trustinsider
English
Try with your brand

AI sampling

How AI answer sampling works

An AI answer is an observation under specific conditions. A useful sample makes those conditions visible and keeps mentions, citations and recommendations separate.

What you get

Define a repeatable observation

Start with a documented question set, market, language and observation period. Record the provider, engine, model and prompt version. Keep repetitions identifiable. If these conditions change, show the change before comparing results.

Conditions accompanying each observation
Question and context
Versioned prompt, language, market and observation objective. Disclose question changes.
Engine and model
Provider, engine and model actually used. An API answer is not equivalent to every user’s experience on the engine’s website.
Time and outcome
Repetition identity, date and complete, incomplete or refused outcome. The sample does not contain only successes.

What you get

Record what was asked and where

The panel design assigns versions to questions and records execution conditions before collection. When a prompt changes, retain its relationship to the previous version and identify unchanged questions: only that subset supports a comparison under compatible conditions. Direct APIs and consumer experiences remain separate methods.

Information context
  1. Question version
  2. Model and mode
  3. Shared questions in comparison

What you get

Exclude incomplete tests from the denominator

Use complete, reviewed answers with confirmed brand attribution. Count duplicates and exclusions separately. A timeout, failed request or unreviewed answer does not demonstrate brand absence. Without eligible answers, indicate insufficient data rather than zero visibility.

Distinct perspectives
  1. Complete, reviewed answers
  2. Errors and refusals kept separate
  3. Empty base: not measurable

What you get

Describe only what was observed

State the number of eligible answers alongside mention, citation and recommendation counts. A cited link is not necessarily an endorsement. Results describe selected questions and observation conditions; they do not measure all AI conversations, audience reached or a universal ranking. A subsequent change alone does not prove that an action caused it.

Information context
  1. Answer actually collected
  2. Events attributed to the brand
  3. Sample limitations

Designing the observation

Questions are part of the measurement.

The panel is a declared set of questions organised around customer choices. Category, comparison, features and support may require different questions. Their selection influences what monitoring can reveal.

Record question origin, language, market, model, method and repetition count. AI-proposed questions and questions observed in customer systems remain distinguishable; neither automatically measures market demand volume. A change creates a new version for review. The calendar retains dates and recurrence, but saving a schedule does not authorise usage or provider calls.

Distinct perspectives
  1. Category questions
  2. Comparison questions
  3. Support questions

Answer provenance

The same question. Different experiences.

A model queried through an API, a web-search-enabled API and a consumer application may produce different answers. Method, tools, personalisation and source availability change what the observation means.

Provenance classification should accompany the data before comparison. TrustInsider’s central assistant and the systems monitored for visibility play different roles: choosing Gemini to assist your team does not limit monitoring to Gemini alone or automatically make other channels available.

Methods to preserve in the sample
MethodPossible observationComparison condition
Model APIAnswer from the queried configuration at collection time.Same panels, model and documented settings; no implicit equivalence with the app.
API with web searchAnswer obtained with the web tool actually enabled.Retain the method, disclosed sources and search conditions.
Consumer interfaceAnswer observed in the interface, if exposed by an authorised provider.State coverage and context; do not generalise to every account.

Outcomes and denominators

Measure the answers. Show the limits.

Mention, recommendation and citation are distinct signals within the same answer. A brand may be named without being recommended; a source may be cited without the answer expressing a preference.

Planned outcomes must remain reconcilable with those collected. Rates are calculated using complete, reviewed answers: a timeout or refusal does not become an absence of the brand. Without eligible answers, the value is not measurable. Providers and models remain in separate groups, avoiding averages of percentages with different bases. Repeated observations illustrate variability within the sample; they do not represent the number of users or conversations in the market.

Versioned panel

Preserve the questions and distinguish every sample change.

Distinct outcomes

Separate complete collections, errors, refusals and outstanding review.

Consistent comparison

Keep model, method, market and period readable alongside the measure.

Related paths

Check your brand’s presence, compare results or choose content improvements: continue with the feature you need.

Apply the method to your brand
How AI answer sampling works · trustinsider