendor pattern targets that vendor’s booking endpoint rather than its marketing site. A "powered by" credit in a footer, a blog paragraph naming a vendor, or a CRM script tag is not a booking widget. Under the earlier whole-page substring match, a test page that merely named four vendors in prose emitted four booking actions and gained six points.

4. Form quality: 20 points #

Assessed over action forms only. A site’s form quality is not scored on the strength of its property-search box.

5. Agent interface: 20 points #

Whether the site declares what it can do to software, rather than leaving software to infer it from the layout.

WebMCP is a proposal. A page declares its tools through document.modelContext, which browsers have not shipped; navigator.modelContext is a deprecated alias, and using either today means loading the @mcp-b/global polyfill.

The ARD catalogue at /.well-known/ard.json, found via a rel="ard" link, was published as a discovery standard in June 2026. Its earlier path, /.well-known/ai-catalog.json, is still served for registries built against it. A persistent MCP endpoint is the third route, and the one that keeps working when nobody has the page open. The scanner finds it either from a <link> to a server card in the page or from a server card published at /.well-known/mcp/server-card.json or /.well-known/mcp.json.

Frontlatch emits to all three from one Action Graph, so a business does not have to bet on one of them.

Agent-callable is a separate answer, and it is not part of the score #

A site can have a contact form and still have no action an agent can call directly, a browser-driving agent might still fill it, but that is a different and weaker thing than a machine calling it on its own. So each detected action carries its own verdict, computed from positive evidence only, and it deliberately does not feed the 0–100 number. The score describes a site; agentCallable describes one action.

WebMCP tool declaration, advertised MCP endpoint, ARD catalog, known booking widget,
or a fully machine-completable form (every field labelled and named, an email/tel
reply channel, a real submit control, no CAPTCHA)

What fifteen real sites scored #

Fifteen service businesses, five trades, five clinics, five real estate agencies, scanned read-only on 25 August 2026. The fifteen sample report pages were retired on 9 October 2026; the run’s figures below stay as its record, and every business in the index now has its own live page.

What the score does not claim #

This section is published because an instrument that hides its error bars is not an instrument. Every item below is recorded in the scanner’s own repository, not discovered here.

• The bare instrument reads one page, fetched once, and does not run JavaScript. A booking flow two clicks deep scores as if it were not there. That is the correct measurement of what an agent gets on arrival to that one page, and it is not a measurement of what the business can do. A live scan run today, at /get-scanned or a rescan from /app, wraps the same instrument with a renderer and a crawl of up to 15 pages, so it already sees more than a bare run of it would; the fifteen worked examples below were scored by the bare instrument and are not rescanned this way.
• It understates most sites, and we know roughly by how much. Opening the same fifteen sites in a real browser, rendering, waiting for scripts, clicking, following links, reached 28 actions this pass did not see, on 14 of the 15 sites.
• A file scan is not a URL scan. Scanning a saved HTML file skips the robots.txt, sitemap, llms.txt and catalogue probes entirely, so the same page can score up to about 11 points higher as a live URL than as a fixture.
• It is not a ranking. A category score is a direction; the fix list is the useful part. Its known measurement defects are recorded in the repository rather than in a footnote, and the fifteen scores in this page’s worked examples are not re-scored when the scanner changes.
• It says nothing about mappability or execution success. Those need a real browser and a sandbox, so the benchmark summary prints n/a for both rather than substituting the detected-action rate as a proxy. The detected-action rate is mappability’s ceiling, not its stand-in.
• A zero in the agent-callable column is a fact about those sites, not a property of the instrument. It is computed from positive evidence, and one site in fifteen did earn it.

The scanner is a read-only visitor. It scans one site at a time, two seconds apart, with a 45-second timeout and one retry. Within a site it fetches robots.txt, the page and six well-known files (llms.txt, the ARD catalogue at its current and earlier paths, two MCP server-card paths, and sitemap.xml) one at a time, half a second apart, so a site sees eight spaced requests and never a burst.

Every request carries a user-agent that names the project and says the scan is read-only. A form is never filled and never submitted.

How the measurement is made #

npm install
npm run scan -- https://somebusiness.com --html report.html --json report.json
npm run scan -- fixtures/bad-plumber.html      # offline fixture mode
npm test                                        # tsc --noEmit, then scoring, callability and publication assertions on 10 fixtures
npm run build:public                            # regenerate publish/public-reports.json from the committed benchmark reports
npm run build:public -- --check                 # fail if that bundle is stale (also asserted by npm test)

You do not install the scanner. To scan your own site, enter your domain at /get-scanned and the report opens in the same tab when the scan finishes. The commands above show how the sample reports were produced, so the method can be checked.