AIO · agent readability
AI Overviews, ChatGPT and Perplexity fetch a page with no browser, no cursor and no login. If the first response is an empty shell, a robots group says no, or there is no llms.txt, the model has nothing to quote. These are the ten checks that decide it.
Free sample report · no signup · about 90 seconds
GET / HTTP/1.1
user-agent GPTBot
accept text/markdown
status 200 · text/html
ttfb 512 ms
llms.txt 404
robots GPTBot disallowed
body <div id="root"></div>
The ten checks
Ten checks in the order an agent meets them. Each one is decided from a raw HTTP response, never from a rendered DOM, so the result can be reproduced and costs no browser.
Per site: 10 checks, a 25 s crawl budget, a 400 ms TTFB budget, 0 browsers.
The AI user-agent groups, read with RFC 9309 semantics: a named group replaces the wildcard instead of inheriting it.
Fail One Disallow line in the GPTBot group removes the whole site from that crawler, and every later check becomes moot.
A markdown index at the root with an H1, at least one link and 50 characters or more of body. An HTML 200 counts as a soft 404.
Fail An agent summarising the site works from scraped fragments instead of the index you meant to publish.
Whether the primary copy is in the first response rather than injected after hydration. An empty root element counts as empty.
Fail A crawler that does not execute JavaScript sees a frame with no text, so there is nothing to quote.
What an Accept: text/markdown request returns, and whether Vary carries the header so caches do not mix the two.
Fail Agents that ask for markdown get full markup, spend their context on tags and truncate the part that mattered.
Organization, Product and FAQPage parse, carry sameAs, and the FAQ answers match the visible text word for word.
Fail The model cannot confirm who you are or which facts are firm, so it hedges or cites a source that will.
One h1, ordered h2 and h3, none used only for styling, and a heading tree that matches the answer the page gives.
Fail The extractor cannot tell which sentence answers the question, so it lifts a neighbouring paragraph instead.
A self-referencing canonical on every URL, no redirect chains, and no parameter duplicates reaching the same page.
Fail Two addresses for one page split the evidence, and the engine keeps whichever one it fetched first.
The sitemap is reachable, lastmod is inside seven days, and every listed URL returns 200.
Fail Stale entries send the crawler into 404s, and the freshness signal the sitemap should carry is spent.
Every money page sits within three clicks of the home page, and the orphan list stays short.
Fail A page nothing links to is a page nothing fetches, and a page nothing fetches is a page nothing quotes.
Time to first byte from a fixed region against a 400 ms budget, measured on the HTML response only.
Fail A slow first byte is a truncated fetch: the agent times out and keeps the empty result.
A check that cannot run is marked skip and left out of the score. We do not write a zero for a page we could not reach.
Why this comes first
AIO is not a second SEO. It is the earlier half of the same question: can this page be reached, parsed and quoted at all.
01
A page that is never fetched cannot be cited.
Crawl access, a fast first byte and server-rendered text decide whether the rest of your work is visible to the engine at all. Optimising a blocked page is invisible work.
Shared ground · crawl, canonical, sitemap
02
Ranking is not quoting.
Answer engines lift sentences. Clear headings, short factual lines and entity schema give the extractor something to lift; a page that ranks but reads as one long block gets summarised, not cited.
Added for agents · llms.txt, markdown, schema
03
A buyer agent may call your site, not read it.
When an agent asks for prices, availability or docs, it needs machine-readable facts and a first byte it will wait for. A funnel built for a human eye returns nothing useful to it.
Not a replacement · SEO still sets the shortlist
What you get
Every failing check ships with the file, header or schema block that resolves it, and the path to place it at. We do not deploy to your infrastructure unless you ask us to.
A markdown index of canonical pages, prices and entity facts, written in the order an agent asks for them.
The AI user-agent groups you are missing, with a Content-Signal line, as a diff against the file you serve today.
Organization, Product and FAQPage, with sameAs filled in and answers that already match the visible text.
Handover is a folder of files, a robots snippet and a schema block, with a one-page change list. You commit them, or we apply them in an engagement where you have given us repository or server access. Nothing is edited on your servers without that access.
Questions
No. Readability is a precondition, not a promise: engines also weigh sources, recency and consensus, and none of those are fixed by a crawl. What the checks remove is the set of reasons a page cannot be fetched, parsed or quoted.
No, it extends it. Crawl access, canonicals, sitemaps, headings and TTFB are shared ground. AIO adds the artifacts agents look for: robots AI groups, llms.txt, markdown negotiation, entity schema and the discovery files under /.well-known.
Not by default. Fixes are delivered as a patch: files, a robots snippet and a schema block, each with the path to place it at. You commit them, or we apply them in an engagement where you have given us repository or server access.
Free · no signup