Core concepts
n-sampling and distributions
Why every result is a range, how the interval is calculated, and how to choose n.
AI answers are non-deterministic. Ask the same question ten times and you will get ten partly different answer sets. Reporting one of them as the truth means reporting noise that flips next week.
How it works
- We execute the prompt n times against the surface, in independent sessions.
- We parse each response separately and record every brand and citation.
- We aggregate into rates and compute a 95% Wilson score interval per rate.
- You receive the aggregate plus, on request, the raw per-sample answers as evidence.
Choosing n
| n | typical interval width | use |
|---|---|---|
| 1 | ±44 pts | not a measurement — never report this |
| 5 | ±31 pts | exploratory prospect checks only |
| 10 | ±25 pts | the lowest count we will put in a client report |
| 25 | ±17 pts | competitive claims and QBRs |
| 50 | ±12 pts | before/after tests on a retainer |
| 100 | ±9 pts | published research and defended claims |
Intervals are 95% Wilson score half-widths computed at a 30% share, which is a typical citation share. They narrow with the square root of n, so halving the interval costs four times the credits — choose n per prompt rather than sitewide.