Skip to content

AI visibility score: what tracking tools actually measure

An AI visibility score counts brand mentions in a vendor's own sample of prompts and assistants. Here is how those numbers are built and what to ask.

By Crawlito · 9 min read

In short

  • An AI visibility score describes how often a brand appeared in answers to the vendor's own prompt sample, collected the vendor's way, not how often real users saw it.
  • Semrush documents its 0-100 AI Visibility score as a mix of Topic Coverage and Mention Consistency across a prompt database built from clickstream data and Google keyword data.
  • Ahrefs documents that Brand Radar runs prompts in public web interfaces without stored user data, detects mentions by string match, and weights mentions by Google search volume to estimate impressions.
  • Google says third-party tools do not have access to its internal ranking data, while Search Console and Bing Webmaster Tools report first-party impressions and citations for their own AI features.
  • Before paying, ask a vendor where the prompts come from, which assistants and modes are sampled, how often, how a mention is matched, and whether you can see the raw answers.
On this page

An AI visibility score is a summary of how often a brand appeared in AI answers to a set of prompts that the tool vendor chose, ran and parsed. It is not a measure of how many people saw your brand in ChatGPT or Google AI Mode, because no third-party tool can see that. What the number means depends on four choices the vendor made: the prompts, the assistants sampled, how often they run, and how a "mention" is detected and turned into a score.

This post reads two vendors' own methodology pages, sets them beside the first-party reports from Google and Bing, and ends with questions to ask before you buy. It does not rank tools. If you want to run a small test yourself first, the manual brand mention test covers that, and AI citations vs referral traffic explains why mentions, citations and visits should stay separate.

What goes into an AI visibility score

Every tracker in this category follows the same broad pipeline, whatever it calls the result:

  1. Pick prompts. From a prompt database the vendor builds, from prompts you type in, or both.
  2. Run them. In one or more assistants, in a particular mode, location and language, on a schedule.
  3. Store the answers. The raw text and any source links.
  4. Detect your brand. Usually by matching a name or domain in the text, and URLs in the source list.
  5. Aggregate. Turn the matches into counts, shares or a 0-100 score, often relative to competitors.

Each step is a choice that changes the number. Two tools can both be accurate about their own sample and still disagree, because they sampled different questions in different places. That is why the methodology page matters more than the headline figure.

Prompt panels and custom prompts measure different things

The biggest difference between scores is where the prompts come from.

A prompt panel is a large set of prompts the vendor collects and runs for everyone. It lets you look up any brand without setup, and it is what powers market-wide share-of-voice charts. A custom prompt is a question you choose, tracked for your account. It matches your customers' language more closely but covers far fewer questions.

The two vendors named most often in this category document both approaches:

Semrush AI Visibility toolkitAhrefs Brand Radar
Panel prompt sourceAI search clickstream data and Google's keyword dataset for AI Overviews, grouped into topics and de-duplicatedGoogle People Also Ask questions and Ahrefs' keyword database, expanded with related sub-questions
Panel size statedOver 317 million prompts and responsesVolumes given per platform per month
Platforms in the panelChatGPT, Gemini, Google AI Overviews and AI Mode (Brand Performance also lists Perplexity)ChatGPT, Perplexity, Gemini, Copilot, AI Overviews and AI Mode (Grok listed, paused for new data at the time of writing)
Custom promptsPrompt Tracking, run dailyCustom prompts, checked monthly up to daily; Claude via the API, custom prompts only
Refresh statedPrompt data "updated daily on a rolling basis" (the same page's summary says monthly); Brand Performance weeklyChatbot data monthly over a 90-day window; AI Overviews and AI Mode continuously (the help article says every few days)

Sources: Semrush's AI Visibility data page and AI SEO metrics glossary; Ahrefs' Brand Radar methodology and help article. These details were current at the time of writing and change often.

Note what the panel sources imply. Prompts derived from search keywords and People Also Ask questions are search-shaped: short, common, and weighted towards what people type into Google. Ahrefs says plainly that it prioritises common, high-demand questions, so long-tail or niche prompts may be missing. If your buyers ask long, specific questions in a chat window, a panel may barely touch them, and custom prompts will tell you more.

How the assistants are sampled

An answer depends on the product, the mode, the location and the person asking. Vendors handle this differently, and some details are not documented.

  • Interface, not API. Semrush says panel responses are "captured from real requests and not via any APIs of LLMs". Ahrefs says it runs prompts through the free, publicly available web interfaces. Both, in other words, collect what the consumer product returns rather than raw model output. Ahrefs makes an exception for Claude, which it covers only for custom prompts, via the API.
  • No personal context. Ahrefs says prompts run in web versions "without any stored user data or prior context", with no personalization applied. Semrush notes that responses are "fast-changing and highly personalized" but its data page does not say how sessions are set up.
  • Mode. Semrush says its ChatGPT data comes from the latest model in search mode. A tool that samples with web search on measures something different from one that does not.
  • Location. Semrush lists 117 regional databases. Ahrefs mirrors the country and language spread of its keyword database and says coverage is strongest in English.

None of this is a flaw. It is the boundary of the measurement. A logged-out, location-fixed sample is consistent and repeatable, and for the same reason it is not your customer's signed-in session with months of chat history.

How mentions become a score

Once answers are stored, the tool has to decide what counts as you. Ahrefs documents that mentions are string matches for a term and citations are linked URLs, and that it does not filter out hallucinated or malformed links "as they reflect real model output". String matching is simple to audit but has known edge cases: a brand name that is also a common word, a competitor with a similar name, or your product named without your company name. Ask how a tool handles those for your brand.

