Share of answer: how to compute it, and where it misleads

AI visibility is often summarized as one percentage per brand. It is easy to read and easy to misread: with a small sample, one answer moves it by a large step, and the count behind it shows how much to trust it. This post sets out the definition, the arithmetic, and the points where the number stops meaning what it appears to mean.

What is share of answer?

Share of answer = answers that recommend the brand ÷ answers sampled, for a stated assistant, a stated set of questions and a stated period.

Three choices inside that sentence change the number, so state them every time:

• Recommended or mentioned. The Dopbot Index counts a brand only when the answer presents it as a pick. A brand named as a comparison or a warning is a mention. Counting mentions gives a higher share and a different meaning.
• Which answers are in the denominator. All answers for a category in an update, across all five question types, or only one type such as budget questions.
• Which assistant. OpenAI and Perplexity are reported separately. In this post, "OpenAI" means OpenAI's API with web search (gpt-6-luna) and "Perplexity" means Perplexity's Agent API with the Sonar model (perplexity/sonar), the two engines of the October 2026 edition.

Share of answer is not share of voice. Share of voice divides a brand's mentions by all brand mentions, so shares across brands add up to 100%. Share of answer divides answers by answers, and an answer can name several brands, so the shares across brands add up to more than 100%. In October 2026, excluding the alternative-to-a-brand question (1,271 answers), an answer recommended 2.3 brands on average, so, across those answers, the brand shares in a category add up to about 230% on average (2.3 brands per answer x 100).

How do you compute it?

1. Fix the questions. Write them once and keep them. Changing a prompt starts a new series.
2. Ask each question on each assistant, the same way every time: same location setting, no logged-in history.
3. Record each answer's recommended brands, each brand once per answer.
4. Count. k is the number of answers that recommend your brand, n the number of answers sampled.
5. Compute the share k ÷ n.
6. Compute an interval. The Wilson score interval is one common choice for a proportion from a small sample. With p = k / n and z = 1.96 for 95% confidence:

• center = (p + z^2 / (2n)) / (1 + z^2 / n)
• half-width = z * sqrt(p(1-p)/n + z^2/(4n^2)) / (1 + z^2 / n)
• The interval runs from the center minus the half-width to the center plus the half-width.

The methodology uses the 95% Wilson interval, and its badge rule is written in terms of the lower end of the 95% Wilson interval. Unlike the simple "plus or minus" formula, Wilson stays between 0% and 100% and behaves sensibly when k is 0 or close to n.

The interval treats answers as independent and covers only the randomness of asking these questions again. Several samples of the same prompt are likely to resemble each other, so the true uncertainty can be wider, and it says nothing about other wordings.

What does a worked example look like?

Take a made-up brand, Brand A. These counts are illustrative, not Dopbot data. Suppose it is recommended in 4 of 10 answers, a 40% share. The interval is 16.8% to 68.7%. If the true rate were anywhere in that range, a sample like this would not be surprising.

Now keep the share at 40% and change only the sample size:

Table: made-up counts at four sample sizes. 5 is one engine's answers per category in the October 2026 edition and 10 is both engines; 20 and 40 are the same for the monthly update of 40 US categories (see the methodology).

Quadrupling the sample from 10 to 40 narrows the interval from 51.9 to 29.1 points, and it is still wide. Two brands do not separate easily either. At 10 answers, 4 of 10 (16.8% to 68.7%) and 6 of 10 (31.3% to 83.2%) overlap almost completely. At 40 answers, 16 of 40 and 24 of 40 (44.6% to 73.7%) still overlap. Overlap is a cautious check, not a significance test: a direct two-proportion test on 16 of 40 against 24 of 40 gives p of about 0.07, so the evidence is suggestive, not settled.

If each share came from its own sample, the Wilson intervals for 40% and 60% first stop overlapping at about 100 answers per brand (97 for exactly 40% and 60%). That is a strict test: intervals that overlap slightly can still hide a real difference, and a standard two-proportion test would call 40% against 60% different from about 50 answers per brand. Either way, 10 or 40 answers cannot settle it.

Zeros need the same care. A brand recommended in 0 of 10 answers is consistent with a true rate up to 27.8%. At 0 of 40, the upper end falls to 8.8%.

