What Is Temperature in an LLM?

Temperature is the parameter that controls how much randomness a large language model applies when choosing its next token. At each generation step the model holds a probability distribution over possible next tokens; temperature reshapes it. Low values sharpen the distribution toward the most likely token, producing consistent, conservative output. High values flatten it, letting less probable tokens through — more varied, more creative, more erratic. OpenAI's API accepts 0 to 2 (default 1); Anthropic's accepts 0 to 1.

How does temperature create answer variability?

At any temperature above zero, generation is a weighted dice roll per token. One early divergence — "The top options include Notion..." versus "Leading tools include Asana..." — cascades, because every subsequent token conditions on the path taken. This is a primary mechanical source of answer volatility: the same prompt, the same model, and a different brand list on consecutive runs. Even temperature 0 is not perfectly deterministic in production systems, where batching and hardware effects introduce small variations.

Why does temperature matter for AI visibility work?

Two distinct roles:

• As noise to measure through. Consumer engines sample at nonzero temperature, so a brand's presence in answers is a probability, not a constant. Measurement must sample repeatedly — running each tracked prompt 5-10 times and reporting frequency — rather than trusting single shots. Visibility platforms build this repetition in.
• As a control to exploit. When you build measurement pipelines via API, you choose the setting: match consumer-like sampling for the answers you score, and pin temperature to 0 for the judge models doing the scoring, so evaluations are repeatable.

Example

A team probing "best data warehouse for a startup" at temperature 1.0 saw their brand in 6 of 10 runs; a colleague testing once at temperature 0 concluded the brand "always appears." Both were looking at the same model — one measured the distribution, the other measured a single draw. The companion dial, nucleus sampling, and the broader measurement vocabulary are defined in this glossary.