AI visibility score: what it measures, how tools calculate it, and how to work out your own.

The short answer.

An AI visibility score tells you how often AI engines mention your brand across a fixed set of questions, expressed as one number. It summarises a sample, not the whole market, and only means something next to the method that produced it.

Three things are true of every score on sale today:

• ✓It starts with a list of questions. Change the questions and the score changes.
• ✓It depends on a formula the vendor chose. Some count answers, some count mentions, some add weights.
• ✓It is an estimate. Semrush’s own documentation says AI answers are “fast-changing and highly personalized, which means no platform can provide exact numbers on visibility.”

None of that makes a score useless. A score built the same way every month shows whether you are gaining or losing ground. It misleads only when the method is hidden, or when numbers from two methods are compared as if they were the same thing. That is why this guide spends most of its time on arithmetic. Our GEO and AEO services work on the same basis: each check records the question, the engine and the date.

The five choices inside every score.

Every AI visibility score is the result of five choices: which questions, which engines, how many runs, what counts as appearing, and what the count is divided by. Two tools that differ on any one of them will give different numbers for the same brand.

The questions

Some tools let you write your own prompts. Others score you against a database they built. Semrush describes its reports as powered by “a collection of over 317 million prompts and responses”. Ahrefs says its Brand Radar queries come from Google’s “People Also Ask” data and its own keyword database. A vendor database describes broad topics; your own questions describe your buyers.

The engines and the runs

Tools cover different engines and refresh at different speeds. Profound’s documentation describes prompts as the queries it “sends to answer engines daily”. Ahrefs says its chatbot question sets are “updated and tested in chatbots monthly, using a 90-day reporting window”. A daily average and a monthly sample will rarely agree to the point.

What counts, and what it is divided by

This is where most of the disagreement comes from. The next two chapters set the formulas side by side.

How six tools calculate it.

The vendors’ own help pages show four different ideas behind the same label: a share of answers, a share of mentions, a benchmark against competitors, and a weighted mix of judgements. The table quotes each vendor’s documentation.

Semrush’s score is relative, not a percentage of answers. Its data page gives an example: Google and Reddit “may get mentioned across many topics but only once per topic, while Lego appears in fewer topics but dominates those topics—earning a higher AI Visibility Score.” Another Semrush help page describes it as mentions compared with the median for your top industry competitors. Either way, a 40 there does not mean “named in 40% of answers”.

Position is a separate measure in most tools. Peec reports Position as “the average ranking across all responses where your brand appears”. Profound says an average position of 1 “means that your brand is typically the first one to be mentioned”. Otterly converts average position with a published formula, 10 + ((5 − average position) ÷ 4) × 90, so first place reads 100% and third place 55%.

HubSpot’s free grader measures something else again. Share of voice is a tenth of its score; sentiment is four tenths.

If you are choosing a product, our comparison of AI visibility tracking tools covers engines, data collection and the questions to ask each vendor.

Why scores from different tools do not match.

Scores disagree because the formulas divide different things by different things. One set of AI answers fed into four published formulas gives four numbers, none of them wrong.

Take an invented set of ten answers. Your company is named in four. Two answers name no company at all. In total, 24 company names appear, four of them yours. Your website is cited in three answers.

Nothing about the company changed between the top bar and the bottom one. Only the sum did.

Weighting moves the number further

Some tools then weight each question by estimated demand. Ahrefs publishes a worked example. Three brands each appear in two of three responses, so each has a simple share of 67%. Once the prompts are weighted by estimated demand, one brand’s share falls to 31%, in Ahrefs’ words “because your mentions come from the prompts people ask least.”

The same article is candid about what this means: “two tools can report different AI SOV for the same brand without either being wrong”. Ahrefs’ methodology page says of its impression estimates that “we don’t claim a validated link between Google search volume and how often a query is asked inside an AI tool.”

So before you compare two scores, check the questions, the engines, the dates, what counted as an appearance, and the denominator. If any differ, they are two different measurements.

Named, cited or absent.

For any one AI answer, a company is in one of three states: named in the answer, cited as a source without being named, or absent. A useful score keeps those apart.

Ahrefs draws the same line: it counts a mention “when a brand appears at least once in an AI generated response” and a citation “when a page appears at least once as a cited source”. Otterly reports brand coverage and domain coverage separately.

The distinction matters commercially. A company that is often cited but rarely named has pages the engines trust and a brand they do not yet connect to the category. Our study of where AI answers get their sources shows which kinds of site the engines link to for B2B buying questions, and our guide to page patterns that earn AI citations covers the cited side in detail.

A do-it-yourself method.

You can calculate a defensible AI visibility score with a spreadsheet: fix a panel of buyer questions, run each one on each engine, mark every answer named, cited or absent, and average the marks into a number from 0 to 100.

Step 1: the panel

Use buyers’ wording and mix the types: category questions (“best … for …”), comparisons, alternatives, problems and requirements such as a location or a certification. Leave your own name out of the question. Our guide to running a manual AI visibility check explains how to word them, and the play on building a fixed question panel gives a short procedure.

Step 2: the engines

For most B2B companies that is ChatGPT with search on, Google’s AI Overviews or AI Mode, and Perplexity. Use a logged-out or temporary session, and note the date and country. We cover how Perplexity finds and cites B2B sources in a separate guide.

Steps 3 and 4: mark and score

Give each answer exactly one mark. If it names you, it is named and earns 1 point, whether or not it also links to you. If it does not name you but one of your pages is in the sources, it is cited only and earns half a point. Otherwise it is absent.

The half point is a judgement, and we would rather state it than hide it: the engine read you, but the buyer does not see your name. If you disagree, set it to zero or one in the calculator below.

