How to Measure AI Visibility: Metrics, Sampling, Mistakes

To measure AI visibility, you ask AI assistants a fixed, representative set of real questions, repeat them enough times to absorb randomness, and count how many valid answers mention or cite your brand. The core metric is a proportion with an honest denominator, reported with its uncertainty, and compared against competitors measured on exactly the same sample. Everything else, from share of voice to position, is a refinement of that basic design.

Key takeaways

The unit of measurement is the individual answer, not a "ranking": AI assistants do not keep a stable list of positions.
Presence rate = answers mentioning the brand ÷ valid answers. Failed, empty or refused answers belong outside the denominator, not counted as zeros.
Because outputs vary from run to run, a single screenshot proves very little; you need dozens to hundreds of answers and a confidence interval.
Competitor comparisons are only meaningful when every brand is scored on the same questions, models, market, language and time window.
Be wary of opaque composite "visibility scores": if you cannot recompute a number from the stored answers, you cannot audit it.

What exactly are you measuring?

AI visibility describes how often, and in what way, a brand appears in answers generated by assistants such as ChatGPT, Gemini, Claude or Perplexity. A generated answer has no fixed slots: the same question can yield five vendors one minute and one recommendation the next. The most robust unit of observation is the individual answer, stored in full together with the sources it cites, the model used, the language, the market and the timestamp.

From that raw material you can derive several distinct signals. They answer different questions and should not be merged too early:

Mention: does the brand name appear in the answer text at all?
Citation: does the answer link to or reference a page from your domain as a source?
Prominence: where does the brand appear (first sentence, first item in a list, a footnote)?
Framing: is the brand recommended, merely listed, compared unfavourably, or described inaccurately?

Research uses similar ideas: Aggarwal et al. (KDD 2024) measure visibility in generative engines with a position-adjusted word count and a subjective impression score. Visibility is multidimensional, so decide in advance which dimension your decision depends on. For background on the discipline itself, see what GEO is.

The core metrics and how to calculate them

Metric	Formula	What it tells you
Presence rate	answers mentioning the brand ÷ valid answers	How often you show up at all
Citation rate	answers citing your domain ÷ valid answers	Whether AI treats your content as evidence
Share of voice	your mentions ÷ all tracked brand mentions in the same sample	Your slice of the conversation versus competitors
First-mention rate	answers where you are named first ÷ answers mentioning you	Prominence once you are present
Accuracy rate	answers with correct facts about you ÷ answers mentioning you	Whether visibility is helping or hurting

The presence rate is usually the headline number because it is simple and easy to audit. Share of voice is useful for competitive context, but note its dependence on which competitors you track: add a new rival to the list and your share falls even if nothing changed in the answers.

Why the denominator matters

Imagine you send 100 questions. Eight return errors, time out or produce an empty response. Of the 92 valid answers, 23 mention your brand. The honest presence rate is 23 ÷ 92 = 25.0%. Dividing by 100 would report 23%, silently treating technical failures as "the AI ignored you". Failure rates differ by model and by day, so a sloppy denominator can create fake trends. Store failures, report how many there were, and exclude them from the ratio.

Sampling: how many questions, and how often?

Language models are not deterministic in practice. Empirical studies show that even with temperature set to zero, outputs can differ between identical requests (Atil et al., 2024; Ouyang et al., 2023). Web retrieval, personalisation and model updates add more variation. A measurement is therefore a sample from a distribution, and it should be treated statistically.

Designing the question set

Start from real intents. Collect questions customers actually ask: category ("best accounting software for small businesses"), comparison ("X vs Y"), problem ("how do I automate invoices?") and local or language-specific variants.
Avoid branded prompts in the main metric. "Is BrandX good?" will almost always mention BrandX; it measures reputation, not discoverability. Track it separately.
Fix market and language. An answer in Turkish for users in Turkey can differ sharply from an English answer for the US. Mixing them blurs both.
Cover each model you care about. ChatGPT, Gemini, Claude and Perplexity use different retrieval and training data; report them separately before aggregating.
Repeat. Ask each question several times per period so that one lucky or unlucky run does not dominate.
Freeze the set. Keep the same core questions over time. Add new ones in a separate segment so the trend line stays comparable.
Store everything. Keep the full answer text and cited URLs, so any number can be recomputed and inspected later.

Sample size and confidence intervals

A proportion based on few answers is very noisy. With the common normal approximation at 95% confidence, the worst-case margin of error is roughly ±13.9 percentage points with 50 valid answers, ±9.8 with 100, ±6.9 with 200 and ±4.9 with 400. For small samples or rates near 0% or 100%, the Wilson score interval behaves better than the simple formula. In the example above, 23 out of 92 gives a Wilson 95% interval of about 17% to 35%. With only 4 out of 20, the interval runs from about 8% to 42%, which is too wide to support any decision.

The practical rule: before reporting that visibility "rose from 25% to 30%", check whether the two intervals overlap substantially. If they do, the honest conclusion is "no clear change yet".

Comparing yourself with competitors

Competitive benchmarking is where most measurement errors become expensive. A fair comparison requires that every brand is scored on the same answers: same questions, same models, same market and language, same dates. Then each brand's presence rate shares one denominator and differences are real differences in the answers, not artefacts of separate runs. This is the logic behind same-sample competitor comparison.

Also look at the sources behind mentions: if competitors are cited from review sites or media where you are absent, that is a concrete lead. Our guide on which sources AI cites explains how to read those patterns, and how ChatGPT recommends brands covers the signals that tend to sit behind recommendations.

Common measurement mistakes

Treating one answer as the truth. A screenshot is an anecdote, not a measurement.
Counting failures as zeros. Errors and refusals are not evidence of invisibility; exclude them and disclose them.
Inventing a composite score. Blending presence, sentiment and position into a single "AI score" with undisclosed weights makes the number impossible to audit or reproduce.
Comparing across different samples. Your figure from March against a competitor's figure from a different tool, question set or language tells you almost nothing.
Drifting question sets. Changing questions every month turns a trend chart into a chart of your own edits.
Ignoring interface differences. API and consumer chat results can differ; state which one you measured.
Confusing visibility with outcomes. Being mentioned does not guarantee clicks or sales. Treat AI visibility as a leading indicator and connect it to referral traffic and conversions where you can.

A transparent methodology that spells out how answers are collected, which ones count as valid and how each ratio is computed is the best protection against all of these. GeoDog, for example, stores every answer with its sources so each figure can be verified.

FAQ

How many questions do I need to start?

For a first read, 20 to 50 well-chosen intent questions per market and language, each repeated a few times across the models you care about, typically give you on the order of 100 to 200 valid answers. That is enough to see large gaps; detecting small changes requires more.

How often should AI visibility be measured?

Monthly measurement suits most brands, because models and retrieval indexes change gradually and noise dominates day-to-day differences. Measure more often around launches, major content changes or known model updates.

Does a high presence rate guarantee traffic?

No. An assistant can mention a brand without linking to it, and users may not click. Presence rate measures exposure in answers; combine it with citation rate and analytics data to understand business impact. You can run a first scan to see where you stand today.

How to Measure AI Visibility: Metrics, Sampling and Common Mistakes