How MentionRadar measures AI citations
Every data figure we publish links here. This page explains exactly what the index is, what it samples, and how to read our numbers — including where they are estimates.
MentionRadar runs a large, continuously refreshed set of real buyer questions through the major AI assistants, records which domains each answer cites or mentions, and stores those question–domain links in an inverted index. Every figure we publish is drawn from that index and is labelled with its sample size (N) and time window, or — where we cite an outside study — carries an inline third-party source and date. We never publish a number we did not measure or cannot attribute. Where index coverage for a slice is still thin, we say so and label the figure as illustrative rather than asserting it as a benchmark.
What is the index?
The index is a record of which domains AI assistants cite or mention in answer to a broad, evolving set of buyer questions. A background worker asks the questions, reads each answer, extracts the domains named (both linked citations and plain mentions), and stores the link from question to domain. Querying that record by domain — rather than by question — is what we call reverse AI search.
Which models do we cover?
We report on the assistants we actively scan. Coverage changes over time as models ship and as we expand the index, so each published figure states which models it covers. Where a finding compares models (for example, how often they disagree on sources), the model set is named explicitly alongside the number.
How do we read each figure?
Every data callout on the site follows the same convention so a figure can be quoted in isolation and still be defensible:
| Element | What it means |
|---|---|
| The figure | The measured value (a share, a ratio, a count). |
| The claim | One sentence stating exactly what the figure describes. |
| N (sample size) | How many queries or citations the figure is computed from. Bigger N, more stable. |
| Window | The time range the data was collected over. AI answers shift, so the window matters. |
| Source | “MentionRadar index” for our own data, or a named third party + date for outside studies. |
When is a number an estimate?
Two cases. First, anything attributed to a third party (for example, published correlation figures from other research) is reported as that source states it and labelled an estimate — we have not independently re-derived it. Second, where our own index coverage for a narrow slice (a small category, a single model, a short window) is not yet deep enough to be a reliable benchmark, we label the figure illustrative and say what would make it firm. We would rather show a labelled, honest number than a confident, fabricated one.
What we do not do
- No invented statistics. If we did not measure it and cannot attribute it, we do not publish it.
- No vanity counters. Numbers on the site reflect real measured values or are clearly marked as illustrative.
- No silent revisions. When a recurring figure changes between reports, we frame it as a change, not a correction of the record.
Want to check your own domain?
The same index powers the free Domain Check: enter any domain and read the real queries the models cite it on, with the competitors named alongside. That is the methodology, applied to you.