AI Agent Readiness Checker
AI agent readiness is how usable a website is to an AI agent: software that arrives without a browser, reads raw HTML and machine-readable files, and tries to complete a task on a visitor’s behalf. This checker approaches a domain the way an agent would — plain HTTP requests, agent user agents, the well-known paths agents try first — and scores what it finds from 0 to 100 across four layers: discovery, access, usability and payments.
Want to know the moment your site breaks for agents — or for everyone? Monitor it free with myeventwatch — expiry and downtime alerts, no card required.
What the scan does, layer by layer
An agent’s visit starts nothing like a person’s. There is no clicking, no scrolling and usually no JavaScript: the agent fetches files and pages directly, decides from what comes back whether the site can be understood and used, and gives up quickly when it cannot. This scan replays that visit in four layers. Discovery asks whether the agent can find the site’s map at all — a parseable robots.txt and how it treats the major AI crawlers, a reachable XML sitemap, an llms.txt file, structured data on the homepage, a correct canonical URL. Access asks whether the agent can actually get in: the homepage must answer a plain HTTP client, must not refuse known AI user agents or throw a bot-challenge interstitial at them, must serve real content in its raw HTML rather than an empty app shell, redirect cleanly to HTTPS, and return honest status codes, including a true 404 for pages that do not exist.
The heavier two layers look past the front door. Usability — the largest layer, worth 40 of the 100 points — asks whether the agent can do anything once inside: is there an OpenAPI specification at the paths agents try, does it parse and describe real operations, are API docs discoverable, does an MCP endpoint answer an initialize probe, is there auth discovery metadata when the site has an API, are signup and contact forms reachable without a challenge and labelled well enough to fill in? Payments asks whether an agent could complete a purchase: an agentic-checkout discovery file for one of the emerging protocols (ACP, UCP, AP2 or x402) and a checkout page an agent can reach. Payments is deliberately conditional — it is only scored when the site shows commerce signals such as cart or checkout links, because a blog or a docs site should not lose points for lacking a checkout it never needed.
How the score works
Each layer carries a fixed share of the 100 points — discovery 20, access 30, usability 40, payments 10 — and each check inside a layer passes, warns, fails or is not applicable. A score of 70 or above earns the “agent-ready” label; the letter grades run A at 85, B at 70, C at 55, D at 40 and F below that. When a layer’s checks are all not applicable — payments on a site with no commerce signals is the common case — the layer drops out entirely and the remaining layers are rescaled to 100, so the score always compares like with like. Every finding names the points it costs: the “fix first” list simply sorts the failures and warnings by the points you would recover, so the order on the page is the order worth working in. And every finding is shown in full — the evidence gathered, what it means, and the fix — free, with no locked sections.
Related checks: the website health check for the classic uptime, certificate and DNS picture of the same domain, the SSL grader when the agent’s problem turns out to be the TLS setup, and the global ping when reachability differs by region. The companion guide, can AI agents use your website?, walks through the layers, how to read a score, and what a real scan of this site found.
Honest limits, stated the way the results deserve. The scan runs from our server in a US Azure region and makes at most about 45 fetches, so it sees the site from one network, on one continent, the way an agent hosted there would. A verdict of “not found” always means “not found at the paths an agent would try, from here, at this moment” — never a claim that a file exists nowhere. A site that behaves differently by geography, or that rate-limits unfamiliar clients, can score differently from another vantage point; re-check after changes rather than treating one scan as a verdict for all time. None of that is a reason to hide the findings: services that sell agent-readiness audits commonly show the score free and charge for the fixes. Here the fixes are the point, so the fixes are free.
The four layers and what each one asks
| Discovery — 20 points | Can an agent find the site and its map? A parseable robots.txt and its treatment of AI crawlers, a reachable XML sitemap, llms.txt, structured data and a self-referential canonical URL. |
| Access — 30 points | Can an agent get in? The homepage answers a plain HTTP client and AI user agents without a bot challenge, serves real content in raw HTML, redirects cleanly to HTTPS and returns a true 404 for missing pages. |
| Usability — 40 points | Can an agent act? A discoverable, parseable OpenAPI spec, API documentation, an MCP endpoint that answers a probe, auth discovery where an API exists, and forms an agent can find, read and submit. |
| Payments — 10 points | Can an agent pay? An agentic-checkout discovery file (ACP, UCP, AP2 or x402) and a checkout reachable without a challenge — scored only when the site shows commerce signals, otherwise not applicable. |
Frequently asked questions
What is an AI agent, in this sense?
Software that visits a website on someone’s behalf and tries to complete a task — research, a booking, a purchase, filling in a form — without a human driving a browser. Agents typically fetch raw HTML and machine-readable files (robots.txt, sitemaps, llms.txt, OpenAPI specs), may identify themselves with user agents such as GPTBot or ClaudeBot, and cannot solve visual challenges or run heavy JavaScript the way a browser does. A site can be excellent for people and nearly unusable for agents; this checker measures the second property.
What score counts as agent-ready?
70 out of 100 or above, which is also grade B or better (A starts at 85, C at 55, D at 40). The threshold is a convention of the agent-readiness category rather than a law of nature: at 70 a site is normally discoverable, reachable without challenges, and exposes enough machine-readable structure for an agent to work with. Read the findings, not just the number — two sites at 70 can be missing entirely different things.
The payments layer says “not applicable”. Is that bad?
No. Payments only applies to sites that show commerce signals — cart or checkout links, or a known shop platform. When it does not apply, the layer is removed from the calculation and the other three layers are rescaled to 100, so a blog, docs site or portfolio is never punished for lacking a checkout. The same applies to individual checks inside other layers: a pricing-page check is not applicable when no pricing page is linked, and a signup-form check is not applicable when there is no signup link.
Why does the scan see something different from my browser?
Because an agent’s visit is different. The scan uses plain HTTP requests, no JavaScript execution, and in several checks an AI crawler’s user agent, all from our server in a US Azure region. If your site serves an empty shell that only fills in after JavaScript runs, or shows a bot challenge to unfamiliar user agents, or behaves differently outside your country, the scan reports that agent’s-eye view — which is exactly the view that decides whether agents can use the site.
Are all the findings and fixes really free?
Yes. Every finding this checker produces — the evidence, what it means, the fix, and the points it is worth — is shown in full on this page, and the report link can be shared with anyone. Some agent-readiness services show a free score and charge per domain for the fix list; we think the fixes are the useful part, so there is no paywall here. The scan is rate-limited because each one costs up to about 45 fetches, but a completed scan is yours.