Methodology
Last updated 2026-08-31
Every AI-visibility vendor generates its own data — there is no Search Console for AI answers. So the honest question is never "what is my score" but "how was this measured". These are the seven questions we think you should ask any vendor, answered in writing for AnswerOrbit. Where an answer names a screen, you can verify it there yourself.
1. Which collection method do you use, per channel?
Answers come from the chat interfaces themselves, never the bare model API — interface and API answers to the same prompt differ materially, so the interface is what we measure. Collection runs through a vendor that drives the real front ends. The method is stored on every stored answer and shown on every screen's disclosure line; when an engine's pipe changes, its stored method changes with it.
2. Which model or product version is queried?
Whatever the interface serves that day — recorded per answer when the interface reveals it. We do not pin models: your buyers do not either.
3. From which countries and languages?
One market per project today (shown on every screen as "1 market"); more markets arrive with the Scale plan. The country each answer was collected from is part of the run record.
4. How many runs per prompt per period, and how are rates computed?
Five to seven samples per prompt per engine, weekly, taken together. Every rate is mentioned-runs over total runs, and every figure on every screen shows its run count. A single run is never presented as a trend.
5. How is a mention detected — and can I see the raw answers?
Substring match against the brand names you configured, position by first mention; citation means the brand's own domain appears among the answer's sources. Every metric clicks down to the individual stored answers it was computed from — the prompt detail page shows each run's full text. If you cannot audit a number, you should not trust it; that goes for ours too, which is why you can.
6. When your method changes, do you say so?
The method lives on each run, so a change never rewrites history: old runs keep the method they were collected with, and the disclosure line reflects the current one. Movement is labeled "real change" or "within noise" by a two-proportion z-test (α 0.05) — a method change cannot masquerade as a ranking change.
7. Can I export the raw data?
CSV on the tables, a read API over MCP for your assistant, and full data export (plus deletion) in Settings — your data is yours.
One more, unprompted: when a check misses an engine, the screens say exactly that — a missed engine is shown as absent, never filled in. Numbers you can defend start with gaps you can see.