Define the scope before measuring
A useful baseline specifies the business, competitors, buyer questions, AI systems, markets, languages, and measurement period. Questions can cover discovery, comparisons, capabilities, factual claims, and purchase considerations.
A result belongs to that scope. It is not a census of everything an AI system knows, and it does not describe all users' conversations. The choice of questions and systems affects the findings.
- State which questions and comparison scenarios were tested.
- Identify the systems, markets, languages, and time period covered.
- Keep valid responses separate from failed requests; disclose missing coverage.
- Distinguish factual checks, observed associations, and analyst interpretation.
What the metrics mean
A model can recommend more than one company in an answer; recommendation shares across companies need not total 100%. A cross-model average should disclose whether models are weighted equally and which systems are included.
The website's illustrative dashboard averages five example recommendation shares equally: (32 + 21 + 17 + 43 + 26) ÷ 5 = 27.8%. Its largest example positioning gap is 90 − 32 = 58 percentage points. Neither number is a measured customer result.
| Metric | Meaning | Interpretation |
|---|---|---|
| Visibility / mention rate | Valid tested responses mentioning the company ÷ valid tested responses × 100 | Presence in the defined sample, not all AI conversations |
| Recommendation share | Valid tested responses recommending the company ÷ valid tested responses × 100 | A recommendation is distinct from a passing mention |
| Observed association | How often a specified category or attribute is associated with the company in the assessed sample | Interpret alongside the attribute definition and tested questions |
| Perception gap | Desired positioning compared with the observed association | A difference in percentage points, not proof of a cause |
| Factual accuracy | Observable claims checked against current supporting information | Separate confirmed errors from claims that cannot be verified |
Compare like with like
Changes are easier to interpret when the buyer questions, systems, markets, languages, and measurement approach stay comparable. If scope changes, explain that change rather than treating every difference as improvement or decline.
AI answers vary. System updates, retrieval changes, timing, and the wording of a question can affect a response. A single answer or one before-and-after example is insufficient to establish a stable trend.
The measurement loop is: establish a baseline, diagnose material gaps, improve relevant information, and re-test. Reports should distinguish observed movement from conclusions about why it happened.
Investigate source evidence
Citations can identify pages associated with an answer. Other relevant material can include official service pages, product documentation, third-party directories, customer reviews, and independent mentions.
A response may combine retrieved material with information learned during training, and some systems do not expose citations. Uncited or inaccessible source influences cannot be verified from the answer alone. Potentially relevant sources should not be presented as confirmed causes.
What an assessment cannot guarantee
Clear content and supporting evidence are useful for buyers as well as AI systems. Google's guidance for its generative search features also emphasizes foundational SEO and unique, reliable content; it does not require a special AI schema or llms.txt file.
- That publishing or editing a page will change an AI system's answer.
- A guaranteed ranking, mention, citation, or recommendation.
- Complete access to a model's training data or internal reasoning.
- That one intervention caused an observed change.
- Identical results across systems, markets, or future measurements.
Further reading
These independent references explain Google's search requirements and the limits of special AI-search markup. They are general guidance, not an endorsement of PerceptCore or evidence of customer results.