Agent view — served from /agent-readiness.md.
How it works
Agent readiness is how well your website works for AI agents — the assistants and crawlers that now read, summarise and act on sites on a person's behalf. It is measured separately from SEO and from page speed, because it asks a different question: not “is this page fast for a human” but “can a machine use this page.”
The four layers
Fifty-five checks, grouped by the four questions an AI has about your site. The first three produce your score. The fourth is reported separately, because it measures the engines rather than your website.
Whether your site publishes the small machine-readable files that tell an AI what you are, what you offer, and what it is permitted to do with your content. Also whether your hosting is quietly blocking AI crawlers — which happens far more often by accident than by decision.
Typical failures: no signposting for AI, blanket crawler blocks inherited from a template, no distinction between crawlers that train on your content and crawlers that cite it.
Whether your pages survive being read by a machine. Does the content exist without JavaScript? Can the site serve a clean text version instead of forcing an AI to wade through markup? Does your structured data match what the page actually says?
Typical failures: pages that are empty without JavaScript, prices in the code that contradict prices on the screen, headings that describe nothing, and content that stops making sense the moment it is quoted alone.
Whether an assistant can complete something on your behalf: submit an enquiry, run a search, add to a cart, begin a booking. This is the layer almost every site fails entirely in 2026, which is exactly why fixing it early is worth something.
Typical failures: forms with no real labels, custom date pickers no machine can operate, bot protection that blocks legitimate assistants alongside malicious traffic.
A defined set of questions, run monthly against the engines that allow it, with every answer stored word for word and graded against your verified facts. We track whether you are mentioned, whether what is said is correct, which sources the engine cited, and whether the tone is shifting.
Reported separately. This measures the engines, not your site, so it never contributes to your score.
Method
If a language model can move the number, the number cannot prove anything about before and after.
So we split the audit. Deterministic checks — does this file exist, does this header return, does this content survive without JavaScript — are the only things that score. They are reproducible: run them twice on an unchanged site and you get the identical number.
The judgement calls, like whether a section really answers its own heading, are made with an AI and reported as advisory flags with evidence attached. They never carry points. That way, when we show you a month-on-month improvement, it is a measurement, not an opinion that happened to shift.
One honest limitation: the advisory half uses AI models that their providers update over time. We pin the version, record it with every flag, and note upgrades in your report rather than letting the analysis change silently.
See our own audited score →Where this differs
There are several free agent readiness scanners now, and they are useful. Run one.
What none of them do is generate the fix. A scanner tells you a file is missing; we build that file from your own site's structure, in the format your framework expects, ready to deploy. Then we re-run the identical audit next month and show you the difference on the same measuring stick.
That is the whole distinction, and we would rather state it plainly than pretend nobody else can count checks.
Populated with your content, not blank templates. Your developer reviews the diff and commits it.