Why recommendation probability needs a range
A point estimate looks precise. An interval tells you how much evidence is underneath it.
AI answer engines vary. Ask the same buyer question more than once and the answer may change because generation is probabilistic.
One run is an anecdote
A single observation cannot tell you how often a brand is likely to be recommended. Outpick repeats a frozen portfolio of buyer questions and keeps the denominator beside every result.
type Measurement = {
recommended: number;
eligible: number;
interval95: [number, number];
};The interval is part of the result
We use a 95% Wilson interval for a binomial proportion. It behaves sensibly at small samples and near the edges.
| Eligible observations | Recommended | Reported result |
|---|---|---|
| 88 | 47 | 43–64% |
| 412 | 224 | 50–59% |
The larger sample supports a narrower range. That narrowing is useful information; hiding it behind a rounded score is not.
What happens when data is missing?
Unavailable is not zero. If an engine cannot be measured, the report says Unavailable and explains why.
Make the claim auditable
Every published range should travel with its dates, prompt portfolio, eligibility rules, method version, and raw evidence.