AI Visibility Audit – GEO Brand Analysis

What an AI visibility audit covers – four layers

A solid GEO audit checks four layers: what assistants answer, which sources they lean on, whether the facts about your brand are true, and whether AI search crawlers can reach your site at all. Measuring answers without the other layers tells you there's a problem, but not where it comes from.

Every recommendation in the report comes from one of these layers, so it's immediately clear whether a fix concerns content, external sources, the brand entity or the tech stack. For the wider context, how assistants pick sources and how GEO differs from SEO, see my guide on GEO – AI search optimization.

How I design the audit question set

The question set determines the value of the whole audit: if it's mostly questions containing your brand name, the result will be inflated and useless. Most questions are phrased the way a customer asks before they even know your company exists.

The proportions I stick to

• About two-thirds of questions without the brand name – recommendations (“which firm should I choose for…”), category comparisons, problems and costs. This is where the customer decides who they'll consider at all.
• Questions with the brand name – “what do you think of X”, “X vs Y”, “how much does X charge”. Mainly for catching errors and sentiment.
• Control questions – a few where the brand has no chance of appearing. They help in later measurements to tell a model change from the effect of your work.

How I phrase questions

• In the customer's words, with the context people actually add: city, budget, company size, industry, constraints. Conversations with AI are longer than the phrases people type into Google.
• No leading questions. “Is X a good company?” almost always gets a positive answer, so it says nothing about real visibility.
• In the language your customers search in. For a company selling in several markets I measure each language separately, because answers and cited sources differ between languages.
• I agree a list of 3–6 competitors with you whose presence I also count. I often add one or two that AI recommends but you didn't consider competitors.

The example is illustrative, but it shows the principle: each question has an intent and a defined thing I read from it. More on finding such questions, including the ones no keyword tool will show you, is in my post on keyword research in the AI era.

Measurement conditions – why the same questions give different results

An assistant's answer depends not only on the question, but also on the model, whether search is on, location, language, account history and the date. If you don't control these conditions, you're comparing random screenshots, not brand visibility.

OpenAI's ChatGPT help pages explain that a temporary chat doesn't use memory or custom instructions by default unless the user chooses the personalized version. That's a good baseline for manual tests, but remember that your customers usually ask from their own personalized accounts. So I treat the audit result as a picture of the “average” user, not an exact replica of every conversation.

How many repeats are enough – a bit of statistics

A few repeats let you tell a brand that appears often from one that appears rarely, but they don't let you measure share of answers precisely. Precision only comes from aggregating many answers across many questions.

Brand share of answers is a proportion, so you can compute a confidence interval for it. Example: a brand appears in 30% of answers. With 10 answers, the 95% confidence interval (Wilson method) runs roughly from 11% to 60%; with 30 answers, from 17% to 48%; with 100, from 22% to 40%. In other words, with 10 repeats a “30%” result could just as well mean 15% or 50%.

What this means for the audit

• For a single question, 5–10 repeats per assistant are enough to classify the result as “never”, “sometimes” or “almost always”. I don't report percentages to the point in that case.
• Aggregate metrics, such as the brand's share across all recommendation questions in ChatGPT, are computed from hundreds of answers, so they're far more stable.
• When comparing with competitors I check whether the intervals overlap. If they do, I don't write that the brand is “ahead”.
• In later measurements, a change of a few percentage points on a small sample is usually noise, not the effect of your work.

Why repeating is necessary in the first place is shown by the SparkToro and Gumshoe.ai study discussed in the guide on GEO – AI search optimization: lists of recommended brands almost never repeat in identical form.

How I score answers – rubric and quality control

I score every answer with the same rubric, and the scoring is done by a language model (LLM-as-a-judge) whose work I check by hand on a random sample. Without a fixed rubric, two people, or even the same person on two different days, would score the same answers differently.

Fields I fill in for every answer

• Brand presence – yes or no, separately for each brand on the competitor list.
• Position in the list – when the answer names several companies.
• Role – recommended outright, mentioned neutrally, mentioned with a caveat, or discouraged.
• Justification – what the model praises or criticizes the brand for. Often the best hint for what to communicate.
• Factual accuracy – whether the brand data matches the fact sheet we prepare at the start.
• Sources – cited URLs and domains.

Why I check the judge

Zheng and co-authors, in “Judging LLM-as-a-Judge” (NeurIPS 2023), showed that a strong model acting as judge agreed with human ratings in over 80% of cases, roughly as often as humans agree with each other. They also described typical biases: favoring answers in a certain position, longer answers, and answers from the judge's own model. So I give the judge narrow, closed questions (is brand X named in the text, is it recommended) rather than a general “how well does the brand come across”, and I verify a random share of the scores by hand. When I find disagreements, I fix the rubric and re-score the set.

Errors about your brand in AI answers – how I classify them and where they come from

Every factual error about the brand gets a severity and a hypothesis about its source, because the source determines the fix. An error taken from a cited page is fixed differently from one the model “remembers” from training data.

How I look for the cause

1. I check whether the answer with the error cites a source. If it does, I open it and look for the wrong information: often an old business directory, an outdated article or a profile nobody remembered.
2. If there's no source, I compare answers with search on and off. An error that appears only without search points to training data and usually fades more slowly.
3. I check data consistency where the brand is described: About page, structured data, Google Business Profile, company registers, LinkedIn, Wikidata.
4. When the brand is confused with another company, I look for what links the two entities: a similar name, the same street, a shared domain in the past.

