Skip to content

Monitoring & Methodology

The same prompt can return different answers: how do you make the data reliable?

We repeat the same prompts for 14 or 28 consecutive days and calculate the statistical citation frequency: for example, how many times out of one hundred your hotel appears in the answer.

Generative models have a controlled degree of variability: the same question can produce different answers at different times. That is why we do not measure single answers but observation cycles (waves) on a constant prompt set. The result is a solid probabilistic figure, such as "cited in 87% of cases", which lets you tell a real improvement from a temporary fluctuation.

Related questions