How we measure AI Visibility

Last updated: 29 September 2026

This page explains where the AI Visibility numbers in Miraqo come from. A measurement of AI assistant answers is only as good as the method behind it, so the method is written down here.

What we query

For each monitored prompt we query up to five sources: Google AI Overviews, Google AI Mode, ChatGPT, Gemini and Perplexity. We use the public versions of the assistants, with the model each one offers by default to an anonymous user. We do not pick the model: when the assistant’s vendor changes the default, the measurement changes with it. That is why we record which model answered each run.

Web search is allowed but not forced: the assistant decides whether to use it, as it would for a real user. If it answers from memory in every run, citing no source at all, we flag it as an answer from memory. In that case a zero in citations only means the assistant did not look at the web.

Claude and Copilot are not among the sources yet.

Where

Each prompt has a country: the request starts from that country and in the language of the prompt. The same prompt can be monitored in several countries as separate rows. AI Overviews and AI Mode are detected where Google actually shows them. If Google generates no AI answer for a query, we record that the answer was missing, and we do not count it as the brand being absent.

How many times

AI assistant answers are not deterministic: the same question, asked twice, can cite different brands. A single answer is not enough to say whether a brand is present. So each prompt runs three times per source at every check, and the measurement is the Share of Voice: the percentage of runs in which the brand appears. Three runs give a scale in steps of one third (0, 33, 67, 100%). That is enough to tell a stable presence from an occasional one without multiplying the cost for whoever pays for the plan.

If a run does not come back, it does not enter the denominator: 2 out of 2 returned runs is 100%, not 67%.

What we count

  • Mention: the brand name appears in the text of the answer.
  • Citation: a link to the brand’s site appears among the answer’s sources, with its position in the list. We recognise the domain even with a different extension (.com and .it) and ignore URL parameters.
  • SERP/AI overlap: the assistant cites the very page that ranks on Google for the keyword linked to the prompt. A prompt with no linked keyword cannot be compared and stays out of the count, without counting as zero.
  • Brands cited: every brand named in the answer, extracted once per check from the representative answer.
  • Prompt intent: informational, comparative, commercial or problem-solving. You can set it when you add the prompt; if you do not, a language model assigns it by reading the prompt text.
  • AI Authority: a score from 0 to 100 with declared weights. Mentions 40%, citations 30%, recommendations as a solution 20%, average position among cited sources 10%, averaged over the monitored sources.
  • Visits from AI assistants: sessions that reached the site with a referrer from ChatGPT, Perplexity, Gemini, Copilot or Claude, read from the project’s Google Analytics 4 or Matomo connection. AI Overviews and AI Mode stay out: Google does not separate that traffic from organic search.
  • AI crawlers on the site: for sites behind Cloudflare, the requests from the ChatGPT, Perplexity, Claude and Gemini crawlers to the pages, read from the edge logs with a read-only token. They are Cloudflare estimates (adaptive sampling) and the user agent is not verified; a request rejected at the edge (WAF, rules) is distinct from a robots.txt disallow, which the audit reads.

Mention and citation are counted separately: an assistant can name a brand without linking it, or link the site without naming the brand.

When

You choose the frequency per project: monthly, every two weeks, weekly or every three days. The check updates all the project’s active prompts together. A newly added prompt is checked right away, then joins the project’s cycle. A prompt already checked in the last 36 hours is not run again. The “next update” date says when the data will change, not when our system runs.

How we build the charts

Each point on the chart is the state of the prompts on that day, that is the last known result of each one, and not only the checks made on that date. A prompt added today does not redraw the curve, and days without checks do not break the line.

When a source does not answer, the data is not invented. The table keeps the last good value with its date and a warning. In charts and averages the result expires after twice the chosen frequency (seven days at least) and leaves a gap. We do not interpolate and we do not repeat the last value, and we do not write zero. When the detection method for a source changes, the chart marks it with a dashed line, so a jump is not read as a result.

What we do not estimate

Prompt volume. There is no public measurement of how many people ask a given question to an AI assistant; the numbers in circulation are modelled estimates. The volume we show is that of the linked Google keyword, labelled as such.

An opaque score. Every AI Visibility number traces back to runs, answers and sources you can open and read in full.

Limits we know about

Three runs measure stability at a coarse grain. A Share of Voice of 33% on a single check can be noise; the series over time says more than a single point.

Very long answers are analysed within the first 12,000 characters for the mention in the text. This happens mostly on AI Mode in German: in those markets the mention may come out slightly low. Citations have no such limit.

Assistants change model and behaviour without notice. We record the model of every run so that a jump in the series can be explained.

If something does not match what you see in your project, write to us and we will check it.