What an AI visibility score actually measures
No AI engine publishes a visibility score, so every number you see is its vendor's own construction, computed over the prompts you chose to track and the competitors you named. Two tools will disagree for that reason alone. Read the trend, freeze the prompt set, and check the source list rather than the score.
By Rajat Kapoor
Updated August 2026
Key takeaways
No AI engine publishes a visibility score, so every figure you see is the tracking vendor's own construction rather than an estimate of one underlying number.
Visibility is calculated over the prompts you chose to track, so adding prompts you already win raises the score without anything changing in the engines.
Share of voice divides your mentions by the mentions of every competitor you added to the tracked set, so naming a large competitor lowers your score the same day.
Freeze both the prompt set and the competitor set whenever you compare two time periods, or the chart is not a measurement.
A mention names your brand in the answer while a citation references your page as a source, and the two have different causes and different fixes.
Visibility rising while average position falls is the normal shape of progress, because newly won mentions tend to appear late in an answer.
If you have one of these tools open, you are looking at a number between 0 and 100, or a percentage, that claims to describe how visible your brand is inside AI answers. Someone in your business is going to put that number in a slide, and treat a change in it as a result.
Before it can carry that weight, it is worth knowing three things about it: what it counts, what makes it move, and what it cannot see. None of the three is obvious from the dashboard, and the first one is not even consistent between vendors. This page covers each in turn, using the definitions the tools publish themselves.
No engine publishes a visibility score
Worth settling before you read any number: ChatGPT, Gemini, Perplexity and the rest publish nothing. There is no underlying figure these tools are competing to estimate accurately, the way two thermometers estimate one temperature. Each vendor has defined a metric, given it a name the other vendors also use, and computed it over a prompt list you assembled.
Their own documentation shows it. Peec AI defines its Visibility Score as the percentage of AI responses that mention your brand, which is an absolute figure: it does not move when a competitor has a good quarter. Semrush defines AI Visibility as a benchmark score from 0 to 100 showing how often your brand appears compared to competitors, which is a relative one and does. Both are called visibility. They are not the same quantity, and no amount of accuracy on either side would bring them together.
This matters most when you are trialling two tools at once, because the explanation usually offered is measurement noise. Noise is real, and it is a separate issue covered on our AEO tools ranking. Two tools can disagree while both work perfectly, because they are not measuring the same thing.
The number moves when your setup moves
This is the part that catches people out, and the one most worth understanding, because it decides whether your reporting means anything. Both headline metrics are computed against a set you control, which means you can move them without anything changing in the engines at all.
Visibility is a percentage of the responses to the prompts you chose to track. The denominator is your own prompt list. Add ten prompts on a topic you already own and the score rises; drop the ones you lose and it rises again. Nothing has changed about what the engines say. It is also why a prompt set full of branded questions produces a flattering number, since a question with your company's name in it will tend to mention your company.
Share of voice has the same property one level down. Peec publishes the formula as your brand's mentions divided by the total mentions of all tracked brands. That tracked set is a list you made. Add a large competitor to it and your share of voice falls the same afternoon, having lost nothing.
Two rules follow, and they are worth building into whatever reporting you put around this.
Freeze the prompt set and the competitor set whenever you compare two periods. A month-on-month chart is only a measurement if the denominator held still.
Change the prompt set when the questions you care about change, and never to move the number. It is the one lever here that improves the metric without improving anything, which makes it the easiest one to pull by accident when a board slide is due. A prompt set is a research instrument. Reweighting it is not optimisation, it is editing the question you asked.
A mention and a citation are different events
The two words get used interchangeably, including by the tools. They describe different events, they have different causes, and confusing them sends you to work on the wrong thing, so the distinction is worth ten minutes.
Peec's documentation draws the line cleanly: sources are every URL a model accessed while generating a response, and citations are the ones explicitly referenced in the answer text. Your brand can be named in an answer with your website never touched, because the model knew about you from training or from somebody else's page. Your site can be retrieved as a source without your brand being named at all, because it informed the answer without earning a mention.
The distinction decides what you do next. Mentioned but never cited, and the work is on your own pages, on whether they answer the question directly and in a form a model can lift. Cited but not mentioned, and the work is the reverse. Neither, and the useful question is which sources are being used for those prompts, because those are third-party pages you may be able to influence by the ordinary means of getting written about.
