How we measure AI visibility
We ask the four major AI assistants — ChatGPT, Claude, Gemini and Perplexity — roughly 48 real buyer questions about a metro and vertical at temperature 0, record every answer verbatim, and count how often each firm is named. Your score is the number of those questions in which an assistant recommended you.
The assistants we test
- ChatGPTGPT-4o mini
- ClaudeClaude Haiku 4.5
- GeminiGemini 2.5 Flash
- PerplexitySonar
We use each platform’s cost-efficient tier — the same class of model that powers everyday consumer answers — at temperature 0 where supported, to keep runs comparable week to week.
How we design the questions
For each vertical we define 12 buyer intents and phrase each one four ways — navigational (“best accountant Austin”), conversational (“who should I use for taxes in Austin?”), and comparison forms. Different phrasings produce different answers, so we mix them deliberately. That yields ~48 questions per run, anchored to the metro and representative ZIP codes.
How we identify firms
We pull the full list of firms in the metro from public map data, then parse each AI answer with a language model grounded on that list — fuzzy-matching loosely-worded names (“the Smith firm”) back to real businesses. Names we can’t match are flagged for review, never silently dropped.
How scoring works
Your AI Mention Score is how many of the ~48 questions named you, across all four assistants. We also compute share of voice (who gets named, and how often) and trace which sources the answers cite. In Austin, only 11 of 241 accountants were ever named.
Why your score can change between runs
AI assistants give different answers from session to session — that’s normal. A single run is a snapshot; the weekly trend is what matters, and it’s the number we help you move. Because we store every verbatim response behind every claim, each report is fully defensible.