How Seige AI measures AI visibility
AI answers change from one run to the next. A single number with no context is not a measurement. This page says exactly what we ask, how often, how a mention is counted, and how sure we are. Method version: neutral-discovery-v3, detection rules: mention-v1.
1. What we ask
We write neutral buyer questions about your category, for example “What are the best tools for X for a small team?”. The questions never contain your brand name or domain, because naming a brand makes an engine look it up and says nothing about whether it recommends you unprompted. We also run a few brand-aware questions (“What is Y?”). Those are shown separately and are not part of your visibility number.
2. How many times
Each neutral question is asked several times per engine. Engines do not give the same list twice: in our own tests the same prompt gave overlap between 50% and 100% from run to run, and different engines overlapped only 36% to 75%. So we report how many of the answers named you (“4 of 12”), with a 95% range (Wilson interval), not a rank. If a range is wide we say so. A free homepage check is one sample per engine and is labelled that way.
3. Which engines, and whether they search the web
We call each engine through its API. Where the engine supports it we switch live web search on, because that is how the consumer apps usually answer.
| Engine | Mode | Model |
|---|---|---|
| ChatGPT | live web search | gpt-4o-mini |
| Gemini | live web search | gemini-2.5-flash |
| Claude | live web search | claude-haiku-4-5-20251001 |
An API call is not identical to what a person sees in the app. Apps add personalization, location and their own settings. We measure a repeatable approximation, not every consumer answer.
4. How a mention is counted
A person-readable rule, not a model grading itself. We look for your brand name, your domain or your aliases in the answer text. A reply such as “I don’t have any information about example.com” does not count, even though it contains the name. Each counted answer gets one of four states: selected (a top-3 item in a list of three or more, or an explicit recommendation), mentioned, hedged (a maybe or an afterthought), or absent. We keep the exact sentence that was counted and show the full answer, so you can disagree with any call.
5. What we do not claim
- We do not claim a rank. Position inside one answer is shown only when the answer is a numbered list.
- We do not claim that a score predicts traffic or sales. We measure appearing in answers, nothing more.
- We do not compare scans made with different methods. Overlapping ranges are not a difference.
- The detection rules are English-only and heuristic. They can misjudge unusual phrasing, which is why the sentence is stored.
- If an engine is down or times out, that answer is left out. It is never counted as a zero.
6. Changes
When we change prompts, engines or detection rules we change the version above, and we do not compare scores across versions.
Questions about the method: reply to any Seige AI email or write to the address on the pricing page.