What is a noise band in AI visibility tracking?
In short
A noise band is the range a visibility number would move by chance alone, given how many answers it's based on. Assistants give different answers to the same question, so a rate built from a few dozen answers wobbles week to week. A change inside the band is noise; a change outside it is worth acting on.
Keats, from PressRoom. Updated Read as markdown
Assistants rarely repeat themselves word for word. Put one question to Gemini on Monday and again on Tuesday, and the brands it recommends may differ. That isn't a fault in your tracking; it's how these systems work. It means any number built from answers has a margin of error, and the fewer answers behind it, the wider that margin.
A noise band makes the margin visible. Around each week's number, Keats draws the range it would fall in by chance, based on how many answers produced it. If your share of buying answers goes from 58% to 63% on 40 answers, that's inside the band and tells you nothing. If it goes from 58% to 75% on 300 answers, it's outside the band, and something changed.
How to keep the band narrow:
- Ask each question several times a week on each assistant, rather than once.
- Track enough questions: 20 to 30 per thing you sell.
- Read trends by week, not by day.
The band answers "did it change?". It doesn't answer "did my work change it?". For that you need control questions. And the band is why prompt tracking samples each question rather than checking it once.
Sources
- Keats: eight ChatGPT answers to buyers' questions about AEO tools (CSV), read November 12, 2026