Key Takeaways
- A standard GEO audit checks your own pages and returns a score, but your pages are the minority of the surface that decides whether an AI engine cites you.
- Audit three layers in order: retrieval substrate (are you indexed and chunk-legible), off-site consensus (does the web corroborate you), and measurement (are you tracking the right signal).
- Google's own docs say there is no special schema and no llms.txt requirement to appear in AI Overviews, so a schema checklist is the smallest lever in the audit.
- A citation is not a passing grade: 61.7% of AI citations never name your brand, and self-cited 'best of' pages are left out of the recommendation 69% of the time.
- The output of a real GEO audit is a baselined prompt set you re-run, not a one-time score.
What a GEO audit is actually auditing
A GEO audit checks whether AI answer engines find, trust, and cite your content when someone asks about your category. The version you will find on the first page of search results does this by running a checklist over your own website: robots.txt rules for AI crawlers, schema markup, question-shaped headings, a direct-answer paragraph under each H2. It ends with a score.
That audit measures the smallest part of the problem. Your own pages are one input into the citation decision, and several large 2026 datasets put them in the minority. The thing a useful GEO audit inspects is the whole surface where the decision gets made, and that surface has three layers: whether the engine can retrieve you at all, whether the rest of the web corroborates you, and whether you are measuring the right outcome. Audit those three in order and you find the actual reason you are missing from AI answers. Run the page checklist alone and you polish the one layer that was probably already fine.
This matters because the score most audits hand back is not a grade you can act on. Being cited is not the same as being named, and being named is not the same as being recommended. An audit that counts citations and stops is auditing a number that does not move pipeline.
Layer 1: the retrieval substrate
Before anything else, an engine has to be able to pull your page into the context it writes the answer from. If it cannot, no amount of schema or formatting matters. This is the layer the checklists are closest to getting right, and also the layer where they add the most busywork.
Start with the part that is genuinely binding: are you indexed and crawlable. Google's own guidance for AI Overviews and AI Mode is blunt about eligibility. To show up in AI Overviews or AI Mode, a page has to be indexed and eligible to appear in Search with a snippet, and that is the requirement. The same document says you do not need to create an llms.txt file or any AI-specific markup, and that there is no special schema.org structured data you need to add. So when an audit spends a page grading your FAQ schema and flagging a missing llms.txt, it is grading things Google has stated it does not use. Check that AI crawlers are not blocked, check that your pages return a 200 and render their content server-side, and move on.
The on-page check that does earn its place is passage legibility. AI engines retrieve and quote in chunks, not whole pages, so the unit that gets cited is a self-contained passage that answers one question without the three paragraphs of context above it. Audit your priority pages by reading the first block under each H2 in isolation. If it states a complete, factual answer on its own, it is retrievable. If it only makes sense after a windup, it fails, and rewriting that block does more for citation than any markup change. This is the part of the page audit worth the hour. The rest of the technical checklist is table stakes you confirm once.
Joseph spent seven years leading marketing at an Inc. 5000 company, and the recurring lesson from auditing content there was the same: teams polish the parts that are easy to see and skip the part that actually gates the outcome. The schema report looks productive. The passage rewrite is the work.
Layer 2: off-site consensus
This is the layer most GEO audits skip entirely, and for a lot of brands it is where the citation is actually won or lost. AI engines decide who is credible by reading the whole web rather than your domain alone, and the signal they lean on is how often and how consistently other sites talk about you.
The size of the effect is hard to ignore. Ahrefs studied 75,000 brands and found that branded web mentions correlate with AI Overview visibility at 0.664, while backlinks correlate at just 0.218. The strongest factors were all off-site, led by branded web mentions, with branded anchors and search volume ranking next. Brands in the top quartile for web mentions earned roughly ten times the AI mentions of the next quartile down. So an audit that never leaves your CMS is blind to the factor with the highest correlation in the dataset.
Auditing this layer means looking at the web the way the engine does. Pull the set of category questions your buyers ask, run them through the engines, and record which third-party sources get cited: the review platforms, the forums, the industry publications, the comparison pages. Then check whether your brand shows up in those sources at all. The audit question is not "is my page good," it is "when an engine assembles an answer about my category, do independent sources corroborate that I belong in it."
There is a trap here that catches sophisticated teams. You cannot corroborate yourself. Amsive's analysis of 100 B2B "best category" queries found that when a brand published its own listicle ranking itself number one, Google's AI cited that page as a source but left the brand out of the actual recommendation 69% of the time, handing the recommendation to the established competitors named inside the list. So a GEO audit has to separate two things your own analytics will blur together: pages where you are the source, and answers where you are the recommendation. The first is easy to manufacture. The second is bought with third-party coverage, which is why this layer belongs in the audit and why a real digital PR program does more for AI visibility than another owned page.
Layer 3: measurement, and why a score is the wrong output
The standard audit ends with a number. The problem is that the number is built on a metric that does not mean what teams assume, measured at a single point in time on a surface that moves week to week.
Start with what a citation actually is. Semrush's ghost citations study found that 61.7% of AI citations are "ghost citations," where the engine links your page as a source but never says your brand name in the answer. The reader gets the information and walks away without knowing you provided it. Across the dataset, citations ran at nearly double the rate of brand mentions. Stack that on top of the recommendation gap from the previous layer and the picture is clear: a high citation count can sit on top of almost no named visibility and almost no recommendations. An audit that reports "you were cited 40 times" has not told you whether anyone heard your name.
So the output of a GEO audit should be a small, fixed prompt set you re-run, scored on three things a raw citation count hides: how often you are named, how often you are recommended, and how your share compares to competitors across the set. Pick ten to twenty prompts that mirror how your buyers actually ask, run them in fresh sessions across the engines that matter to you, and log the results on a schedule. The first run is your baseline. Every run after is the audit. You are watching the trend and the volatility, not chasing a single top spot, because AI answers shift enough that any one snapshot lies.
You do not have to build this from scratch. Google launched generative AI performance reports in Search Console in June 2026, giving site owners a dedicated view of impressions and pages inside AI Overviews and AI Mode in Search. That is a free baseline for the Google surface. For the cross-engine picture and the share-of-voice comparison, a dedicated tracker fills the gap, and choosing one is its own decision worth getting right before you commit. We walk through that in our guide to AI search visibility tools.
How to actually run it
Run the three layers in dependency order, because fixing a lower layer while an upper one is broken wastes the work. Confirm the retrieval substrate first: if you are not indexed or your passages do not stand alone as chunks, off-site mentions have nothing to attach to. Then audit off-site consensus, since that is the highest-correlation factor and the one your own tools cannot see. Then stand up the measurement loop so every later change is graded against a baseline instead of a vibe.
The reason this sequence matters is that AI visibility is not a separate channel you bolt on. The substrate is the same indexed, retrievable content that strong organic SEO produces. The off-site layer is the same brand-mention work that earns trust with human buyers. The measurement loop is how you tell whether either is working. When the three run together, they compound: indexed pages get retrieved, third-party mentions get you recommended, and the prompt set tells you which lever to pull next. That is the difference between a coordinated growth system and a folder of disconnected audit reports.
A GEO audit is worth running when it changes what you do on Monday. The checklist-and-score version rarely does, because it grades the layer that was already fine and stays quiet about the two that were not. Audit the substrate first, then the consensus, then the measurement. Done in that order, the audit tells you the one thing you actually wanted to know: where, exactly, the AI answer is leaking your brand. If you want a faster read on which layer is costing you, our SEO and AEO checklist is a quick place to start.