A real case from the data, as an example and not a finding: in the air purifier category the most recommended brand appeared in 6 of 8 answers (excluding the alternative-to-a-brand question, October 2026, both engines). That is 75%, with an interval of 40.9% to 92.9%.

What is one answer worth?

The step size is 100 divided by n. At the October 2026 edition's 10 answers per category, one answer is 10 points of share. In the monthly update of 40 US categories, with 40 answers per category, one answer is 2.5 points. Per engine the steps are 20 points (5 answers) and 5 points (20 answers). The methodology states the same: at 40 answers, a brand recommended in 20 has a true rate of about 35% to 65%.

So a move from 30% to 40% in a 10-answer category is one answer. It is inside the noise of asking the same questions again.

How noisy is the most-recommended brand?

Small samples exaggerate whoever happens to come out on top. To see how much, we simulated a category in which nothing real favors any brand: 11 brands, each named in any one answer with the same probability (2.3 ÷ 11, about 21%), so that answers recommend 2.3 brands on average and a category shows about 9.3 distinct brands in 8 answers. That matches the October 2026 sample, where categories averaged 9.2 distinct brands and 7.9 answers (excluding the alternative question, 160 categories). The simulation ran 200,000 categories.

Table: a simulation under equal chances, beside the October 2026 numbers (excluding the alternative-to-a-brand question). The simulation assumes independent answers and equal brands, which real answers are not.

Two readings. First, at about 8 answers per category, the most recommended brand reaches half or more of the answers in about 52% of categories even when no brand is favored, so one category's top brand tells you little by itself. Second, the real sample is more concentrated at the top (62.3% vs 45.1%) and has more one-answer brands (48.4% vs 38.2%) than equal chances would give, so the sample is more concentrated than this equal-chance model produces. The index cannot say why. With 40 answers per category, equal chances would almost never show a brand in half of the answers, so a brand that does is more likely to be a real pattern. This is why the index reports counts rather than ranks, and why the update samples more.

Where does share of answer mislead?

• Prompt choice. The share belongs to the questions you asked. In the October 2026 edition, the brand lists of the budget and premium answers overlapped by 8.2% per category on average (Jaccard, usually 2 budget and 2 premium answers per category, 155 of 160 categories). A share measured on one question type is therefore not a good guide to another.
• Assistant choice. The two engines' brand lists overlapped by 47.0% on average per category (Jaccard, equal weight per category, excluding the alternative question). Averaging a brand's share across engines hides which assistant it appears in; report each.
• API versus app. Perplexity's documentation says its API uses the same search system as its interface, with differences in configuration, so outputs may differ. See API answers vs app answers.
• Endpoints change. The October 2026 edition was collected from October 23, 2026. Its Perplexity answers were requested through Perplexity's Agent API with the model string perplexity/sonar, which Perplexity's models page lists as of October 2026. Perplexity's documentation says support for its older Sonar Chat Completions endpoint ended on September 27, 2026, and that synchronous and streaming requests keep working while they are reformulated as Agent API requests, rolling out gradually by model. The index does not use the older endpoint. Record the endpoint and model string with every sample, so a later change is visible.
• Classification error. A model reads each answer to decide which brands it recommends. Name variants and ambiguous phrasing cause misses, so a hand-kept merge list and spot checks matter.
• Presence is not influence. A source in an answer's list does not show that it caused the recommendation, and share of answer says nothing about sales.
• Change over time. A share from a 10-answer sample compared with one from a 40-answer sample is not like for like. For change, see how many answers to trust a change.

How should you report it?

Show the count, the denominator, the assistant, the question set and the interval, not only the percentage:

Table: a made-up report line, to show the layout. The interval assumes independent answers; repeated samples of one prompt are likely to resemble each other, so it can understate the uncertainty.

Add the date, the prompt text (or a link to it) and the number of failed answers. If you only have the budget to sample once, say so. For a step-by-step manual version, see how to measure your brand's AI visibility, and for the term itself, the glossary.

Frequently asked questions

Can I average share of answer across assistants?

Only with stated weights, and not as the headline. The assistants return different numbers of brands, record their sources differently, and their brand lists overlap by 47.0% on average per category (Jaccard, excluding the alternative-to-a-brand question), so one blended number hides where a brand is and is not recommended.

Should I count mentions that are not recommendations?

Count them separately. A mention in a warning or a comparison is a different event from a pick, and merging them raises the share but counts something other than a recommendation.