Can AI read
your website?

What canaisee measures

canaisee grades any public web page on how readable it is to AI agents. It runs the same battery of checks a modern language-model crawler runs: it fetches the page with a real HTTP client, inspects the HTML, pulls robots.txt, sitemap.xml, llms.txt, and the /.well-known/ discovery files agents read, and compares what the raw response contains to what a headless browser sees after JavaScript runs. The result is a single headline grade backed by a transparent, versioned sub-score breakdown you can argue with.

Checks span five categories. Crawler accessibility asks whether agents are welcomed in the first place: whether robots.txt opens the door to the crawlers that decide citations and the fetchers that produce referrals, whether the edge lets an honest crawler through, whether the main content survives a plain HTTP fetch without running JavaScript, whether a login or paywall blocks the route. Content readability asks how much of the page is real content versus chrome, what the page costs an agent in tokens, and whether anything is hidden from readers. Semantic structure covers the hooks agents use to orient themselves: headings, structured data, machine-readable dates, a sitemap, a declared language. Agent operability asks whether a browser-driving agent can actually work the page: named controls, labelled fields, landmarks, and a first load that is not gated by a consent wall or a CAPTCHA. Agent-native surface is the frontier: Accept: text/markdown negotiation, an AI catalog at /.well-known/ai-catalog.json, a usage-policy line in robots.txt, advertised markdown alternates, a valid llms.txt, WebMCP tool declarations, A2A agent cards, and, for stores, the UCP and ACP commerce protocols shopping agents transact through.

Why agent-readability matters

AI systems increasingly answer questions about your site without sending a human visitor. Whether those answers are accurate, attributed, and up-to-date depends on whether the model can actually read the page. If your robots policy blocks the crawler, if your edge challenges it, if your content only materializes after a React bundle executes, if your copy is buried under a kilometer of analytics scripts — the model guesses, or hallucinates, or cites someone else. The cost of being invisible to agents is no longer theoretical: in competitive categories the citations a model issues are already shaping where traffic, attention, and authority flow.

canaisee is opinionated about this. We believe the web of the next decade is being read by agents first and humans second, and that sites which invest in agent-readability now will compound that investment the same way sites which invested in mobile-readability a decade ago did. The grade is a snapshot of how ready a site is for that transition. The rubric is public, versioned, and designed to be argued with. This section is visible on purpose: prose that only agents can see is the kind of signal the rubric no longer rewards.

How to read your scorecard

Every scan returns a composite grade (A+ through F) on a 100-point scale. The composite blends Hygiene (the table-stakes tier: crawler accessibility, content readability, semantic structure, agent operability) with Frontier (the agent-native tier): 0.90 × Hygiene + 0.15 × Frontier for content sites, 0.85 / 0.20 for interactive and commerce sites, clamped to 100. A site with perfect Hygiene and zero Frontier lands at B (85), the top of the expected range for a well-built modern site. Frontier features push a site into A and A+.

Every check surfaces its evidence and its flags, so the grade is inspectable: you can see exactly what the crawler fetched, what it extracted, what it flagged, and what it would take to move the number. The full rubric — every check, every weight, every cut-point — is published at /rubric and versioned as the web changes.