Stability and sampling

AI is non-deterministic, which means if you ask the same engine the same question twice, you will get slightly different answers. One answer is therefore a single sample, not a verdict.

Stability comes from sampling over time: your tracking cadence asks each question repeatedly and accumulates the responses. As a rule of thumb, we treat roughly eight accumulated responses as a stable read of how an engine treats your brand; a daily cadence gets there in about a week, a weekly cadence over several weeks.

Below that, a number is honest but noisy, so we say so on the page. A metric built on fewer than about eight responses carries a small chip showing the count it rests on, like n=5. Read a chipped number directionally, as “we are showing up”, rather than precisely, as “we appear in exactly 47.2% of answers”. Staying quiet would be the dishonest option: with no chip, a figure drawn from two answers would look every bit as settled as one drawn from two hundred.

The count in a chip is that metric's own denominator, not a generic response count, because the metrics do not share one. Position is averaged only over the answers that mention you, so a brand named in 2 of 20 answers has a Position resting on two observations, and its chip reads n=2 even though twenty answers were collected. Each metric's “How it's calculated” table names the denominator its chip counts.

A small sample also changes what your report says, not just how it is labelled. A brand that appears to have moved on a handful of observations is left out of the headline changes rather than printed with a caveat, because a move that small is not yet news.