How we measure

Each week we ask every engine each of your buyer questions several times, 3 to 5 depending on your plan (see Pricing); the free check asks once. One ask is one sample. Results are grouped in rolling 4-week windows, in UTC, and the latest run day is also shown on its own.

Each answer shows how we collected it, wherever we recorded that:

• Captured from the app: the answer the engine's consumer app gave when we asked, not signed in as you.
• Via API: the answer from the engine's developer API, which can use different settings or search than the app.
• Search results page: the AI answer shown on Google's results page (AI Overviews, AI Mode).

Answers from the API and the app can differ, and even an app capture can differ from what you'd see, because answers change from one ask to the next and can depend on location and account. So we don't treat the routes as interchangeable: when we change how an engine is reached, we raise that engine's methodology version, and charts and comparisons don't cross that change (see below).

How often answered samples name your brand, by name, another name you gave us, or your domain. A question that already names a brand (like “alternatives to Orbit”) isn't counted for that brand, because the answer would almost always name it. Matching is by text, so a lookalike product name can count; we haven't yet measured how often that happens on real answers.

For one question on one engine, the range is a 95% Wilson interval. Across questions, answers to the same question aren't independent, so the range comes from resampling whole questions (a cluster bootstrap), and it always contains the pooled Wilson interval. With 5 or more questions each question counts equally, so the rate is the average of your questions' rates and can differ from the raw share of samples. With fewer than 5 questions we use the pooled share and its Wilson interval. These ranges describe your tracked questions, not every question a buyer might ask.

Of the links an engine cites (each page counted once per answer), the share that point to your own site. Questions that name you are included here. Citations from one answer arrive together, so the range comes from resampling whole questions, and it always contains the 95% Wilson range at that share over the answers behind it: answers, not citations, carry the uncertainty, so it is never shown as more certain than they allow. We show citation share only when at least 5 questions have answers citing any source. The same method gives each engine's mix of source types.

When an engine shows no AI answer (for example, no AI Overview appears), or a sample fails, it's left out of the rate and shown separately. It is never counted as a miss. We show a rate only from 6 samples; below that you'll see “Not enough data”.

A change is called a real change only when the 95% interval for the difference (Newcombe's method, on sample counts) excludes zero and the change is at least 5 percentage points. Otherwise it's within noise, which doesn't mean nothing happened, only that the data can't tell. With fewer than 20 answered samples in either window, we say there isn't enough data. A real change says the rate moved, not what moved it.

The examples below use that plain test, which treats samples as independent. Fix results and source shifts use the stricter tests described next.

We compare the change on the fix's questions with the change on your other tracked questions on the same engines, from days 14 to 27 after you mark it done against the 4 weeks before (minus the answers that picked the fix). It's a real change only when that difference's 95% range excludes zero, it's at least 5 points, and the fix's own questions moved the same way. The range is widened when your questions disagree with each other more than the answer counts suggest, so a change that hits a whole run at once, such as an engine update, mostly cancels out.

“Not enough data” also covers fewer than 2 runs in either period, no engine with at least 5 other questions to compare against (engines with fewer are left out), or a change in how answers were collected. A fix is measured once, and the result doesn't change afterwards.

When a fix's questions would have too few answers after it to give a result, the first weekly run after you mark it done decides, from the 4 weeks before only, to ask those questions a few extra times in the weekly runs of days 14 to 27, within your plan's limit for extra answers. They're used only for that fix's result, which then shows a larger n than the dashboard for the same questions, and an extra that gets no answer is never replaced. Extras make a result possible more often. They don't push it either way: for a fix on one question, most results will still be within noise.

A site is listed when the change in the share of an engine's answers citing it, on the questions answered throughout, has a 99% range that excludes zero and is at least 7 points, and the same comparison a week earlier went the same way by at least 7 points. We use 99% because we check many sites each week, so a 95% test would list some by chance.

If we change how an engine is reached or how answers are read, that engine's methodology version goes up. Each window uses only each engine's newest version. Charts break the line where any engine's version changes, and we don't compare values across that break.