What is temperature in a language model?

When a language model generates text, at each step it's choosing among many possible next words, each with some likelihood. Temperature controls how sharply that choice favors the most likely option. At low temperature, the model almost always picks the highest-probability word, producing repetitive, predictable output. At higher temperature, it more often picks less likely words, producing more varied output, and, along the way, more varied citations.

This matters for AI visibility because temperature is usually set by the platform, not the user, and most consumer AI products run at a temperature well above zero for the sake of natural-sounding answers. That means the same prompt run twice through the same model can surface a different set of sources, not because anything about the web changed, but because the model's word-by-word choices landed differently both times.

This is the mechanical reason a prompt basket needs to be run more than once to mean anything: a single run at nonzero temperature is one sample from many possible answers, not the answer.