How to audit your brand's visibility in AI

An honest method for measuring whether AI assistants recommend your brand: questions, engines, metrics and the mistakes that invalidate an audit.

Auditing AI visibility means measuring whether, and how, the assistants recommend your brand when a buyer asks. It is done in five steps: define the questions, choose the engines and the modes, run and record, calculate the metrics and diagnose the cause. An initial snapshot can be done by hand today; continuous measurement needs a tool.

What auditing AI visibility means

An AI visibility audit answers a simple question: when a buyer asks an assistant for a company that does what yours does, does yours appear? And if it appears, in what position, in which engines, with what description? It is not an impression, it is a measurement. And like any measurement, it is only worth something if it is done with method.

The most common mistake is confusing an audit with curiosity: typing the company name into ChatGPT once and drawing conclusions. That measures nothing. The assistants give different answers from one another and even between two identical questions. A serious audit has five steps.

Step 1: define the questions

Everything rests on a fixed set of questions, the ones your buyers actually ask. Not your company's name, but the need it solves. Cover the various stages: generic category questions (who does this in Portugal), comparative ones (which to choose between X and Y), and decision ones (which do you recommend hiring). It is these last, the bottom-of-funnel ones, that convert most and where appearing matters.

Twenty to thirty well-chosen questions are enough for a first reading. What matters is fixing them: they are the same in every subsequent measurement, otherwise no comparison is possible.

Step 2: choose the engines and the modes

One engine is not enough. Google's AI mode, ChatGPT, Perplexity, Claude, Gemini and the others train on different data and search in different ways; being strong in one and invisible in another is the norm. Test in the ones your buyer uses, and measure several.

Within each engine, distinguish two modes. In the memory answer, the assistant answers without the internet, from what it learned in training; that measures the strength of your entity. In the answer with live web search, the engine searches and cites sources at that moment; that measures your presence on the web. The same brand can be strong in one mode and absent in the other, and the diagnosis depends on separating the two.

Step 3: run and record

Ask each question, in each engine, more than once, because the answer varies between runs. For each answer, record three things: whether the brand was mentioned, whether it was recommended as an option (which is different from mentioned in passing), and in what position it appeared. Also keep the other brands cited and the sources the engine used, because they are the map of where the recommendation is born.

A note that avoids false positives: if the brand only appears in a context of negation, of the kind "I could not find information about that company", that is the opposite of a citation. It counts as absence, not as presence.

Step 4: calculate the metrics

From the records, three metrics are calculated, always with an explicit denominator:

• Citation rate: in what percentage of the answers the brand appears. It is the primary metric.
• Share of voice: what slice of the category's mentions is yours, compared with the other brands.
• Average position: when it appears in lists, in what place, knowing that a lower position is better.

The words that matter are fixed and explicit. Saying "we appear 30% of the time" only means something if you say in how many questions and in how many engines. Without a denominator, the number is not auditable.

Step 5: diagnose the cause

An audit does not end at the measurement, it ends at the diagnosis. If you are absent, the question is why, and there are three typical causes that are told apart by the data you have already collected:

• A weak entity: absent mostly in memory mode. The models do not fix you because your identity is thin or contradictory across platforms.
• Missing content: absent mostly in search mode. You have no pages that answer the questions and no presence in the sources the engine reads.
• A technical barrier: the site is not legible to AI crawlers, the content only exists after JavaScript runs, or the crawlers are blocked in robots.txt.

The diagnosis is what turns the audit into a plan. Without it, you have numbers; with it, you know where to act first.

By hand or with a tool

An initial snapshot can be done by hand, and it is worth it: define the questions, run them, record, and you already know where you stand today. It is work, but it demands nothing beyond time and method.

What the hand does not sustain is continuity. AI visibility moves: week by week competitors come in, the models update, the sources change. Measuring once is a portrait; measuring always is a time series, and only the series shows whether you are improving. It is the difference between knowing where you are and knowing where you are going. That is also why we built a tool that runs this measurement across every engine and surface, every week.

The mistakes that invalidate an audit

• A single engine. The answers vary too much between them.
• A single run. The same question gives different answers.
• Confusing a mention with a recommendation. Appearing in passing is not being recommended.
• Not fixing the questions. Without the same set, there is no comparison.
• Numbers with no denominator. A percentage with no base is not auditable.
• Stopping at the measurement. With no diagnosis of the cause, the audit generates no action.

Frequently asked questions

Read next

• How to measure citation rate and share of voice, the metrics in detail.
• Six commands to find out whether the AI can read your site, for the third of the three causes.
• Knowledge vs augmented, the two modes explained.