Home / Generative engine optimization / GEO Audit Checklist: How to Audit a Site for AI Search
GEO checklist

GEO Audit Checklist: How to Audit a Site for AI Search

A GEO audit checks whether AI tools like ChatGPT, Perplexity and Google’s AI Overviews can reach your pages, read them, understand who you are, and find an answer worth quoting. This is the checklist we work through, in order, on every GEO engagement we take on. Work down it and you will end up with a real list of problems. Order matters. If the bots are blocked at step one, nothing below it counts. For the basics first, read what generative engine optimization is.

Step 01

Crawler access and robots.txt

Start here. Crawlers are the bots that visit and read your pages. If they cannot fetch a page, nothing below matters.

  • What to check

    Which bots (called user agents) your robots.txt file allows, and whether AI crawlers are on that list. Look for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended, Bingbot and CCBot.

  • How to check it

    Read yoursite.com/robots.txt line by line. Then check your server or CDN logs to confirm those bots are actually showing up. An open robots.txt means nothing if a firewall blocks them before they reach the site. Perplexity publishes firewall settings and IP lists for its bots, covered in our Perplexity SEO guide.

  • What a fail looks like

    A sitewide block nobody knew about, an AI crawler blocked by a plugin’s default setting, or a robots.txt that allows everything while the CDN returns 403.

Letting AI crawlers in is a business decision, not a default. Some publishers block them on purpose, because their content is what they sell. But you cannot block a tool and expect it to cite you. Decide crawler by crawler and write down why. Also note that Google-Extended is a separate setting from Googlebot, so blocking one does not block the other.

Step 02

Rendering: does the content exist without JavaScript?

  • What to check

    Whether your text, headings and internal links are in the raw HTML the server sends, or only show up after JavaScript runs in the browser.

  • How to check it

    Use view source instead of inspecting the page in the browser’s DOM view, or fetch the page with curl. Compare that to what the browser shows. Test one page of each type: home, service, article, location.

  • What a fail looks like

    A nearly empty div and a script tag. Text you can see in the browser but not in the source. Many AI crawlers run less JavaScript than Googlebot, so they hit this problem first.

Step 03

Page structure and heading order

  • What to check

    One H1 (the main page title) per page, headings in order with no skipped levels, and headings that describe what sits under them.

  • How to check it

    Use a heading-outline browser extension or a site crawler that pulls out headings. Read the outline without the body text and see whether it sums up the page.

  • What a fail looks like

    Multiple H1s from a slider. H3s with no parent heading. Headings that are really just styled divs. Headings like ‘Our Approach’ that say nothing about the section.

Step 04

Can your answers stand on their own and be quoted?

This is the item most sites fail. It is the difference between a page that ranks and a page that gets quoted.

  • What to check

    Whether each section answers its heading in the first sentence or two, in words that still make sense when pulled out and dropped into someone else’s paragraph.

  • How to check it

    Copy the first two sentences of a section and read them cold. Do they answer the question without the setup above them? Watch for words like ‘it’ or ‘this’ and phrases like ‘as mentioned above’, which stop making sense once the passage is pulled out.

  • What a fail looks like

    Answers that show up in paragraph four. Sections that build up to a conclusion instead of stating it. Text that only works if you read it in order.

For how this plays out across featured snippets, voice and AI answers, see our guide to answer engine optimization.

Step 05

Clear business identity

  • What to check

    That your business name, location, service description and contact details read the same everywhere you appear, and that a reader could tell you apart from a business with a similar name.

  • How to check it

    Search your brand name and read the top twenty results. Compare your site, Google Business Profile, LinkedIn, directories and review sites, and note every difference in name, location or description.

  • What a fail looks like

    Three versions of your name across profiles. One city on the site and another on a directory. A description that changes with every listing. When things are unclear, an AI model hedges or names a competitor instead.

Step 06

Structured data

  • What to check

    Whether pages carry valid JSON-LD, the behind-the-scenes code that tells machines what a page is about, and whether it matches what is on the page: Organization, LocalBusiness where it applies, Article or Service on the right page types, sameAs links pointing to your real profiles, and FAQPage only where a real FAQ exists.

  • How to check it

    Run pages through Google’s Rich Results Test and the Schema.org validator. Then read the JSON-LD next to the page and confirm a human can see every claim it makes.

  • What a fail looks like

    Code describing content that is not on the page. Two conflicting Organization blocks, one from the theme and one from a plugin. sameAs links pointing to dead profiles.

