How to Run an AI Visibility Audit in an Afternoon
An AI visibility audit is a structured count. You put a fixed list of buyer questions to the AI assistants, record whether your brand appears in each answer, and note which sources got used instead of you. It takes an afternoon the first time and about an hour a month after that.
There is no tool you have to buy to do it. There are tools that make it faster, and I get to them at the end, but the method below is the whole thing and it runs on a spreadsheet.
I ran this on my own site and scored zero. That is a real result rather than a rhetorical opening, and I have published the full baseline separately. What follows is the method that produced it.
What questions should you audit?
The ones a buyer types before they know your name. If your brand appears in the question, the question is worthless, because only someone who already found you would ask it.
Write twenty to forty of them. The useful ones sound like problems, not categories: “how do I compare my site against a competitor without paying for an enterprise tool”, “what should a small team measure instead of ROAS”, “who does competitor research for small agencies”. Pull them from the questions clients actually ask you on calls, from your inbox, and from the People Also Ask box on your main topics.
Keep the list fixed once you have written it. This list is your baseline instrument, and changing it later destroys your ability to compare readings.
Which AI assistants should you test?
Test the ones your buyers use, which for most people means starting with ChatGPT and Google’s AI answers, then adding Perplexity and Gemini if your audience skews technical. Testing more assistants makes the audit longer without usually changing the decision.
| Assistant | Why it earns a place | When to add it |
|---|---|---|
| ChatGPT | The default for most people asking an open question | First pass |
| Google AI answers | Sits above the results your buyers were already going to see | First pass |
| Perplexity | Cites sources heavily, so it exposes who is trusted on your topic | Second pass |
| Gemini | Worth adding if your audience skews technical or Android-heavy | Second pass |
Two are enough to find out whether you have a problem. If you are invisible on both, adding a third will not rescue you, and the fix is the same regardless. Breadth matters later, when you are tracking improvement and want to know which surface moved first.
What do you record for each answer?
Five columns, and the fifth is the one people leave out and then wish they had. Whether you appeared at all is only the first of them.
| Column | What you write | Why it matters |
|---|---|---|
| Appeared | Yes or no | The headline number, and the only one that matters at zero |
| Position | Where in the answer you were named, first or fifth | Being mentioned last in a long answer is closer to invisible than it looks |
| Accuracy | Whether what was said about you is correct | A confident wrong description is worse than no mention |
| Sources cited | The domains the answer used | The list of who the model currently trusts on your topic |
| The answer itself | Paste it whole | Answers vary between runs, and you cannot re-read a summary |
That fourth column is the actual output of the audit. The domains cited in your place are not trivia, they are the specific pages standing between you and the answer. Some of them will be publications you could realistically be mentioned in. Some will be competitors. A few will be forum threads, which is a cheaper opening than it looks.
How do you score an AI visibility audit?
Divide the number of answers you appeared in by the number of questions you asked. That single percentage is your visibility rate, and for a first audit it is enough.
Resist building a weighted composite on the first pass. A score that blends position, sentiment and citation share into one number feels rigorous and hides the thing you need to see, which is usually that the raw appearance rate is very low. Add nuance when the headline number is no longer the story.
What does a zero actually mean?
It means you have not been established as a source on those questions yet. It does not mean your site is broken, penalised or badly built, and it is not a judgement on the quality of your work.
This distinction matters because the fixes are completely different. A technical problem gets a technical fix. Absence from an answer is a content and reputation problem, and it gets solved by being genuinely useful on the specific questions, in enough places, for long enough that a model has reason to associate you with them.
Do you need an AI visibility checker tool?
Not for the first audit, and possibly not for the second. What a tool buys you is scale, scheduling and history, which are all things you do not need until the manual version has told you whether there is anything to track.
A free AI visibility checker will give you a number in thirty seconds. Treat that number as a smoke alarm rather than a diagnosis: it tells you something is worth looking at, and it will not tell you which questions you lost or who won them, which is the part that changes what you do on Monday. I have written up what the paid categories actually measure in LLM SEO tools and what they measure.
What do you fix first?
The questions where a competitor is named and you are not, on topics you genuinely know more about. That is the shortest distance between the audit and a result.
Resist the technical checklist. Google’s own guidance on generative AI features states that “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add”, and that there is “no requirement to break your content into tiny pieces for AI to better understand it”. A week spent adding markup because a checklist told you to is a week not spent answering the questions you lost.
What does help is writing the answer so it can be lifted. Put the question in the heading, answer it completely in the first two sentences underneath, and let the detail follow. The reasoning behind that, and the four metrics that measure the result, are in LLM visibility and how to measure it.
How often do you repeat it?
Monthly, same questions, roughly the same date. Model answers move on their own, so a single reading is a data point and two readings a month apart are the first thing that resembles a trend.
Keep every raw answer you pasted. In six months the interesting question will not be your score, it will be how the description of you changed, and you cannot reconstruct that from a percentage.
If you would rather have the competitive half of this done for you, Vantage compares your site against up to three competitors and includes how each brand shows up inside AI answers, returning a graded report in about thirty minutes. The findings are AI created, so read them with your own judgement before acting on them. The first analysis is free, which is enough to see where you stand before you commit to anything.
The audit is not complicated. It is just specific, and almost nobody does it, which is most of why it is worth doing.
Fifteen years in growth: VP Digital Lead at Vodafone Idea, digital lead for Nestlé Indonesia at dentsu, and nine years at Performics on brands like Citibank and Taj Hotels. Now co-founder at ShopLoco and Adroit Digital. He writes here and builds tools in the Lab.