Methodology 2026.10.2, in force since 23 October 2026. Every answer and every number records the version it was made under.
We ask AI assistants the questions your customers ask, many times, and read every answer. Every figure is a share of valid answers: answers that were collected and not empty. Failed or empty answers are left out of every count.
Mentioned: the answer names the business.
Recommended: the answer recommends the business, backed by a sentence in the answer that says so.
Your site cited: the answer cites a page on one of the business's own domains.
In the first three named: the business is among the first three businesses the answer names. We never call this a rank.
Figures cover the last 28 days, for each question and each assistant, and all of them together. Competitors are shown side by side, each with its own figures; we do not compute a share of voice.
Today every answer comes from an assistant's official API with web search turned on. API answers can differ from what people see in the consumer apps, and every answer says how it was collected. See the assistants page for each one.
Where a licensed data provider can collect what people see in the consumer apps, we use it, and label those answers too. We never use anyone's personal account.
Each question is asked on a schedule (daily for most assistants), at a randomly chosen time within each period, and stored with its full text, the sources it cites and its cost.
A language model reads each answer and lists the businesses it names, whether it recommends them, and the sentence that shows it.
Then fixed checks, not the model, decide what counts: a name must appear in the answer; a recommendation and a tone need a sentence quoted word for word from the answer that names the business. Claims that fail the checks are dropped and kept for review. Names with and without accents are different businesses.
An answer that says it could not find, does not know or has no information about a business does not count as a mention of it.
Every figure comes with its 95% interval and the number of answers behind it, for example “31% (±5, 336 answers)”. We use the Wilson score interval:
(p + z²/2n ± z·√(p(1−p)/n + z²/4n²)) / (1 + z²/n), z = 1.96
We show “insufficient data” instead of a number when there are fewer than 30 answers, or when the interval is wider than ±15 points.
Figures for all assistants together add up the answers of each, so assistants with more answers weigh more; the figures for each assistant are always shown beside them.
We compare the last 28 days with the 28 days before using a two-proportion z-test. A change is significant only when p < 0.05 and it is at least 5 percentage points. Anything else is described as within normal variation.
Only significant changes appear in change badges and in the weekly email.
Asking the same question many times does not give fully independent answers, so our intervals are somewhat narrower than they should be. We say so rather than hide it, and the calculation is built so a method that accounts for this can replace it.
Answers from APIs are a good guide to, but not the same as, what someone sees in an assistant's app, where the account, location and history change the answer.
A free one-off report asks only a few questions, so most of its figures are marked insufficient data. That is the honest result of a small sample.
Tone is recorded for each mention as positive, neutral or negative, with its sentence, and shown as counts, never as a score.
2026.10.2 · 23 October 2026
An answer that says it could not find a business, does not know it, or has no information about it no longer counts as a mention of that business. Answers are re-read under this rule; nothing else changes.