How to measure AI visibility without mistaking noise for progress
A report says your AI visibility rose from 30% to 60%. That sounds encouraging, but what doubled: the proportion of answers mentioning your brand, your share among competitors, or the number of customers?
Start with what was counted. An AI visibility score describes the answers you collected. Check the denominator—the total you divide by—and how the answers were collected. These tell you what the percentage means and whether two results can be compared.
Decide what to count before choosing the metric
To measure how often a brand is mentioned, count completed AI answers. Record whether each answer names the brand. If the name appears three times in one answer, that still counts as one answer containing the brand.
Under Supasaito’s visibility definition, the calculation is:
Visibility = completed answers containing the brand ÷ completed measured answers × 100.
A mention simply names a brand. A recommendation suggests choosing it. A citation links to a page presented as a source for the answer. Count these separately. An answer can name a brand without recommending it. A cited page is a source link, not automatically a recommendation of the organisation behind it. If your question concerns recommendations or citations, define a separate counting rule rather than treating every mention as the answer.
Keep unavailable measurements separate. A completed answer that does not mention the brand counts towards the total, but not towards answers mentioning the brand. A failed answer or disabled model gives you no completed answer to count. If no answers completed, there is no denominator from which to calculate visibility. Supasaito’s Prompts & answers help explains how to inspect those statuses and the dates behind results.
Visibility and share of voice answer different questions
Visibility asks: in what proportion of the measured answers did this brand appear?
Mention-based share of voice asks: what share of the appearances of a defined group of brands belonged to this brand?
For the second calculation, count each comparison brand at most once per answer. Divide the target brand’s appearances by the total appearances of all brands in that comparison group.
Consider this illustrative example, not a customer result:
| Brand | Answers containing the brand |
|---|---|
| A | 30 |
| B | 50 |
| C | 20 |
There are 50 completed answers, and several brands can appear in the same answer. Brand A therefore has:
- 60% visibility: 30 ÷ 50.
- 30% mention-based share of voice among A, B and C: 30 ÷ (30 + 50 + 20).
Both figures are correct. Visibility divides by 50 answers; share of voice divides by 100 brand appearances. If no comparison brand appears, this share of voice is undefined.
The comparison group matters: changing which brands are included changes the question the share answers. State that group whenever you report the result. The share-of-voice glossary explains this method and its conditions.
Do not assume that every tool uses this formula. Ahrefs Brand Radar uses “impressions” weighted by search demand to calculate AI share of voice. Weighting means that answers associated with more search demand contribute more to the total, rather than every brand appearance counting equally. These impressions add up Google search volumes associated with prompts whose answers mention the brand; they do not count people who actually viewed those AI answers. The result can therefore differ from the equal-count method above. Before comparing tools, check what each counts, how much each appearance contributes and which brands are included.
Keep presence, position and source evidence separate
Visibility tells you whether the brand appeared. It does not tell you where it appeared or whether a linked source supports what the answer said.
Supasaito reports average position separately, as the brand’s position among brands mentioned in measured answers. It does not weight visibility and is not a Google search ranking. Read the underlying answers before interpreting the position.
Likewise, open a citation and find the passage that supports the claim in the answer. A link on the same topic is not sufficient support. Check three things: is there a source link, does the linked passage support the claim, and is the claim factually correct? These checks help explain the answer without redefining its visibility score.
Build a comparison you can describe
A prompt is a question or instruction sent to the AI. A prompt panel is the set of these questions used for the measurement. Its scope is part of the result, not a detail to hide below the chart.
Before comparing measurements, retain:
- the questions and counting rule;
- the language, AI provider and search mode—for example, whether web search was enabled;
- the model or version and available settings, where known;
- collection dates and the completed, failed or unavailable results;
- how many questions were covered and how often each was repeated;
- the brands included in the comparison.
Do not invent a model version when it is unavailable. Record the gap and consider how it limits the comparison.
Breadth and repetition serve different purposes. Adding questions changes what the panel covers. Repeating a question lets you observe variation under its recorded conditions. More repeats of one question do not broaden the question set, and answers to the same or closely related questions should not automatically be treated as separate, independent checks of the result.
Research by Blackwell, Barry and Cohn examines repeated benchmark evaluations: tests that assess large language models (LLMs), the models behind AI assistants, on specified tasks. The study looks at how results vary between runs and how confidently they can be interpreted. It supports documenting conditions and variation. It does not tell marketers how many questions or repeats to use, or how large an increase must be to count as an improvement.
A useful comparison therefore explains what stayed comparable and what changed. If you added prompts, switched modes or changed the comparison brands, do not silently present the new percentage as the same measurement.
Report a change without claiming its cause
Suppose an illustrative panel contains a brand in 6 of 20 completed answers in one round and 12 of 20 in another. Assume the panel and counting rule are unchanged and the collection conditions are documented as comparable.
The observed visibility rises from 30% to 60%. The difference between the two percentages is 30 percentage points: 60 − 30. The second proportion is also twice the first: 60 ÷ 30 = 2.
Put the calculation and its conditions together in a reporting sheet. This example is fictional; fields that have not been supplied remain unknown.
| Report field | Illustrative entry |
|---|---|
| Measure | Mention-based visibility: answers containing the brand ÷ completed measured answers |
| First round | 6 of 20 completed answers; 30% |
| Second round | 12 of 20 completed answers; 60% |
| Observed change | +30 percentage points; twice the initial proportion |
| Questions and counting rule | Same panel and rule in both rounds, as stipulated for this example |
| Provider, model, search mode, language and dates | Not supplied; record the actual values, or identify what is unknown, in a real report |
| Unavailable results | Not supplied; list failed or unavailable results separately |
| Comparability | Assumed for this illustration; check against the recorded conditions in a real comparison |
| Conclusion supported | The brand appeared in a larger proportion of the collected completed answers |
| Conclusions not established | Whether the increase lasts, whether a statistical test can distinguish it from answer-to-answer variation, what caused it, whether it applies beyond these questions, or whether customers increased |
| Next evidence | Collect more answers under comparable conditions to see how the figures vary; check website visits, enquiries and customer records separately for business results |
For a real report, attach the questions and the underlying answers so the reader can check the counts. Explain any changed conditions alongside the result. If you offer a possible reason for the movement, label it as an interpretation and state what evidence would test it.
Connect the result to the next question
An AI answer naming your brand, a recorded website visit, an enquiry and a confirmed customer are different things to count. Measuring one does not measure the others.
AI search can produce website visits: Google’s documentation describes traffic from AI Overviews and AI Mode within overall Search Console Web reporting. Those combined figures do not identify every AI visit separately or show whether AI search brought visits or customers you would otherwise not have gained.
Keep answer measurements beside separately defined visit, enquiry and customer records. Use the reporting sheet to identify which observation is still missing for the business question.
For the broader reporting context, see AI-answer metrics for executive reporting. The practical next step here is narrower: show which answers were counted, what total you divided by and how the answers were collected for every visibility figure. Then state the next evidence needed for the conclusion you want to draw.
A useful report lets the reader distinguish three things: what was observed, what could explain it and what remains unknown.
