Temperature (AI model)
Temperature is a sampling setting on a language model that controls how much randomness goes into choosing each next token: near zero the model takes the most probable token almost every time, and higher values let less probable tokens through more often. It is set by whoever calls the model, not by the person reading the answer.
What the setting does to the next word
At each step a language model holds a probability for every token it could produce next. Temperature rescales that distribution before one token is drawn from it. A low value sharpens the distribution so the front-runner wins nearly always, which makes output repeatable and conservative. A high value flattens it so alternatives are drawn more often, which reads as variety or, past a point, as drift. Anthropic’s glossary describes the effect in those terms; each provider’s API documents its own range and default.
Why the same question can return different names
A consumer assistant runs at whatever temperature its operator chose, and you cannot read the setting. Ask the same buyer question twice and the list of businesses can differ, in order and in membership, with nothing about the businesses having changed. One answer is therefore a sample, not a fact about your standing. This is why the measurement puts a panel of questions rather than one, keeps every answer, and reads visibility as a share across them, re-measured on the same questions once the fixes are live, and every month with monitoring. Movement in that share is signal; a single run is weather.
A common question
Why can the same buyer question return a different list of names each time?
Because each answer is sampled rather than looked up. A consumer assistant runs at whatever temperature its operator chose, a value nobody outside can read, and that setting governs how often a less probable token gets through. Ask twice and the businesses named can differ in order and in membership with nothing about those businesses having changed. One run is a sample, which is why a whole panel of questions is put, every answer kept, and visibility read as a share across them, measured again after the fixes and every month with monitoring.
Related terms
Sources
Definition of the setting: Anthropic’s glossary of model terms. No figures on this page.