How motive measures

A set of questions for one brand is its panel. motive proposes it from search demand: what people in your market type into Google about your category, in their language. From those searches it writes questions a buyer would ask an assistant. The questions never name a brand and are not copied from your own pages, so they cannot favour you.

Every question has a type:

• Who to pick: asks the assistant to name providers. Most of a proposed panel is this type, because only these questions can show who gets recommended.
• How to choose: asks how to choose, or compares kinds of options.
• How it works: asks for knowledge. It shows which sites the assistant trusts on your topics.

You review the panel before anything runs. You can edit, remove or add questions, including your own list. A question keeps its history for as long as its wording stays the same.

Each question goes to each assistant word for word, with no instructions added, through the provider's official API with web search on. The model is the one the assistant's free plan uses by default, the largest group of users and the floor of what your buyers see.

Every run asks each question 2 times on each assistant. Runs repeat every week, so each week adds new answers to the history.

Named

The answer text mentions your brand, by its name or one of the names you list for it.

Cited

The answer cites a page on your site as a source. An answer can cite you without naming you, and the reverse.

Searched

The assistant ran a web search before answering. Assistants decide this on their own.

Share of voice

How often the answers name each brand you track, you and your competitors, over the same answers.

Sources

The pages answers cite, each marked as yours, a competitor's, or a third party's. Your pages include the ones you list on other sites, such as a YouTube channel or a marketplace store.

Every rate counts successful answers only. A failed request is never counted as "not named".

Ask an assistant the same question twice and you can get two different answers. A single ask tells you little. motive reports every rate with a 95% interval: the range the true rate is likely in. More answers make it narrower. The same rate of 30%:

• from 10 answers: between 11% and 60%
• from 40 answers: between 18% and 45%
• from 160 answers: between 23% and 38%

Answers to the same question tend to agree with each other, so they carry less information than answers to different questions. The interval accounts for this: it is computed from the number of independent answers the data is worth, not the raw count. This is why asking more questions narrows the interval more than asking the same question again.

When the intervals of two runs overlap, treat the difference as noise. The interval method is the Wilson score interval, adjusted using question-clustered variance estimated from your own answers.

The headline rate weights assistants by your market settings. Countries sharing one market contribute by their country weights. An all-markets figure or a named group weights its markets by your split, or equally when no complete split is set.

Samples are split across countries, not multiplied. Intervals account for repeated questions across assistants and countries. Counts shown are raw successful answers. When no answer (or every answer) names you, motive cannot tell how much answers to one question agree, so the interval assumes they always agree and counts each question once. When part of the weight has no answers, the rate uses the measured part and reports its coverage. Each run keeps the assistant and country weights it started with.

• Every answer in full, with the searches the assistant ran, the pages it read and the pages it cited.
• The complete response from the provider, with a SHA-256 fingerprint, so any number can be traced back to the answers behind it.
• The method version and the exact models of every run. Evidence is never edited after it is recorded.

You can open every answer in the app and download a run as a CSV file.

• We do not read anyone's conversations with an assistant, and we do not say which questions real users ask.
• We do not add instructions to the questions or steer the assistants toward any brand.
• We do not sign in to consumer apps or use real people's accounts.

• The consumer apps add their own hidden instructions, memory and personalization. API answers with web search are close to what users see, not identical.
• Gemini's API has no location setting, so its answers follow the question's language only.
• A brand counts as named when its name or a listed alias appears. A misspelled or unlisted name is missed.
• Google AI Overviews and AI Mode are not observed yet.

Any change to how motive asks, which models it uses, or how it reads answers is a new method version. Every run records its version.

1. Method 5 · 8 November 2026

   Intervals never claim more precision than the weighted answers allow, and brand names in Chinese, Japanese and Korean are matched without word spaces.

2. Method 4 · 8 November 2026

   Headline rates weight each assistant, and each country of a multi-country market, by the market's weights. Intervals use a question-clustered variance that supports those weights.

3. Method 3 · 12 October 2026

   ChatGPT Free and Gemini Free changed their default models; motive follows them.

4. Method 2 · 12 October 2026

   Each assistant uses the model its free plan gets by default. Every provider response is kept in full, with the searches, the pages read and the pages cited.

5. Method 1 · 11 October 2026

   First version: four assistants with web search, sampled questions, named rate with a 95% interval.

Questions about the method: founders@yempik.ai. See also the privacy notice.