AI Answer Variability Across Runs (Jaccard)

Paste repeated answers to one question and see how much the set of brands named changes from run to run.

How to use

What it checks and how it works

For each pair of runs the tool takes the set of brands named in each and computes the Jaccard index: the number of brands in both runs divided by the number in either. A score of 1 means the two runs named the same brands; 0 means they share none. The mean, the lowest and the highest pair summarise the whole set, up to 12 runs. The index looks only at which brands appeared, not at their order, wording or sentiment. Two runs that name no brand at all have no defined score and are left out of the mean. A low overlap on a few runs is a reason to run more, not a finding about an assistant. The result describes the pasted runs only.

What counts as a high overlap?

There is no universal threshold. Compare your own questions with one another: a question whose runs keep naming the same brands is stable, and one whose brands change each time is a lottery for anyone hoping to be listed.

Why does the same question give different brands?

AI assistants generate answers with some randomness, and what they retrieve can change between runs. That is why Zaher repeats each question and reports a range rather than a single answer.

How many runs should I paste?

Three is the least that shows a pattern, and five to ten gives a steadier picture. The tool reads up to 12. Use the census sample-size calculator when you need to plan the number of answers.

Sources

Related tools

Privacy

Runs in your browser; nothing is uploaded. The interactive tool needs JavaScript.