Step 5: average

For each engine: points divided by answers, times 100. The headline score is the plain average of the engine scores, so each engine counts equally. We do not weight by position, sentiment or estimated audience; each adds an assumption that makes the number harder to check.

A worked example.

Here is the method on an invented panel: ten questions, three engines, 30 answers. The result is a score of 42.

The headline score is the average of the three engines: (40 + 25 + 60) ÷ 3 = 41.7, reported as 42 out of 100.

Two figures go beside it: a named rate of 10 in 30 answers (33%) and a cited-only rate of 5 in 30 (17%). Half the answers do not include the company at all.

What the grid tells you that the score cannot

The single number hides three findings. Perplexity names this company more than twice as often as Google does. The company is absent from the “alternatives” question, which usually means no page or third-party list places it against its competitors. And question 7 is cited by all three engines but named by none: a trusted page that does not connect the advice to the company.

That is why the engine scores belong next to the headline. The play on keeping a per-engine scorecard argues for showing leadership the range: a headline of 42 should travel with “25 to 60 by engine”.

The score calculator.

Enter your own counts and the calculator applies exactly the method above: named scores 1, cited only scores ½, each engine is scored out of 100 and the engines are averaged. Nothing you enter is sent anywhere.

Save the inputs with the date. The score becomes useful the second time you run it, on the same questions.

How far to trust the number.

Trust the direction over several months, not the digit in any one month. AI answers change from run to run, so a small panel run once can move ten points without anything real having happened.

The best public evidence is research by SparkToro and Gumshoe.ai, reported by Search Engine Journal on 30 January 2026. The team ran 2,961 prompts across ChatGPT, Claude and Google’s AI answers, each 60 to 100 times per platform. The same list of brands came back “less than 1% of the time”, and the same list in the same order “less than 0.1% of the time”. The report’s more encouraging finding is that “how often a brand appears across many runs of similar prompts is more consistent”.

Google says something similar about its own products: “AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.” The mechanics are in our guide to how ChatGPT picks which companies to name.

Three ways to make your score steadier

• ✓Repeat the runs. Ask each question several times per engine and count every run. The play on counting across repeated runs suggests five as a practical minimum for a small team.
• ✓Know what one answer is worth. With ten questions, one answer is ten points of an engine’s score. With ten questions run five times each, it is two points.
• ✓Freeze the panel. Add new questions as a separate set rather than editing old ones.

The same research found the order of names to be the least stable thing measured, which is one reason our method leaves position out.

What a good score looks like.

There is no universal good score, because a score depends on how hard the questions are. Compare with your own past scores and with named competitors on the same panel.

A company can score high on questions that describe it exactly and low on the broad category questions its buyers ask first. Both numbers are true, which is why we do not publish score bands here. Three comparisons are worth making:

• ✓You against last month. Same questions, same engines, same number of runs.
• ✓You against two or three competitors. Score their names in the same answers with the same sum.
• ✓Engine against engine. A gap of 20 points or more usually has a specific cause, such as a source one engine reads and another does not.

To see where you sit before building a panel, request our free AI visibility report. For a public reference point, the B2B AI Shortlist Index publishes its questions, engines and counting rules in full.

When the score goes to senior people, do not send it alone. Our guide on presenting AI visibility to a board sets out the measures that belong on the same page.

From a score to a plan.

A score is only worth calculating if it changes what you do next. Each pattern in the grid points to a different kind of work.

If you are mostly absent, read why brands are left off AI shortlists. If other sites do the talking, see the off-site sources AI engines rely on.

Connect the score to money

A visibility score is a leading indicator. It says buyers could be seeing your name, not that they enquired. We look at one anonymised client result on the revenue side in a separate article. Our model for AI visibility ROI walks from share of answers to enquiries and pipeline, and the guide to tracing AI referrals through to pipeline shows where those visits appear in your analytics.

Calculate a baseline you can explain, fix the largest gap the grid shows, and run the same panel again a month later. Our answer engine optimisation guide sets out the wider method step by step. If you would rather have that done for you, our generative and answer engine optimisation work starts from the same kind of dated, recorded checks.

Questions owners ask.

What is an AI visibility score?

It is a single number, usually 0 to 100, summarising how often a brand appears in AI-generated answers to a set of questions. Each tool defines it differently: a share of answers, a benchmark against competitors, or a weighted mix.

Why does my score differ between two tools?

Because the tools use different questions, engines, run schedules and formulas. Ahrefs says two tools can report different figures for the same brand “without either being wrong”.

Is there a free AI visibility score checker?

Yes. HubSpot’s AI Search Grader is a free one-time check that scores sentiment, presence, recognition, share of voice and competition. The calculator in this guide scores your own questions with a published formula. Read what each one counts before acting on it.

What is the difference between an AI visibility score and AI share of voice?

A visibility score usually measures how often you appear at all. Share of voice measures your appearances as a share of all the brands appearing, so it always depends on which competitors are being counted.

Is an AI visibility index the same thing?

Not quite. “Index” usually describes a published benchmark across many brands or a composite chart, such as Otterly’s Brand Visibility Index, which plots brand coverage against average position. A score describes one brand on one set of questions.

Sources: Semrush, data sources; Semrush, metrics; Semrush, toolkit; Profound, Answer Engine Insights; Profound, glossary; Peec AI, visibility; Peec AI, position; Otterly, KPI definitions, 5 Oct 2026; Ahrefs Help, metrics; Ahrefs, AI Share of Voice, 22 Oct 2026; Ahrefs, Brand Radar methodology, 26 Feb 2026; HubSpot, AI Search Grader; Search Engine Journal, 30 Jan 2026; Google Search Central, AI features. All accessed 12 November 2026.