GEO Audit: How to Audit Your AI Visibility in 9 Steps

You want to know whether ChatGPT, Gemini, or Perplexity cite your website — and if not, why. A GEO audit answers both questions. It evaluates your site against the criteria AI engines use to select their sources, identifies blockers, and gives you a prioritized action plan.

This article walks you through a complete GEO audit, step by step, that you can run yourself or accelerate with an automated tool.

What a GEO audit evaluates (and what it does not)

A GEO audit is not an SEO audit. The two complement each other but measure different things.

What a GEO audit measures:

• Can your content be extracted and cited by an AI? (citability)
• Can AI bots access your website? (AI accessibility)
• Is your site understood by machines? (structured data)
• Is your content verifiable? (figures, named sources)
• Do you exist outside your own website? (external presence)
• Is your content up to date? (freshness)

What a GEO audit does not measure:

• Your Google rankings (that is SEO)
• Your page speed (that is technical SEO)
• Your backlink profile in detail (that is off-page SEO)
• The exact number of times AI engines cite you (no tool can measure this with certainty)

A GEO audit measures your citation potential — the technical and editorial conditions that maximize your chances of being selected as a source by AI engines.

→The 6 Detekia score criteria and how they are measured

Step 1 — The citation test (15 minutes)

Before diving into the technical side, start with the most concrete test: ask AI engines directly whether they know your website.

How to do it

Open ChatGPT, Gemini, and Perplexity. For each, ask 5 questions:

Recommendation questions (your market):

1. "What's the best [your service/product] in [your city]?"
2. "Recommend a [your service/product] for [specific need your clients have]"

Expertise questions (your domain):

3. "How do I choose a good [your industry]?"
4. "What's the difference between [A] and [B] in [your sector]?"

Brand question (your name):

5. "What does [your company name] do?"

How to analyze the results

For each response, record in a spreadsheet:

• Cited? — your site or brand appears in the response (yes/no)
• Competitors cited — which names appear instead of yours
• Type of source cited — official website, blog post, forum, directory?

Calculate your presence rate: number of citations / (5 questions × 3 engines) = X%.

Below 20%, you have a significant AI visibility problem. Above 50%, you are in a good position. Above 80%, you can shift to fine-tuning mode.

What the results tell you

If you appear nowhere but your competitors are cited, the problem is likely technical (AI accessibility, structured data) or editorial (citability, content quality).

If nobody appears in your sector, that is an opportunity: the first to optimize will capture the citations.

If you appear on some queries but not others, compare the corresponding pages: the one that gets cited probably has something the others lack (better structure, more data, FAQ...).

For the brand question, go further: our guide explains how to check methodically what ChatGPT says about your brand (neutral session, repetitions, recording sources). Also keep in mind that 15 answers are a small sample, and that answers vary from one run to the next. To measure your share of voice across all the queries in your market, in ChatGPT, Gemini, Claude and Perplexity, you need an AI visibility audit such as the one run by Beeleven, the agency that built Detekia.

Step 2 — The AI accessibility audit (10 minutes)

This is the binary criterion: if AI bots cannot access your site, nothing else matters.

Check robots.txt

Go to your-website.com/robots.txt and look for directives targeting these user-agents:

• OAI-SearchBot — ChatGPT web search (blocking it removes you from answers with sources)
• GPTBot — training of OpenAI models
• ClaudeBot — Claude
• PerplexityBot — Perplexity
• Google-Extended — use of your content to train Gemini and ground its answers (it has no effect on your presence in Google Search, which includes AI Overviews)

Details for each crawler are in OpenAI's documentation and Google's.

If you see Disallow: / for any of these bots, you are actively blocking them.

Expected result: no AI bot should be blocked (unless it is a documented strategic decision).

Check the llms.txt file

Go to your-website.com/llms.txt. This file, if it exists, guides AI crawlers to your most important content.

Expected result: the file exists and lists your main pages.

→llms.txt, robots.txt and AI accessibility: the technical guide

Check JavaScript accessibility

Disable JavaScript in your browser (Chrome DevTools → Ctrl+Shift+P → "Disable JavaScript") and navigate your site. If the main content disappears, AI crawlers probably cannot read it either.

Expected result: main content is visible without JavaScript.

Check the sitemap

Go to your-website.com/sitemap.xml. Verify that it is accessible, lists all your important pages, and that <lastmod> dates are up to date.

