Methodology

How the AI Visibility Score actually works

We ask the AI engines the real questions your buyers ask, several times each, then measure how often you get cited, mentioned and recommended. That rolls into a score from 0 to 100 with a confidence interval on it. No fake precision, and no number at all when the sample is too thin to mean one.

PerplexityAI OverviewsGeminiChatGPTCopilot

Engines measured

Every plan, no per-engine add-ons

Confidence interval

95%

On every score we publish

Runs per prompt

Repeated

One run is an anecdote, not data

The problem

Why checking once tells you almost nothing

Traditional rank tracking works because Google is broadly stable. Search the same phrase twice in a minute and you get the same ten links, so one look is a fair measurement.

Generated answers do not behave that way. The model samples as it writes, so the same question can return a different answer, a different set of sources and a different opinion of you within minutes. Ranking first on Google no longer guarantees the answer even mentions you, because the engine is choosing what to quote rather than what to list.

That makes a single check actively misleading. Ask once on a good run and you look visible. Ask once on a bad run and you look invisible. Both are the same brand on the same day. The only way to say anything true is to ask repeatedly and report the spread, which is what the rest of this page describes.

The method

Five steps, in order

We start from your real buyer questions

Vanity terms tell you nothing. Nobody asks an assistant for your brand name and then decides whether to buy from you. They ask the question that sits one step before the purchase, like which tool suits a small agency, or who is worth hiring in their city.

So we read your site first and work out what you actually sell, then generate the questions a real buyer would ask on the way to choosing. Those are the moments where you either appear or a competitor does. Everything else is decoration.

We sample repeatedly, because one answer is an anecdote

Ask an engine the same question twice and you can get two different answers with two different sets of sources. That is not a bug in the model, it is how generation works. Any tool that asks once and hands you a number is reporting a coin flip.

We repeat every prompt across every engine and look at the pattern instead of the roll. More runs mean a tighter range and a result you can act on. Fewer runs mean we say so rather than pretending.

We weight earned citations above easy mentions

Being named in a sentence and being used as a source are different things, and they are worth different amounts. A citation means the engine reached for your page to build its answer. A mention is a passing reference that can come from almost anywhere.

The score leans hardest on citations because they are earned and hard to fake. After that comes coverage, whether you appear at all, then prominence, how early in the answer you land, then sentiment, whether the framing helps or hurts you.

Every number carries its uncertainty in public

Each score ships with a confidence interval. When the sample is too thin to mean anything, we print no score at all and tell you why, because a made up number is worse than an honest gap.

The same rule governs movement. A week over week change only counts as real when the intervals do not overlap. That single rule stops most of the false victories this category is built on.

We keep the receipts

Every raw answer is archived exactly as the engine returned it, with the prompt, the timestamp and the cited sources attached. You can open any result and read the answer we scored.

It also means we can re-score history without paying to ask again. When the scoring improves, we recompute from what we already captured, so your past numbers stay comparable instead of quietly shifting under you.

How to read it

What the range around your score means

A score of 34 with a range of 28 to 40 is not a worse measurement than a flat 34. It is the same measurement, told honestly. The range says where the true value most likely sits given how much the engines disagreed and how many times we asked.

Three rules make it useful rather than decorative:

A wide range is a finding. It usually means the engines have no settled view of you, which is common for newer brands and contested categories.
Overlapping ranges are not a change. If last week was 28 to 40 and this week is 31 to 43, nothing has been proven. Treating that as growth is how people celebrate noise.
No range means no score. When the sample is too small, we show the gap instead of filling it with a number that would look confident and mean nothing.

Limits

What this does not do

A measurement tool that will not name its own limits is not measuring, it is marketing. Here are ours.

We cannot read your private chats

We measure the public style answers engines give to category and brand questions. What one logged in person sees in their own session, with their own history, is not something we can or should observe.

We cannot promise you a citation

Answers move week to week and no one controls them. Anyone guaranteeing a permanent spot in an AI answer is selling something. We measure where you stand and name the specific work that tends to earn a place.

A score is a measurement, not a verdict

A low score on a young site usually means nobody has written about you yet, not that your product is worse. The number is a starting point for work, not a grade.

Questions

The things people ask us

Why does the score come with a confidence interval?

Because AI answers are not deterministic. Ask the same question twice and you can get different sources back. We run each prompt several times and report a range, so what you are looking at is signal rather than one noisy roll of the dice. When the range is wide, that itself is the finding: the engines do not agree about you yet.

What actually counts as a citation?

Your own domain appearing as a cited source in the engine's answer. We track that separately from a mention, which is your name appearing in the text, and from a recommendation, which is the answer actively telling someone to pick you. Most brands score far better on mentions than citations, and confusing the two is how people convince themselves they are visible.

Which engines do you measure, and how honestly?

Perplexity, Google AI Overviews, Gemini, ChatGPT through OpenAI web search, and Microsoft Copilot through Bing Copilot Search. We label the ChatGPT one as a web search proxy rather than claiming we can read inside the app, because we cannot, and a tool about honest measurement should not start by overstating its own reach.

How many runs is enough?

It depends on how much the engines disagree. A question where every run returns the same three sources settles quickly. A contested question where the answer reshuffles each time needs more sampling before the range means anything. Rather than fixing a number and hoping, we widen the interval when the answers scatter and tell you when the sample is too thin.

Can you guarantee I will get cited?

No, and anyone who says otherwise is selling something. We show where you stand, who gets cited instead of you, and which specific sources the engines are pulling from for your category. Earning a place in those sources is the actual work, and it is work, not a switch.

Why does my score differ from another tool's?

Usually because the other tool asked once, asked different questions, or counted a mention as a citation. There is no shared standard in this category yet. That is exactly why we publish the method, show the raw answers, and put the uncertainty on the number instead of hiding it.

See it run on your site

The free check uses the same method described on this page. No login, no card, and a real confidence interval on the number it gives you.

Run a free check