Fixing errors is usually work on the brand entity, which is what knowledge graph engineering covers. What to do step by step when ChatGPT says something untrue about your company is in my post on brand reputation in AI.

Source analysis – the part of the report that pays off most

The list of domains cited in your category usually turns out to be the most practical part of the audit, because it shows directly where you should be. In the report I compile it across all questions and separately for each assistant.

• Domain ranking – which sites are cited most often and for which question types.
• Breakdown by category – your site, competitor sites, media and rankings, communities, knowledge bases, directories.
• The “cited, but without you” list – specific URLs that assistants cite for recommendation questions, which feature your competitors but not your brand. A ready target list for digital PR.
• Your cited URLs – which pages already work as sources and for which questions. Often not the pages the company considers most important.

Each platform has different favorite sources, so a list for ChatGPT won't replace one for Perplexity or Google AI Overviews. How I turn such a list into an off-site plan is described in my AI citation strategy and ChatGPT SEO service.

The technical part of the audit – can assistants read you

Before you start improving content, rule out the simplest cause of missing citations: an AI search crawler can't fetch the page, or the page isn't in the index the assistant uses. In the audit I check this in several places, because one data source is often not enough.

• robots.txt – whether you let in search crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot, Bingbot, Googlebot), regardless of your decision on training crawlers.
• Firewall and CDN – whether bot protection rules block those same crawlers. Since July 2025 Cloudflare blocks known AI crawlers by default on new domains (MIT Technology Review), so a correct robots.txt isn't enough.
• Server logs – which bots actually arrive and which response codes they get.
• Bing's index – in tests ChatGPT relied heavily on Bing for search, and Copilot runs on it directly. I check indexing in Bing Webmaster Tools.
• Content without JavaScript – whether the main content, prices and contact details are in the HTML, not only after scripts render.
• AI Performance report in Bing Webmaster Tools – since February 2026 it shows how often your pages are cited in Copilot and Bing's AI summaries, the cited URLs and a sample of the queries that led to them.
• GA4 – whether assistant traffic is separated. ChatGPT adds utm_source=chatgpt.com to its referrals, as OpenAI describes in its publisher FAQ.

Crawler configuration is covered in my post on AI crawlers and robots.txt, and traffic measurement in how to measure traffic from ChatGPT, Perplexity and AI Mode. If the audit finds problems in this layer, I fix them as part of Perplexity, Copilot and ChatGPT Search optimization.

What the report contains and how I set priorities

The report ends with a list of actions ranked by impact, effort and who can carry them out. The measurement results are only the justification for that list.

Report structure

1. A one-page summary: the key numbers, the three biggest problems and the first three actions.
2. Brand share of answers against competitors, by assistant and intent, with measurement uncertainty marked.
3. Source analysis with the “cited, but without you” URL list.
4. A list of factual errors with severity, example answers and a hypothesis about the source.
5. Technical findings.
6. The action plan and raw data: questions, answers and scores, so you can repeat the measurement yourself or with another provider.

How I rank actions

A one-hour DIY AI visibility mini audit

A simple test on your own won't replace an audit, but it's enough to see whether the topic matters for your company at all. You need a spreadsheet and three assistants.

1. Write down 10 questions: 7 without your brand name (recommendations, comparisons, costs) and 3 with it.
2. Pick ChatGPT, Google (AI Mode or AI Overviews) and one assistant your customers use, e.g. Perplexity or Copilot.
3. Ask each question three times in each assistant, each time in a new, clean session without history or personalization.
4. In the spreadsheet, record: whether your brand came up, which competitors came up, which domains were cited, and whether the brand description contains an error.
5. Count how many of the 63 answers to the non-brand questions included you, and how many included each competitor.
6. List the domains that appeared in citations at least three times.

How to read it: if competitors show up clearly more often and you barely appear, the problem is real and worth investigating properly. Keep the limits in mind, though: 9 answers per question is too few to track month-to-month changes, and manual logging is easy to get wrong. The next steps for measuring on your own are in my post on AI visibility audits and AI share of voice.

Monitoring tool or audit – which to choose

An AI visibility monitoring tool measures regularly but doesn't tell you why the result is what it is. An audit measures once, more thoroughly, and ends with a diagnosis and a plan. In practice they complement each other.

A comparison of popular tools (Profound, Peec, Otterly, Semrush) and when building your own makes sense is in my post on AI visibility monitoring tools.

What happens after the audit

An audit only makes sense if something changes afterwards. The results usually point to one or two main directions, and the action plan assigns them to specific tasks.

• Brand missing from recommendations, with rankings and media as the sources – ongoing citation and mention work as part of an AI citation strategy.
• Factual errors and confusion with other companies – knowledge graph engineering.
• Good visibility in Google, none in ChatGPT, Perplexity and Copilot – Perplexity, Copilot and ChatGPT Search optimization.
• Falling Google clicks where AI summaries appear – Google AI Overviews and AI Mode optimization.
• Local and voice questions – voice search optimization.

After the first changes are live, I repeat the measurement on the same question set and under the same conditions, usually after about three months. Only then does a before-and-after comparison mean anything.