Can AI search read your page?

How to read your result

Five rows, one question each. A row passes, needs review, needs a fix, or couldn't be checked. That last state is deliberate: a fetch that fails tells you nothing, so it never shows as a pass and never as a failure.

• Crawler access. The site's robots.txt, evaluated for the page's path against every AI agent in the table below. One blocked retrieval agent is a fix; a blocked user-triggered agent is a review; blocked training agents are noted and never fail the row.
• llms.txt. Present or not, and whether it has the shape the convention describes: a title, a summary line, sections, and links with descriptions. Missing is neutral.
• Reachability. The page must answer 200 in at most one redirect, its canonical must point at itself, and it must not carry noindex, nosnippet or a tight max-snippet. Any of those on a page you want cited is a fix.
• Server-rendered text. How many words are in the raw HTML, fetched without running JavaScript. Under 120 words inside a large document is the signature of a JavaScript shell.
• Quotability. Seven mechanical dimensions, two points each: a 40 to 80 word lead answer right after the H1, headings phrased as questions, a date signal, JSON-LD schema, tables or lists, at least 300 words, and one H1 with a title. Below it, the three weakest passages, because an engine lifts a passage, not a page.

Retrieval, user-triggered, training

The agent list is grouped by what blocking each one costs, and getting this backwards is the most common mistake in the whole subject. Teams block retrieval agents thinking they are protecting content from training, and teams block training agents thinking they are opting out of AI answers. Neither does what they meant.

• Retrieval fetches at answer time, or builds the index the engine searches when answering. Blocking it removes you from that engine's answers.
• User-triggered fetches one page because a person asked the assistant to read it. Blocking it breaks "read this link for me" on your site.
• Training collects content to train future models. Blocking it is a legitimate decision either way and does not remove you from today's answers.

Agent names and roles verified against each vendor's published crawler documentation on 2026-09-11. Names change and vendors add new ones; an agent your robots.txt names that isn't in this list is reported as unrecognised, not ignored.

What the check can't see

Three things, said plainly so nobody quotes a pass as more than it is.

• Blocks that don't live in robots.txt. CDN bot rules, WAF rules, rate limits aggressive enough that a crawler backs off, and country blocks that happen to cover a crawler's egress. If your logs show zero hits from an agent this report lists as allowed, look there.
• Anything JavaScript adds. By design. The check reads the page the way a script-less agent does. If your content only exists after hydration, that is the finding, not a limitation of the tool.
• Substance. The quotability score is structural. It can't tell whether a passage says something no other page says, and that is what earns a citation. A high score on a page that restates the consensus is a well-formatted page nobody will quote.

A worked example

We ran the check on our own home page on 2026-09-11. Crawler access passed: robots.txt names no AI agents, so all 25 may fetch it. No llms.txt. Reachability passed: 200 in 141 ms, no redirects, canonical pointing at itself. Server-rendered text passed with 765 words in the raw HTML. Quotability scored 8 of 14: the first paragraph after the H1 is 51 words and stands alone, but the page carried no JSON-LD, no date signal and no heading phrased as a question. The three weakest passages all opened with a sentence that named nothing specific. That is the to-do list, in order, and the same one the check will hand you.