What an AI visibility score measures, and what it hides

What an AI visibility score is

An AI visibility score is one number that sums up how present your brand is in AI answers. It usually runs from 0 to 100, and a higher number is meant to be better.

The idea is borrowed from SEO tools, which have scored search visibility for years. The difference is in what sits underneath. A search ranking is one fixed list for one query. An AI answer is written fresh each time, names a handful of brands, and may or may not link to a source.

So an AI visibility score has to make choices that a search score did not. Which answers count, which brands count, and how much each signal weighs are all decisions the vendor makes. Most of this post is about those decisions.

Scores come from three kinds of product. SEO suites have added them beside their search reports. Dedicated AI visibility tools put them on the first screen. Free graders, such as the one HubSpot offers, give a one-time score at no cost, with the brand checked once rather than tracked over time.

What AI visibility scores typically combine

Most scores are built from some of six signals. The table names each one and what it counts, using the definitions vendors publish on their own pages.

Some vendors use only one of these signals and call it a visibility score. Others blend several. The name tells you nothing about which kind you are looking at.

Same name, different denominators

The published definitions disagree on the most basic question: visible out of what? The number on the bottom of the fraction, the denominator, decides what the score means.

Look at the first two rows. Peec AI's metrics page divides by every response. Profound's help centre divides by the responses that name at least one brand. An answer that names nobody counts against you in the first and is left out of the second. The same answers can give two different percentages, and both can be correct by their own definition.

Semrush's knowledge base calls its AI Visibility score a benchmark from 0 to 100 that shows how often a brand appears compared to competitors. The page does not give the formula. A separate Semrush article says its share of voice counts how many times a brand is mentioned and how high it appears, and that for ChatGPT it is also weighted by topic search volume.

HubSpot's AI Search Grader is unusual because it publishes its weights. Sentiment carries 40 of the 100 points, so a brand that is rarely named but described warmly can score well. That is a choice HubSpot has made openly, and it is a fair one to know before you compare its score with anyone else's.

Ahrefs' methodology page says its impressions weight mentions by Google search volume, and that this is a modelling choice, not a measured relationship. It adds that Ahrefs does not claim a validated link between Google search volume and how often a query is asked inside an AI tool. That is an honest caveat, and few vendors write one.

Why a composite score hides its denominator

A composite score blends several counts into one number, and the blend throws away the counts. Once mentions, sentiment and position have been weighted and added, you cannot work back to how many answers named you.

Here is a made-up example to show how. Two brands get the same score of 50 on a tool that, for this example only, gives half the points to how often a brand is named and half to sentiment. No real vendor's formula is being described.

The two brands have the opposite problem, and the score cannot tell them apart. Brand A needs better things said about it. Brand B needs to be named at all. A marketer shown only "50" would not know which work to do.

There is a second loss. A score of 50 from 100 answers and a score of 50 from 10 answers look the same, but the second could swing wildly on one answer. The count beneath a figure is what tells you how much to trust it, and a composite drops it.

A figure you cannot recompute is a figure you cannot check. Ask for the counts underneath any score before you act on it.

Why the score moves when nothing changed

AI assistants answer the same question differently each time, so any score built on their answers moves on its own. A rise or a fall can be noise rather than the result of anything you did.

SparkToro's study, published in January 2026, had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. It found less than a 1 in 100 chance that ChatGPT or Google's AI would give the same list of brands in any two responses, and roughly 1 in 1,000 for the same list in the same order.

The study did find a measure it trusted: a visibility percentage across many prompts, each run several times. It was blunt about rank, saying any tool that gives a ranking position in AI is full of baloney. That matters for scores that weight position heavily, because position is the least stable thing in an answer.

The practical answer is to learn the normal swing before reading a change. Ask the vendor how many times each question is run per reading. Then watch the score for a few weeks while you change nothing. A later move smaller than that resting wobble is not evidence of anything.

What to ask a vendor before you trust the score

Seven questions will tell you what a score means and whether you can rely on it. A vendor with a sound score can answer each in a sentence.

1. What does the score divide by: all answers, answers that name some brand, or something else?
2. How many answers is my score based on, and can I see that count beside it?
3. Are questions that already contain my brand name included? Those are answered with my name almost every time.
4. Which signals go into the score, and what weight does each one carry?
5. Which assistants are asked, and is each one reported separately as well as in the total?
6. How much does my score move between two runs when nothing on my site has changed?
7. Can I open the actual answers behind the score and read them?

The third question deserves its own warning. A question like "Is Example Co any good?" names the brand, so the answer names it back. If such questions are mixed in, the score rises without a single new buyer finding you. Our guide on how to report AI visibility honestly works through this with a made-up example.

The fifth question matters too. A single score averaged across assistants can hide one that never names you. If ChatGPT names you often and Perplexity never does, the average looks middling. The real problem, one assistant that ignores you, disappears from view. Ask for one row per assistant, each with its own count.

A vague answer to any of the seven is not proof the score is wrong. It does mean you cannot check it, and you should treat it as a rough signal rather than a measurement.

What a score you can check looks like

The most useful AI visibility figure is a count over a stated total, for questions that do not contain your name. "Named in 6 of 120 answers" can be checked, compared and argued with. "Visibility 42" cannot.

A score still has a use. It can be a quick way to see that something moved. The count is what tells you what moved, and by how much.

How Agentcites reports visibility

Agentcites does not give a composite score, by design. Its visibility figure is the share of non-branded answers that named you, shown with both counts beside it, for each assistant and in total. Branded questions are reported separately and never blended in. Today it asks ChatGPT, Perplexity, Gemini and DeepSeek each week, and every figure opens to the answers behind it. The method page sets out how each number is counted.