Methodology
A visibility score is worthless if it cannot survive scrutiny. This page is the whole method: what we ask, how often, how the score is built, and where its limits are.
The questions
Twelve queries per check, fixed so scores are comparable across businesses: six unbranded discovery questions (“best dentist in Pasadena”), three service-specific questions with buying intent, two direct questions about the business, and one comparison. Phrasing varies deliberately, because assistants answer “best” and “who do you recommend” differently, and that spread is part of the measurement.
The surfaces and runs
Each query runs against ChatGPT, Claude, and Perplexity through their official APIs with live web search enabled, twice per surface, because large language models answer differently run to run. A typical check captures roughly 75 answers, about 25 per assistant. Every capture stores the exact prompt, the full response, cited sources, the model, and a timestamp, and captures are immutable once written.
The score
| Visibility | 45 | Whether and where you appear in answers to unbranded customer questions, weighted by intent. Sole or top recommendation scores full; a passing mention scores low. |
| Site readability | 25 | Deterministic site checks: structured data, AI crawler access in robots.txt, question-shaped citable content, sitemap, content depth. Reproducible by anyone. |
| Accuracy | 15 | What AI gets wrong about you on direct questions: hours, services, location, or confusing you with a similarly named business. |
| Entity footprint | 15 | The off-site signals AI actually cites for local answers: Google Business Profile completeness, review presence on cited platforms, name and address consistency. |
Scoring is arithmetic over the captured evidence. No model chooses the score. The rubric is versioned, and a score always names its version.
Variance, honestly
Visibility is reported as a rate with a 90 percent confidence interval, and the headline score carries that range: “34, range 28 to 40 across 130 captured answers.” The report never makes single-run claims. It does not say “you never appear”; it says “you appeared in 0 of 3 runs of this question on this surface on this date,” with the captures attached.
Limits
Answers are generated through the providers’ official APIs with live web search, which can differ from their consumer apps; the report names the exact surface used. Personalization and device location cannot be fully simulated. AI answers change continuously, so every report is stamped with its measurement date. AI Overviews do not appear for every query, and a run where none appeared is recorded as data, never discarded. Where a dimension was not measured, the report says so instead of estimating.