One constraint is worth knowing before you spend a quarter on this. Models read HTML. Peec's documentation states plainly that they cannot read content behind a paywall, or content that only exists once JavaScript has run. If your best material is client-rendered, it may be invisible to the thing you are trying to be visible in, and no amount of prompt tracking will surface that as the cause.
Position is not a search ranking
Same word, different objects again. Peec measures position as the average rank of your brand among the brands named in a response. Semrush's average position is where a citation of your domain appears among the citations. One is about your name in prose, the other about your URL in a reference list.
Read it alongside visibility rather than on its own, because the two move independently and each is misleading by itself.
What the score cannot tell you
Three things, and none of them is a defect in any product.
It cannot tell you why. A tracker reports that you were absent from an answer. It has no view of which page would have won the mention, which is why the useful screen in any of these tools is the source list rather than the headline score.
It cannot tell you whether a mention produced a customer. AI assistants do not reliably pass a referrer, so the link between an answer naming you and somebody arriving on your site is mostly unobservable. Some tools watch requests hitting your own infrastructure instead, which is a real capability and a different measurement, and a good deal of what it sees is automated agents rather than people.
It cannot tell you what your buyer saw. These tools query from a clean context. Real users carry conversation history, saved memory, a location and an account, and answers vary with all of it. The reading is a laboratory measurement of a system that behaves differently in the field, which is workable as long as the trend is the signal and no single answer is treated as evidence.
Frequently asked questions
What is a good AI visibility score?
There is no benchmark, and any figure quoted as one is describing a single vendor's metric computed over somebody else's prompt list. Because the denominator is the set of prompts you chose, a score of 40 percent on twenty tightly focused buying questions and a score of 40 percent on two hundred broad ones are not comparable achievements. Judge it against yourself instead: the same prompt set, the same competitor set, measured a month apart. The useful question is not whether the number is good but whether it moved, and whether anything other than the number changed to make it move.
Why do two AI visibility tools give completely different numbers for the same brand?
Usually for two independent reasons, and the first is the one people miss. The tools are not calculating the same quantity: one vendor may define visibility as the plain percentage of answers that mention you, while another defines it as a 0 to 100 score benchmarked against competitors, and both publish those definitions openly. On top of that, each tool sends your prompts to the engines on its own schedule and records what comes back, so results vary between runs. The first difference is definitional and will never converge. Check each vendor's own documentation for how the metric is computed before concluding that either tool is wrong.
Does adding more prompts improve my AI visibility?
It changes the score without changing your visibility, which is not the same thing. The metric is a percentage of responses to the prompts you track, so the prompt list is the denominator. Add questions you already win and the percentage rises; add hard ones and it falls. Neither movement reflects anything the AI engines did. Add prompts when the questions your buyers ask have genuinely changed, and accept that doing so breaks comparability with earlier months. If you need the trend to stay readable, keep the original set intact and track new prompts as a separate group.
What is the difference between a mention and a citation?
A mention is your brand named in the text of an answer. A citation is your page referenced as a source for that answer. They are independent: a model can recommend you from what it already knows without ever fetching your site, and it can read your page as background while naming somebody else. The fix differs accordingly. If you are mentioned but not cited, the gap is usually that your own pages do not answer the question in a liftable form. If you are cited but not mentioned, your content is useful but your brand is not the one the answer is about.
Can a visibility tool tell me whether an AI mention brought me a customer?
Not reliably, and it is worth planning around that rather than waiting for it to be solved. AI assistants largely do not pass referrer data the way a search engine does, so a visit that began with a chatbot recommendation often arrives looking like direct traffic. Some tools observe requests reaching your own infrastructure, which is a genuine measurement and a different one from prompt tracking, and much of what it records is automated agents rather than human visitors. In practice, treat AI visibility as an upper-funnel measure and expect the attribution to stay partial.
My visibility went up but my average position got worse. Is that bad?
Usually the opposite. Average position is calculated only across the answers where you appear, so it is an average over a changing population. When you start being mentioned for prompts you used to lose entirely, you tend to enter those answers late, as the fourth or fifth name rather than the first. Those late mentions pull the average down while your overall presence grows. Both numbers are accurate and the combination is the normal shape of early progress. It is worth worrying about only when visibility is flat and position is still sliding.