How to run an AI SEO audit

A conventional SEO audit still catches most of what matters, and it misses the four things that decide whether an assistant quotes you. This AI SEO audit is the delta: what to add, in what order, and how to tell a real problem from a crawler report entry nobody should act on. It uses Search Console, your server logs, a crawler and about an hour of reading. Google's documentation on AI features is the reference for the parts involving its own surfaces.

Step 1: can AI crawlers reach your pages?

Start in your server logs, not in a tool. Filter for GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot and Google-Extended over the last 30 days and answer three questions.

• Are they arriving at all? If not, check robots.txt first, then your firewall or CDN bot rules, which block far more AI agents by default than most teams realise.
• What status are they getting? A crawler receiving 403s is being blocked by infrastructure regardless of what robots.txt says. This is the single most common finding in this audit and it is invisible from outside.
• Which pages do they fetch most? That list tells you what the systems currently consider relevant about you, which is useful information you cannot get anywhere else.

Then check robots.txt deliberately rather than by inheritance. Our robots.txt generator lists all fifteen agents with what each one does, so the decision is explicit. Whichever way you decide, decide: a default template blocking OAI-SearchBot because someone copied it from a publisher is not a strategy.

Step 2: can they read what they fetch?

Reaching the page and reading it are different problems. Many AI crawlers do not execute JavaScript, so anything injected client-side is invisible to them even though it renders perfectly in your browser.

Test it properly: fetch your own page with JavaScript disabled, or use curl with a plain user agent, and read what comes back. Then check three specific things.

• Is the main body content in the server-rendered HTML? On many React and Vue sites it is not, and nobody notices because Googlebot renders JavaScript and the AI crawlers quietly do not.
• Are headings in the HTML as headings? Divs styled to look like headings carry no structure, and structure is what passage extraction relies on.
• Is anything important behind an interaction? Accordions, tabs and 'read more' toggles that load content on click are frequently absent from what a crawler receives.

Any crawler from our technical comparison will do this at scale, with rendering disabled, and the diff between the rendered and unrendered crawl is the report you want.

Step 3: can a passage stand alone?

Retrieval systems extract passages, not pages. Audit your top twenty pages by asking one question of each section: would this paragraph make sense pasted into an answer, with no other context?

The recurring failures are consistent across sites:

• The answer is buried under 300 words of preamble, so the first extractable passage is context rather than substance.
• Sections open with 'as we saw above', which makes them unusable in isolation.
• Headings are clever rather than descriptive, so a machine cannot tell what the section answers.
• Claims have no numbers, so there is nothing specific to quote.
• Numbers have no sources, so there is nothing to corroborate.

Our GEO content scorer automates most of this check and names the fix for each failure. Run it on your five highest-value pages before deciding whether the problem is site-wide.

Step 4: is your entity consistent?

Assistants assemble answers from many sources. Inconsistent descriptions of your company across those sources produce several weak claims instead of one strong one, and the strong claim usually belongs to a competitor.

Open your site, LinkedIn, Crunchbase, the review platforms your category uses, and the top three roundups that mention you. Then check:

• Is the company name identical? Including legal suffixes, capitalisation and spacing.
• Is the one-line description the same? Not similar. The same.
• Is the category the same? If you are an 'AI SEO platform' in one place and a 'content marketing tool' in another, you have told the system you are two different things.
• Is your pricing public? Assistants asked what something costs will name whoever published a number.
• Does Organization schema exist site-wide with sameAs links to every profile that corroborates you?

This is the slowest layer to fix and the one that most determines whether you get recommended rather than merely mentioned. Our diagnostic on why AI recommends competitors is essentially an expanded version of this step.

Step 5: what do the assistants actually say about you?

The measurement step, and the one people do first and should do last, because until steps one to four are sound the answer is predictable.

Build a panel of 20 to 50 questions your buyers genuinely ask, taken from sales calls and support tickets rather than a keyword tool. Run them across ChatGPT, Perplexity, Gemini and Copilot, and record three things separately for each: are you mentioned, are you cited with a link, are you recommended as the answer.

Then record the thing most tools bury: which sources got cited instead of you. That list is the most actionable output of the entire audit, because it names the specific pages standing between you and the recommendation. Some will be competitor pages you cannot influence. Many will be roundups, review platforms and community threads you can.

Do this manually the first time, even if you intend to buy a tracker. Reading the actual answers teaches you more about how your category is described than any dashboard, and it takes about ninety minutes. The paid options, from $20 a month, are in our tracker comparison.

How do you prioritise what the audit finds?

An audit that returns 200 findings and no order is worse than no audit, because it produces paralysis and a document nobody opens twice. Sort into three buckets.

Blocking means something prevents you appearing at all: crawlers blocked at the CDN, content not in the HTML, a stray noindex, a redirect chain on a key page. Fix this week, regardless of effort, because everything else is worthless until it is done.

Compounding means it improves your odds everywhere over time: passage structure, entity consistency, schema, third-party corroboration. Schedule over a quarter. This is where most of the actual value is and where most audits lose momentum.

Marginal means it is real but small: a missing alt attribute, an imperfect heading, a slightly long title. Batch it into whatever you are doing anyway. Do not let it consume the meeting.

The failure mode of technical audits is spending the quarter on bucket three because it is easy to tick off. The failure mode of AI SEO audits specifically is spending it on step five because it is interesting, while a CDN rule quietly blocks every crawler you were measuring.

Conclusion

Run the five steps in order and the audit takes a day. Access, rendering, passage structure, entity consistency, then measurement. The order matters more than the thoroughness: teams routinely discover, three months into an AI visibility programme, that their CDN was returning 403 to every AI crawler the whole time. Check the logs first. Everything else is content strategy, and content strategy applied to pages nobody can fetch is an expensive way to learn about infrastructure.

Start with the structure check Free SEO tools and calculators

Frequently asked questions

Sources

Every figure in this article traces back to one of these. We link them so you can check the original rather than take our summary of it.

• Google Search Central: AI features and your website (gov)
• OpenAI: bots and crawler user agents (gov)
• Google Search Central: overview of Google crawlers (gov)
• Google Search Console Help: performance report (gov)

Free tools for this

Everything below runs in your browser, with no signup and nothing uploaded.

Definitions: Content brief · Internal linking · Content cluster · Structured data

Related reading