Temperature is a sampling parameter used when a language model chooses each next token. At low temperature the model picks high-probability tokens more consistently, producing similar output each time. At higher temperature it samples more freely, producing more varied and sometimes more creative output. Temperature and related sampling settings are one reason the same prompt can return different answers, including different brand recommendations, on different runs.
How it works
At each step the model produces a probability for every possible next token. Temperature reshapes that distribution before a token is sampled. A value near zero sharpens it so the top choice almost always wins. A higher value flattens it so lower-ranked tokens get chosen more often. Related settings, such as top-p, cut off unlikely tokens in other ways.
p_i = exp(z_i / T) / Σ_j exp(z_j / T)
// z = model scores (logits), T = temperature
// lower T sharpens, higher T flattensWhy it matters for AI visibility tracking
Consumer AI products do not let users set temperature, and providers do not publish the settings they use. But some randomness in generation, combined with changing search results, personalization and model updates, means a single check of a prompt is not reliable evidence. A brand might be named in one run and absent in the next.
Good prompt tracking accounts for this by running each prompt several times and reporting how often a brand appears, rather than treating one answer as the truth.
Example
Five runs, three answers
Ironvale, a fictional field service platform, asks the same 'best field service software for HVAC companies' prompt five times in one afternoon. Ironvale appears in three of the five answers and in a different position each time. The team stops reporting single screenshots and starts reporting appearance rates across repeated runs.
Common mistakes
Drawing conclusions from one screenshot
Variation between runs is normal. Repeat prompts and look at rates over time.
Using high temperature for factual tasks in your own tools
For extraction, classification or support answers, low temperature usually gives more consistent results.
Questions
Not always. Even at zero, some systems produce small variations because of how computation is batched and parallelized. Vendors note this in their docs.
Not in the consumer apps. It is a setting in the developer APIs.
