Which topics this audience is in far LESS than everyone else, and far more.

Inverts the usual question. The rest of the platform builds an audience from topics you name; this starts from an audience you already have and reports what its members are unusually absent from. Absence is often the more reliable signal: given a list of converters, predicting who else will convert is hard, while ruling out who will not is frequently easy.

Read significance_z BEFORE the rows. A sweep compares hundreds or tens of thousands of topics at once, so some differences look convincing by chance alone. significance_z is the |z| a row must clear across this sweep (Bonferroni corrected) and significant_rows counts those that do. Rows are returned raw and never pre-filtered, so a row can appear here and still be noise.

detectable_feature_size is a guide to size, not a cutoff: roughly the reach a topic needs for this audience to show it at half the usual rate. Smaller topics can still produce significant rows when the difference is stronger.

universe_scope names the denominator. It changes with the tiers swept and moves every lift, so results with different scopes are not comparable.