Methodology
How we measure AI visibility.
We publish our method because the method is the product. Every number in a report should be traceable to a captured response, a date and a platform.
What we measure
Whether an agent or team is named, described accurately and supported by citations when AI systems answer recommendation-oriented questions in a defined market.
Prompt universe construction
We build a fixed set of buyer- and seller-intent questions for the market, segment and price tier being evaluated. The set is locked for the measurement window so results remain comparable across retests.
Model selection
Coverage is determined by relevance to the client's market, technical feasibility and stability of access. We report which systems were tested and when, rather than implying universal coverage.
Testing procedure
Prompts are run repeatedly rather than once. Each response is captured with its text, timestamp, model and any cited sources, because AI outputs vary between runs.
AI Recommendation Share
The percentage of relevant tested recommendation responses in which the brand appears. Responses that do not qualify — refusals, non-answers, off-market answers — are excluded and reported separately.
Citation and source analysis
We record which sources are referenced in recommendation answers for the market and which entities each source is associated with, then compare coverage between the client and competitors.
Entity analysis
We assess how consistently the team, brokerage, markets and specialties are described across the public information environment AI systems can access.
Competitive benchmarking
Competitors are selected from those actually surfaced by the prompt set, not from a list of who the client considers a rival.
Diagnostic framework
The AI Visibility Index is a diagnostic composite used to summarize observable dimensions of visibility. It is not a proprietary ranking factor and does not represent knowledge of model-ranking systems.
Intervention tracking
Every implemented change is documented with a date, owner, rationale and dependency so later measurement can be interpreted against what actually changed.
Retesting
The same locked prompt set is rerun on the same platforms. Differences are reported alongside the confidence we have in attributing them.
Confidence classification
Every statement in a report is classified. We do not report numeric confidence percentages unless a statistical model supports them.
Known limitations
AI outputs are probabilistic and change over time. Personalization, geography and account history can affect answers. Correlation between a signal and an outcome is not causation. We label uncertainty rather than smoothing it over.
Methodology versioning
The methodology is versioned. When it changes, reports state which version produced the numbers so historical comparisons stay honest.
Classification
How every statement is labelled.
Observed fact
Directly captured.
Supported finding
Backed by multiple pieces of evidence.
Working hypothesis
Plausible explanation requiring further testing.
Unknown
Evidence is insufficient.