AI Readiness Check

AI readiness is the set of technical conditions an answer engine needs to reach your site, read your pages and attribute what it read to you. ChatGPT's search, Claude, Perplexity and Google's AI Overviews read web pages when they answer a question. A page they cannot read, they cannot use as a source.

Each of the three steps can break in its own place. Access closes in robots.txt rules or in the CDN in front of the server. Reading fails when content only exists after JavaScript runs, or when an address returns an error shell instead of the page. Attribution weakens when a page does not say plainly whose it is: organisation name, logo, author and date.

Readiness is a precondition, not an outcome. A site with an open door may still not be named; a site with a closed one is never among the pages that engine reads. This check answers the first question, not the second.

AI companies do not come to your site with a single bot. By their own documentation most have three jobs: a bot that collects pages for a search index (OAI-SearchBot at OpenAI, Claude-SearchBot at Anthropic, PerplexityBot at Perplexity), a bot that opens a page the moment a user asks (ChatGPT-User, Claude-User, Perplexity-User), and a bot that collects content to train models (GPTBot, ClaudeBot). Google's AI Overviews use the ordinary Google Search index.

They all need the same things: no block against their name in robots.txt, a server that returns the real page to a request carrying that name, and content present in the first HTML response. OpenAI documents that sites which disallow OAI-SearchBot will not be shown in ChatGPT's search answers. Blocking training bots is a data policy decision and not the same thing as search visibility, which is why our scan does not count it as a defect.

One more detail for user bots: OpenAI and Perplexity write that robots.txt rules may not apply to these user-started requests. So robots.txt on its own is not an access list. Knowing which bot comes for what lets you make the decision in the right place; the bot list and example rules are in our AI bot access guide.

AI Bot Access Guide

According to Vercel's measurement published in December 2024, none of the major AI crawlers, OpenAI's and Anthropic's included, execute JavaScript: they may download JavaScript files but they do not run them. The exceptions in the same measurement were Gemini, which uses Google's infrastructure, and Applebot.

In practice, a page whose content is built in the browser by JavaScript looks complete to a human visitor and sends these bots an empty shell. The title, the meta tags and even the structured data may be there; the product description, the price or the article text are not. Such a site can pass most other checks and still not be read.

The fix is to build the main content on the server, or to prerender pages and serve them as HTML. To check for yourself, view the page source in your browser (not the live view in the developer tools) and look for your main text there. Our scan reads the page without running JavaScript and flags it as a critical item when very little text is readable.

The score is the result of the checks that could run, and it always says how many checks it rests on. A check is in one of three states: passed, to fix, or not measured. A check that could not run (say robots.txt did not answer in time) is not counted as a defect and does not enter the score at all, because we do not know what is true.

The same distinction holds for engine answers. If the engine did not name you, that is a finding. If we could not reach the engine or the daily free allowance ran out, that is our limit, and the page says it could not be measured; it does not say you were not named. When your name does come up we ask a second question: is the answer really about you, or about another company with the same name? When we cannot confirm it, we say unconfirmed.

A high score says the plumbing is sound. It says nothing about the quality of your writing or about any engine naming you. When the score is high and the engine still does not recognise you, the result page states that contradiction separately.

The two overlap a great deal. Google writes that there are no additional technical requirements to appear in AI Overviews and AI Mode: the page must be indexed and eligible to be shown with a preview in Search. Title, description, canonical tag and sitemap help in both worlds.

Where they part is where most classic SEO audits do not look: AI bots other than Google's being addressed by name in robots.txt, CDNs that can stop those bots by default (Cloudflare announced in July 2025 that it was changing its default in that direction) and crawlers that do not run JavaScript. A page that ranks well on Google can be unreadable to ChatGPT or Claude.

So our scan includes some SEO checks, but its score is not an SEO score. It does not measure rankings, keywords or link profiles.

A check that could not run is not a failure. If your robots.txt times out, we do not know whether GPTBot is allowed, so we say we do not know, and leave that check out of the score entirely. The alternative is a number that drops for reasons you cannot act on, then silently recovers on a re-run.

A 200 is not proof. A single-page app answers 200 for every path, so a sitemap that “exists” may be your homepage HTML wearing a different URL. We read the body, not the status code.

Structured data is read the way engines read it. schema.org markup is commonly wrapped in an @graph container, where the root object carries no @type of its own. That is correct markup and the pattern most SEO plugins emit. A checker that only inspects the root object calls it broken and sends you to damage working markup.

A deliberate block is not a defect. Plenty of sites block a particular crawler on purpose. Where we find one, we state what it costs you and leave the decision where it belongs.