Temperature (LLM)

At each step of inference, the model assigns probabilities to every possible next token. Temperature reshapes that distribution before sampling: near zero, the model almost always picks the most probable token, yielding nearly deterministic output; higher values flatten the distribution, letting less likely tokens through and increasing variety.

Consumer chat products typically run at moderate temperatures for natural-sounding conversation, which means answers, including which brands get named, vary meaningfully between runs of the identical prompt.

For recommendation queries, several brands often sit close together in probability. At non-zero temperature, sampling decides which subset appears in any given answer: you might be named in six runs out of ten. Your true visibility is that frequency, not the binary outcome of one check, the statistical root of answer volatility.

A strong brand association in training data and retrieved sources raises your baseline probability, so you survive sampling more often. Weakly associated brands flicker in and out; strongly associated ones appear consistently at any reasonable temperature.

First, never judge AI visibility from single queries, sampling noise will mislead you in both directions. Second, aim to raise your underlying probability: dominant share of voice in source material makes you the high-probability token sequence that low and high temperatures alike select.

Measurement must match the mechanism: repeated daily sampling of the same prompts, aggregated into mention rates. Geonimo's tracking is built on exactly this principle, surfacing your stable visibility level beneath run-to-run randomness.