Repetitions and the probabilistic nature of AI answers

AI answers aren't fixed

Ask an AI model the same question, as the same simulated person, more than once, and you will not always get back the exact same answer. Models work probabilistically: they have a sense of which options are more or less likely, but not a single fixed answer they repeat identically every time. Ask a hundred times and a handful of brands will keep coming up as the leaders, but the exact order between them can shift from one run to the next.

A single response is therefore one draw from that distribution, not the distribution itself. Treating it as the whole picture risks reading noise as signal, especially for anything close to a tie.

Repetitions smooth that out

This is what the repetitions setting is for: asking the same question to the same simulated person more than once and averaging across the results, rather than relying on a single draw. More repetitions produce a more statistically stable picture, particularly for brand-mention counts where you care about the overall pattern rather than any single response. It comes at a real cost, though, since each repetition is a full additional run, so it is worth setting deliberately rather than maxing it out by default. A handful of repetitions is usually enough to smooth out the noise without materially increasing cost.

The one setting that cancels this out

Repetitions only reveal something if the model is actually allowed to vary its answer. If the model's temperature is set fully deterministic, it will tend to return close to the same answer on every repetition, in which case running it multiple times mostly confirms the model is consistent rather than surfacing genuine variability. If you want to see how much an answer actually varies, pair repetitions with a non-zero temperature; if you specifically want one deterministic read, a single repetition is enough and further ones will not tell you much more.

Related reading

This is the same instinct behind not over-loading a segment's persona with attributes: both are about getting a result that reflects the broader pattern rather than leaning on a single, possibly unrepresentative voice.