Then the matches are aggregated. The vendors define their headline numbers differently:

MetricVendor definition (summarised)
Semrush AI VisibilityA 0-100 benchmark combining Topic Coverage (how many topics include your brand in AI answers, compared with all other domains) and Mention Consistency (how often you are mentioned across responses within those topics)
Semrush MentionsThe number of prompts in which a brand is included in AI responses
Semrush CitationsThe number of AI responses that cite your domain as a source
Semrush Share of VoiceThe percentage of mentions your brand receives in AI answers compared with competitors in your market
Ahrefs Share of VoiceBased on how often brands are mentioned or cited; the exact formula is not published
Ahrefs Estimated ImpressionsMentions weighted by Google search volume to model potential exposure

Two points deserve attention. First, a score built from Topic Coverage is relative: it moves when competitors move, even if nothing changed on your side. Second, Ahrefs is candid that its impressions figure is a modelling choice, saying it does not claim "a validated link" between Google search volume and how often a query is asked inside an AI tool, and that Brand Radar shows what appears, not who saw it. Treat any "impressions" or "reach" number from a panel the same way: an estimate of potential exposure in a sample, not an audience count.

What Google and Bing report directly

Two search engines now publish first-party data on their own AI features. These are not visibility scores, but they come from the operator's own systems rather than a vendor's prompt sample, so they are a useful check on a tracker's story.

Google's guidance on third-party SEO tools says they "don't have access to our internal ranking data", that any predictions are their own, and that some tools' data is misread as coming from Google. It encourages using Search Console alongside any tool.

  • Google Search Console has a generative AI performance report that counts impressions, defined as how many times links to your site were shown to a user in AI Overviews or AI Mode. Google says it rolled out to all websites worldwide as of August 31, 2026, and it excludes Search Labs experiments. Our guide to the report covers the views and filters.
  • Bing Webmaster Tools has an AI Performance report, announced as a public preview in February 2026, that counts citations in Microsoft Copilot, AI summaries in Bing and select partner integrations, with grounding queries. Bing says the data shown is a sample of overall citation activity. A June 2026 preview update added Citation Share, the percentage of citations attributed to your site out of all citations shown for the same grounding query. Bing calls it an observational metric, not a ranking. See Bing's AI Performance report for setup.

Neither is documented as covering ChatGPT, Perplexity or Claude, and neither tells you why a number changed. For those assistants, a third-party sample or your own test is the only view of the answers, and your logs and analytics are the only view of what reached your site.

SourceWhat it countsCoversSampled or measured
Third-party trackerMentions and citations in answers to the vendor's promptsWhichever assistants the vendor runsSampled
Search Console generative AI reportImpressions of your linksAI Overviews, AI ModeReported by Google
Bing AI PerformanceCitations of your pagesCopilot, Bing AI summaries, some partnersReported by Bing from a sample of citation activity

Questions to ask a vendor before you pay

Use these on a sales call or against the methodology page. A good vendor answers them in writing.

1. Where do the panel prompts come from, and can I see the prompts behind my score?
2. Can I add my own prompts, and how often are they run?
3. Which assistants, modes (web search on or off), models and locations are sampled?
4. Are answers collected through the public interface or an API, signed in or out?
5. How is a mention matched? How are similar names and products handled?
6. Is the score relative to competitors? Which competitors, and who chose them?
7. Can I export the raw answers and source URLs for each run?
8. How many runs sit behind each data point, and how is run-to-run variation shown?
9. When the methodology changes, is history recalculated or marked as a break?
10. How do your numbers compare with my Search Console and Bing Webmaster Tools data?

Questions 7 and 8 matter most. A score you cannot trace back to saved answers is hard to trust and impossible to debug. If your number drops, you need to see which prompts changed and what the assistant said instead. When you report the figures internally, follow the denominator rules in the AI search reporting dashboard guide.

Finally, be careful with any vendor that promises the score will rise if you buy a service. Google's guidance on third-party tools makes a similar point for search: using a service or tool "doesn't guarantee ranking success". No tool controls what an assistant chooses to say.

Frequently asked questions

Why do two AI visibility tools give my brand different scores?

They sample different prompts, assistants, locations and dates, and they define the score differently. Each can be correct about its own sample, so compare the methodology, not just the number.

Is share of voice in AI answers the same as market share?

No. It is your share of mentions or citations within the vendor's prompt sample, usually against a chosen competitor set. It says nothing direct about sales or audience size.

Does an AI visibility score come from ChatGPT or Google data?

No. Third-party tools run prompts and parse the answers themselves. Google says third-party tools do not have access to its internal ranking data; first-party numbers come from Search Console and Bing Webmaster Tools.

Can I trust estimated impressions in an AI tracker?

Treat them as a model of potential exposure. Ahrefs, for example, weights mentions by Google search volume and says that link to AI usage is not validated.

Is a readiness score the same thing as a visibility score?

No. A readiness score checks whether your pages can be reached and read, and how the Crawlito score works explains one: it is computed by rules and measures published checks, not the chance of being cited.

Rule out access problems before reading the trend

A falling visibility score can have a plain technical cause, such as a robots.txt change or a firewall challenge that stops AI search crawlers from reading your pages. Before you debate methodology, run the free check on an important page to confirm the main AI crawlers can fetch and read it.

Filed under Strategy and audits. Topics: analytics, ai-search, measurement, strategy