AI agent readiness checker
Can AI agents find your site, read it and use it? A level from 0 to 5, and what each level still needs.
- Checks your page
- 3 browsers
- 17 rules
- Last updated
What we keep
The result appears here
Enter your page’s address and press “Check the page”. This is what the tool runs on it:
- robots.txt answers with an HTML pagenot started
- No sitemap for search engines to findnot started
- The page's response points agents to nothing the site offersnot started
- The site's DNS points agents to none of its servicesnot started
- The page gives agents no Markdownnot started
- The Markdown for agents drops the Arabicnot started
- llms.txt cannot serve as a guidenot started
- The page's content is written by scriptsnot started
- robots.txt blocks AI search crawlersnot started
- The site does not say how AI may use its contentnot started
- Content signals in robots.txt cannot be readnot started
- The site answers bots with a challengenot started
- The site publishes no file an agent can start fromnot started
- The API catalog cannot be usednot started
- A file for agents is published and cannot be readnot started
- The page's WebMCP tools cannot be used by an agentnot started
- WebMCP tools with undescribed parameters, or too manynot started
What it checks
- Access and discovery: your
robots.txtand sitemaps, theLinkheader that points agents to your APIs, and DNS records for agents (DNS-AID). - Content for machines: whether the page answers in Markdown to whoever asks, whether that Markdown keeps your Arabic text, your
llms.txt, and whether your text shows without JavaScript. - Crawler policy: whether
robots.txtkeeps AI search crawlers from your page, content signals (Content-Signal) that say how your content may be used, and whether your site answers bots with a challenge. - What you offer agents: an API catalog (RFC 9727), OAuth metadata, an MCP server card, an A2A agent card, a skills index, an agent catalog (ARD), and WebMCP tools on the page in three engines.
- Commerce (optional): agent payment and shopping protocols, x402, MPP, UCP, ACP and AP2, shown and never in a share.
Example
User-agent: * Content-Signal: search=yes, ai-trian=no Allow: /
User-agent: * Content-Signal: search=yes, ai-input=yes, ai-train=no Allow: /
How to fix
Start with the easiest, which takes you up one level after another:
- Presence: a text
robots.txtwith aUser-agentgroup, a sitemap it names, and aLinkheader to what you offer. - Policy: allow AI search crawlers, and add a Content-Signal line to the
*group (below the list). - Reading: answer
Accept: text/markdownwith a Markdown version of the page that keeps all of its Arabic text, withVary: Accept. - Interfaces: publish at least one that fits you: an MCP server card at
/.well-known/mcp/server-card.json, a skills index, or an API catalog at/.well-known/api-catalog. - Auth and identity: OAuth metadata (RFC 8414 and RFC 9728) and, if your site sends bots, a Web Bot Auth key directory.
The policy line in robots.txt:
In the result, each check you did not pass says what to add, where, and gives text to copy.
User-agent: *
Content-Signal: search=yes, ai-input=yes, ai-train=no
Allow: /FAQ
No. What matters most to a content site is presence, policy and reading: robots.txt, the sitemap, content signals and a Markdown version. MCP and A2A cards and OAuth metadata are for those who offer a service agents use. That is why most of these checks are information, which takes little from Arablyzer's score, and only when a file is published and broken.
The level is a ladder: you reach a level only when you pass everything below it. The share is how many of the counted checks you pass, without commerce and without the informational checks. We show both by our verdict and, beside them, as the public tool many compare with counts them, under "the reference's count".
For three reasons, given beside each check. We honour robots.txt: when it asks us not to fetch an address, we do not, and say "not checked". We judge AI search crawlers by the rule's effect: a site that blocks them does not pass with us, even when it names them. And we read WebMCP tools in three engines, not Chromium alone, and content signals in the response header too.
Only GET requests with no data, as ArablyzerBot: the page once as Markdown and about 25 fixed addresses, four at a time, none that robots.txt asks us not to fetch. We never call a WebMCP tool or fill a form. DNS questions go to a DNS resolver, not to your site.
What this tool runs
- robots.txt answers with an HTML pageminor
robots-txt-not-text - No sitemap for search engines to findminor
sitemap-missing - The page's response points agents to nothing the site offersinfo
agent-link-relations-missing - The site's DNS points agents to none of its servicesinfo
dns-aid-missing - The page gives agents no Markdowninfo
markdown-negotiation-missing - The Markdown for agents drops the Arabicminor
markdown-arabic-loss - llms.txt cannot serve as a guideminor
llms-txt-invalid - The page's content is written by scriptsmoderate
js-only-content - robots.txt blocks AI search crawlersmoderate
robots-blocks-ai-search - The site does not say how AI may use its contentinfo
content-signals-missing - Content signals in robots.txt cannot be readminor
content-signals-invalid - The site answers bots with a challengeinfo
bot-challenge - The site publishes no file an agent can start frominfo
agent-discovery-missing - The API catalog cannot be usedminor
api-catalog-invalid - A file for agents is published and cannot be readminor
agent-manifest-invalid - The page's WebMCP tools cannot be used by an agentminor
webmcp-tools-invalid - WebMCP tools with undescribed parameters, or too manyinfo
webmcp-tools-unclear
How we measure
We fetch robots.txt, then the page, as ArablyzerBot, follow redirects, and check the final page's origin. We ask for each agent file with an Accept that fits it, and none with an Accept that names Markdown, since some services convert any address to Markdown and answer for an address that does not exist. With each type we also ask for an address no site has: when the site answers it with 2xx, it answers every address, and only a file that passes its content test counts as published. We render the page in Chromium, Firefox and WebKit, with a recorder that sees WebMCP tools before the page's scripts run, and apply Lighthouse 13.6's errors and warnings to them. We ask DNS over HTTPS with the DNSSEC bit. We compute the level and share with the ladder and sums the public tool uses, which matched its results in all seven scans we saw, and show our verdict beside its count.