How are mention rate and citation share calculated?

Mention rate is the count of valid answers where the brand appears divided by all valid answers; for a brand appearing in 50 of 200 answers, that is 25%. Citation share is the count of answers linking to your own domain divided by answers that contain at least one web source; in the example that is 20 / 150 = 13.3%. When the denominator changes the result changes, so the report should state which denominator was used.

How do cached answers and technical errors affect measurement?

If the same answer is served again from cache, it should not be counted as a new answer; otherwise the sample looks like it grew when it did not. Attempts that returned no result due to a technical error should also not be treated as valid answers where the brand was absent. If failed attempts are not reported separately, a provider that errors more often will make the brand look less visible than it is.

Why do small samples show large percentage swings?

In a sample of 20 answers, one additional appearance moves the rate by 5 percentage points, while in 200 answers the same appearance moves it by only 0.5. That is why a single answer can swing the score heavily in small samples. When you grow the sample, read providers and prompt sets separately, and make the period, new-answer count, prompt count, and coverage change visible in every report.

How to Measure AI Visibility: The 7 Metrics That Matter

Q: Is AI visibility measured only by how many times a brand is mentioned?

No. A mention count alone is an incomplete picture; a brand can appear in an answer but not be recommended, or be described without any source link back to its site. A useful report defines share of voice, mention rate, citation share, answer position, prompt coverage, sentiment distribution, and recommendation rate separately. Each of these seven metrics reflects a different need.

Measuring AI visibility is more than counting how many times your brand is mentioned. A brand can appear in an answer, not be recommended, or be described without any source link back to your site. Each of these outcomes reflects a different need.

A useful report defines share of voice, mention rate, citation share, answer position, prompt coverage, sentiment distribution, and recommendation rate separately. When the formulas are clear, you can understand why a score changed and decide what to improve.

Start by setting a shared measurement baseline

The calculations in this guide rely on a single hypothetical example: assume we collected 200 fresh, valid answers across 40 different prompts. The brand is mentioned in 50 answers, recommended in 30, and receives a source link to its own site in 20. A total of 150 answers contain at least one web source. These are not any customer's actual results.

If the same answer is served again from cache, we do not count it as a new answer. Nor do we treat an attempt that returned no result due to a technical error as a valid answer where the brand was absent. Failed attempts need to be reported separately; otherwise a provider that errors more often can make the brand look less visible than it really is.

In this guide we define share of voice as your share of the answer-level mentions across the tracked brands. Each brand is counted at most once in the same answer.

Formula: The number of answers where your brand is mentioned ÷ the total answer-level mentions across all tracked brands.

In our example the brand is mentioned in 50 answers; if the total mentions across all tracked brands is 250, the share is 20%. Because more than one brand can appear in a single answer, it is possible for 250 to be greater than the 200 answers.

This metric is sensitive to the competitor list. If you add five brands to the list, your share can fall even if your own mentions are unchanged. Compare against the same competitor set from the prior period as well.

2. Mention rate: in how many answers do you appear?

Formula: Valid answers where the brand is mentioned ÷ all valid answers.

In the example, 50 ÷ 200 = 25%. Repeating the brand name within the same answer does not raise the rate. A similarly named company or another meaning of the word should not count as a mention of your brand.

Track this rate separately for branded and non-branded prompts. Appearing when you have already put the brand into the prompt is not the same success as being recommended on your own in a category question.

Formula: Answers that link to your own domain ÷ answers that contain at least one web source.

In the example, 20 ÷ 150 = 13.3%. If you want all valid answers as the denominator, the result is 20 ÷ 200 = 10%. Both calculations are usable; the report heading should state the denominator. Here we use the first as the "citation share among answers that contain a source."

Alongside the domain rate, track the target page too. Your home page being a source does not mean the pricing guide you just published is a source. For this distinction you can look at source analysis.

4. Answer position: where do you appear in the list?

In clearly ordered option lists, record the brand's position. For example, across the eligible lists where the brand appears, the average position might be 2.4. When you arrive at this number, state which answers were accepted as ordered lists.

Do not always treat the first brand name mentioned in a free-form paragraph as the "top recommendation." And when the brand does not appear at all, do not assign an imaginary last place. Absence belongs in the mention rate; position in an ordered list should stay in a separate metric.

5. Prompt coverage: for which needs do you appear?

Formula: Distinct prompts where the brand appears at least once ÷ distinct tracked prompts.

If the brand appears in 15 of 40 prompts, coverage is 37.5%. This calculation prevents appearing frequently in just a few prompts from dominating the report. Still, appearing once is not the same as appearing regularly; also check how many different days each prompt produced a positive result.

The day you add new prompts, the coverage rate can change. Separating new and old prompt sets lets you read the change correctly.

6. Sentiment distribution: how are you described?

Suppose you classified 35 of the 50 answers where the brand appears as positive, 5 as negative, and 10 as neutral. The suggested net sentiment calculation is (35 − 5) ÷ 50 = 60%.

Add example text next to the labels. Describing a product's limitation in a neutral way is not always negative sentiment. If you use automatic classification, have a small sample reviewed by a human and do not change the definition between periods.

7. Recommendation rate: does being mentioned mean being chosen?

Among the answers where the brand is mentioned, the recommendation rate is 30 ÷ 50 = 60%. Across all valid answers, the recommendation rate is 30 ÷ 200 = 15%. Do not present these two results under the same name.

Distinguish a positive product description from a recommendation aimed at the user's need. The sentence "X is a product like this" is not the same behavior as "you could consider X for this need of yours."

Why do a single measurement and small samples mislead?

In 20 answers, one additional appearance moves the rate by 5 percentage points. In 200 answers, the effect of the same additional appearance is 0.5 percentage points. That is why it is easy to see large percentage swings in small samples.

As you grow your sample, do not blindly combine different platforms into a single total. Read providers and prompt sets separately first. Serving the same answers again does not truly grow the sample either. In every report, make the period, new-answer count, prompt count, and coverage change visible.

Tie the score to the work to be done

If mention is high but citation is low, review your own source pages. If citation is high but recommendation is low, look at the content's role in the user's decision. If brand information is wrong, check product descriptions and third-party sources. If visibility rises but revenue does not change, do not interpret this alone as failure or success; visits and conversion data should be measured separately.

When you review Maya's approach to AI visibility tracking, evaluate not just the score but the answers and sources that produce it. The value of a report is being able to explain which concrete decision each number helps with.