How we measure AI visibility
No black box: this page explains exactly what happens during a scan, how each number in your report is calculated, and what a scan can — and cannot — tell you.
1. We ask real buyer questions
For your industry and location we generate the questions real customers type into AI assistants — like “Who is the best {industry} in {location}?” or “Who should I call for an emergency?”. You can add up to two of your own questions. Each question is asked exactly as a customer would ask it, with no hints about your business.
2. We query multiple AI models via API
Every question is sent to several large language models — the same kind of AI technology behind ChatGPT-style assistants — via API. Answers vary between models and between runs; that's why we ask many questions across several models instead of relying on a single chat. We show every raw answer in the report, unedited.
3. What we measure in the answers
- •Mention rate — the share of answers that mention your business (by name or a known alias).
- •Average position — where you appear in the recommendation list when you are mentioned (#1.0 = always first).
- •Competitor radar — which other businesses the AIs recommend, and how often.
- •Sentiment — how the AIs describe you when you are mentioned (positive / neutral / negative).
- •Cited sources — the websites and directories the AIs reference, i.e. where their knowledge comes from.
4. How the A–F grade works
The grade is a simple function of your mention rate: A = mentioned in 60% or more of answers, B = 30–59%, C = 10–29%, D = 1–9%, F = 0%. No weighting tricks — the raw answers behind the number are always in the report.
5. Limitations — what a scan can't tell you
AI answers change over time, differ slightly by user location and phrasing, and models are updated regularly. One scan is a snapshot, not a guarantee — that's why monitoring plans re-scan weekly and track the trend. We currently query leading large language models; coverage expands as other assistants offer stable APIs. Automated sampling covers direct and search-grounded channels today; consumer surfaces like ChatGPT web are sampled through a manual human-collection channel, and coverage of China's leading assistants is on our roadmap — we add a channel only when it is really implemented.
6. Two answer channels: direct and search-grounded
Each question is sampled on two channels. Direct: the model answers purely from its trained knowledge — like chatting with an assistant without browsing. Search-grounded: we run a real web search for the question and ask the model to answer only from those results, citing them — the cited URLs come from the search results, not from the model's memory. This simulates how search-backed assistants answer, but it is a simulation (retrieval + model synthesis), not a direct measurement of ChatGPT search, Perplexity or Google AI Overviews. Reports label each channel separately.
7. What our data represents — and what it doesn't
We would rather under-claim than oversell. Our numbers are API samples of AI models at a point in time, and that has three honest limits:
- •A mention is evidence that a test produced a mention — it is not proof that every real user sees the same answer. AI answers vary by user, session, location and phrasing.
- •Answers vary run to run even with the identical prompt. A single yes/no is nearly meaningless; that is why we show hit rates across repeated scans (e.g. mentioned in 3 of 5 runs) and why re-running the same prompt set on a schedule is the correct way to measure.
- •Visibility is not traffic, and traffic is not conversions. Being recommended by an AI model is an upstream signal — it tells you where you stand in the answer, not how many customers it sends you.
This is also why every sampled answer in Geovory can be opened as a public evidence page — raw answer, prompt, model and UTC sample time — so you never have to ask anyone to take our numbers on faith.