Structured data is not a confirmed ranking factor in any AI answer tool, and nobody should tell you it is. It is a cheap way to state facts a machine would otherwise have to guess. Our guide to schema for AI search covers which types to use. The same caution goes for llms.txt, a new file some sites use to describe themselves to AI models. It is not a standard, and no major AI platform has committed to reading it. It is cheap to publish, but do not count on it to move results. Our llms.txt guide covers what goes in one.

Step 07

Author credentials and E-E-A-T

  • What to check

    Whether important pages have a named author with credentials anyone can verify, a bio on your site, and a presence that exists beyond your site. Google calls this E-E-A-T, short for experience, expertise, authority and trust.

  • How to check it

    Pick five content pages. For each, ask who wrote it, what qualifies them, and whether a stranger could confirm it with one search. Make sure the code and the byline agree.

  • What a fail looks like

    Posts signed ‘Admin’. A bio with no credential anyone can check. An author who shows up nowhere outside your site. In regulated fields, claims about results with no clinician or attorney named.

Step 08

Freshness and dates

  • What to check

    Whether published and updated dates are visible, accurate and match the structured data, and whether pages that go out of date have been reviewed.

  • How to check it

    Compare the date shown on the page, the dateModified field in the JSON-LD and the last time the content really changed. Spot-check pages that mention prices, laws or how a platform works.

  • What a fail looks like

    A script that stamps today’s date on every page. A 2019 article about a platform that has changed twice since. A dateModified that moves when nothing changed, which hurts trust instead of showing freshness.

Step 09

What other sites say about you

  • What to check

    Whether trusted outside sources describe you accurately, in the terms you want to be known for.

  • How to check it

    List the places an AI model is likely to pull from for your field: industry publications, association directories, review sites, local listings and press coverage. Check what each one says about you.

  • What a fail looks like

    You only exist on your own website. Listings that describe an old version of the business. Review profiles you have never claimed.

Step 10

The spot-check: are you actually being cited?

Everything above is a stand-in. This item measures the real result directly, and it is done by hand.

Build a set of 20 to 30 real buyer questions. Pull them from what clients ask on calls, your sales inbox and your Search Console query report, not from a keyword tool. Cover four kinds:

  • Category questions

    How a buyer describes the service before they know the industry terms, plus your location if you serve one.

  • Problem questions

    The problem they type in before they know what the fix is called.

  • Comparison questions

    Two approaches, two kinds of providers, or you against the alternative.

  • Branded questions

    Your name alone, your name plus reviews, and your name against a competitor.

Run every question through the tools your buyers use. That is usually ChatGPT, Google AI Overviews and Perplexity, plus Copilot in markets that run on Microsoft. For Google, Search Console now reports AI Overview impressions too, covered in how to rank in AI Overviews. Record four things each time: the question, whether you were mentioned, which sites were cited, and which competitor showed up. For ChatGPT, links in its search answers carry a utm_source=chatgpt.com tag, covered in our guide to ranking in ChatGPT.

Two rules keep this honest. Run each question at least twice in separate sessions, because these tools give different answers each time and one run proves nothing. And run them logged out when the tool allows it, so your chat history is not feeding your own site back to you. Repeat every month and keep the results in one sheet, so you can see real change instead of guessing.

Pair this with whatever your analytics can see. The method is in our guide to tracking AI search traffic in GA4, including why AI tools only show up partly in referral reports.

11 Questions

Common questions

How long does a GEO audit take?

Items one through nine take a day or two of focused work on a typical site, and longer on a large or multi-location one. Item ten takes a few hours the first time, then under an hour each time you repeat it. Fixing what you find is the real timeline. For what that costs, see how much GEO costs.

Do I have to allow AI crawlers?

No, and some publishers have good reasons not to. Just decide on purpose, crawler by crawler, and write down why, so nobody undoes it by accident.

Is any of this a confirmed ranking factor in ChatGPT or AI Overviews?

No. These tools do not publish how they pick sources, so every item here is based on how they seem to find and cite pages. What each item does reliably is remove a reason for an AI model to skip you. That is the part you control.

We ran the checklist and found twenty problems. What first?

Fix them in list order. Crawler access and rendering decide whether anything else gets seen, so they come first no matter how small they look. Quotable answers and a clear identity come next, because they decide whether the content you already have can be used.

Four audit findings: zero AI citations, 81 percent duplicate titles, a mobile speed score of 27 and insecure links.
Real findings from recent audits. Yours will be specific to your site.

See where you stand today

Want something to print and work through offline? The GEO readiness checklist is the downloadable version, cut down to the checks that catch the most problems.

Get a free digital audit

Working with businesses across the US