Expected result: sitemap accessible and up to date.

Step 3 — The citability audit (30 minutes)

This is what the heaviest criterion of the GEO score measures, Answer structure (40 points out of 100), and it is the one that fails most often.

Analyze your 5 main pages

Take your 5 most important pages (homepage, main service/product page, key blog post, FAQ, About page). For each page, evaluate:

The first 100 words — do they contain a direct answer to the question the visitor is asking? Or do they start with vague marketing copy ("Welcome to...", "Leading innovation...")?

Test with this method: copy the first 100 words of your page and paste them into ChatGPT, asking "Based on this text, what does this company do and what does it offer?" If ChatGPT cannot answer clearly, your intro is not extractable.

H2/H3 subheadings — are they phrased like your customers' questions? "How much does our service cost?" is good. "Learn more" is bad. "Our advantages" is mediocre. Under each heading, a short answer (one to three sentences, 15 to 70 words) should come before the details: in our calibration study of September 30, 2026 on 317 pages, this is the signal that most clearly separates cited pages from non-cited pages on the same site.

Paragraph autonomy — pick a random paragraph from the page. Is it understandable out of context? AI engines often extract a single paragraph to answer a question. If that paragraph requires the previous paragraph for context, it loses its value.

Structure — are comparisons, prices and features laid out in tables? Tables are among the signals of cited pages in the study. Structured information is more easily extractable than continuous text.

Quick scoring

For each page, assign a subjective score:

• Good (4-5/5): factual intro, descriptive subheadings, autonomous paragraphs, lists/tables
• Average (2-3/5): mix of informative and commercial content, partial structure
• Weak (0-1/5): commercial intro, no structure, vague continuous text

Step 4 — The structured data audit (15 minutes)

Test with the Rich Results Test

Go to Google's Rich Results Test and enter your homepage URL. Note:

• How many rich result types are detected?
• Are there any errors or warnings?

Repeat for your FAQ page and your main blog post.

Essential schema checklist

Check for these schemas:

• Organization on the homepage — name, logo, URL, address, contact, sameAs (links to LinkedIn, etc.)
• Article on editorial content — title, author (with name and LinkedIn URL), dates, description
• FAQPage on pages containing Q&As
• BreadcrumbList on internal pages
• WebSite with SearchAction on the homepage

Expected result: Article markup with accurate dates (dateModified) and, if the page has a visible FAQ, FAQPage markup with the same questions. The rest is useful but not decisive: Google states that no special markup is required to appear in its AI features, and in our study, Organization schema did not separate cited pages. Do not spend more time on it than on the structure of your answers.

→Schema.org and AI: ready-to-copy code examples

Step 5 — The verifiability audit (20 minutes)

Scan content for evidence

Go through your 5 main pages and count, for each page:

• Quantified data — statistics, percentages, dollar amounts, time frames (target: 1 data point per 200 words)
• Named sources — "according to Gartner", "per a McKinsey study" (not "according to experts" or "studies show")
• Reference dates — "in 2026", "since March 2025" (not "recently" or "for some time now")
• Concrete examples — client case studies, usage scenarios, measured results

Identify unsourced claims

Look through your pages for claims that are not backed by evidence:

• "We are leaders in..." — leaders by what metric?
• "Exceptional results" — which results exactly?
• "Many satisfied customers" — how many exactly?

Every unsourced claim is a negative signal for AI engines.

Step 6 — Trust signals: useful, not decisive (5 minutes)

About page, named authors, social links, legal notice: they are often presented as citation levers, in the name of E-E-A-T. Our calibration study does not confirm it: named authors, Organization schema, Contact, About and legal pages, sameAs links and expert quotes did not separate cited pages from non-cited pages on the same site. Google itself notes that E-E-A-T is not a specific ranking factor.

Two caveats. The study compares pages within the same site: it shows these signals do not separate two pages, not that they are useless to a site. And they still matter to your readers, some being legally required.

So check quickly:

• A complete legal notice: company name, registration number, address, publication director (required in France)
• Signed articles, where readers expect to know who is writing
• An About page that clearly says who you are and what you do

Expected result: complete legal information and an identifiable author. Beyond that, keep your time for the structure of your answers and your presence on third-party sites.

Step 7 — The comparisons and claims audit (10 minutes)

You often read that AI engines discard promotional content and that superlatives should be banned. Our calibration study says the opposite: "best…" comparison pages are among the most cited pages, and the former neutrality criterion of the Detekia score penalized exactly those pages. It has been removed.

What matters is not tone but structure and proof. Reread your key pages and look for:

• Comparisons without explicit criteria: a good comparison answers the question asked ("what is the best option for…?") with named criteria and, often, a table
• Claims without proof: "market leader", "exceptional results" (leader by what measure? which result?)
• Pressure tactics (artificial urgency, fear) that carry no information

The goal is not to remove the commercial dimension, but to back it with facts: "Solution trusted by 312 companies since 2019" rather than "Market-leading solution". Step 5 (verifiability) measures this too.

Step 8 — The external presence audit (15 minutes)

Check your web footprint

Search for your brand on Google (in quotes): "your company". How many results come from third-party sites (not your own)?

Also search for:

• Your name on Reddit (via Google: site:reddit.com "your company")
• Your company LinkedIn profile — is it active?
• Your Google Business Profile — does it exist and is it verified?
• Press mentions — articles that talk about you?

Evaluate diversity

Ideally, you should be mentioned on several types of sources:

• Industry press/media
• Forums/Reddit
• Professional directories
• Google Business Profile
• LinkedIn (active company page)

Expected result: at least 3 different types of external sources.

Step 9 — The freshness audit (5 minutes)

Check the dates:

• When was your last blog post published?
• Do your main pages display a visible last-updated date?
• Are there outdated dates in your content ("2024 trends", "2023 new releases")?
• Are the dateModified fields in your Article schemas up to date?

Expected result: content updated within the last 3 months, no outdated dates.

Synthesize the results

How to prioritize

Prioritize blocking criteria first — AI accessibility and citability (a short answer under each heading, question headings, tables). If bots cannot access your site or if your content is not extractable, nothing else matters.

Next, focus on high-impact criteria — freshness, content depth, verifiability and external presence. In our study, 79 to 89% of the sources cited by AI engines were third-party sites: your presence beyond your own site matters as much as your pages.

Finally, hygiene: structured data and trust signals (author, About page, legal notice). They do not separate pages, but a serious site keeps them in order.

DIY audit or professional audit: what changes

The nine steps above are enough to spot a site's technical and editorial blockers. What they do not tell you is where your brand actually stands in AI answers compared with your competitors. That is where a professional AI visibility audit changes scale, on four points:

• Query volume. A manual test covers 10 to 15 questions. A professional audit covers hundreds, grouped by topic and intent (Beeleven's audit covers at least 100 queries, up to 2,000), which shows the topics where you are missing.
• All four AI engines. A manual test often stops at ChatGPT. Yet each engine has its own source preferences: the audit asks ChatGPT, Gemini, Perplexity and Claude the same questions, so you can compare.
• Repetitions. AI answers vary from one run to the next. A single answer is an anecdote; only many questions asked systematically give a mention rate and an average position you can compare over time.
• Cited sources. The audit lists the pages AI engines cite instead of you (media, comparison sites, forums, competitors). That is your list of sources to target, which a manual test only glimpses.

The DIY audit remains the right free starting point to fix the basics. When the question becomes "where does our brand stand against our competitors, across our whole market?", you need a different tool: that is what Beeleven's AI visibility audit is for.

Automate the measurement

The manual audit we just described takes about 2 hours. It is an excellent exercise for understanding the stakes, but it is not sustainable for regular monitoring.

Detekia automates the measurement for free on its 6 criteria, drawing on the latest AI visibility measurement tools and methods: it scrapes the actual DOM of your page, analyzes each criterion, and gives you a score out of 100 with recommendations prioritized by impact. It also includes a first citation test: 5 queries sent to an OpenAI model to see whether your site comes up.

The automated score is especially useful for:

• Initial diagnosis — where do you stand exactly?
• Post-optimization tracking — did your fixes improve the score?
• Page comparison — also test a competitor's equivalent page to see what it lacks, or what yours lacks.

The Detekia score measures a page's citation potential, not your actual presence in answers. For that, Beeleven can identify the competitors and sources AI engines cite instead of you on ChatGPT, Gemini, Claude and Perplexity. And if you'd rather hand the audit to a provider, here are the criteria to choose a GEO agency.

→GEO: the complete guide to being cited by AI in 2026

→SEO vs GEO: what are the differences in 2026?

→Get your free GEO score