Designing segments that don't overfit

How segments work

Segments work by creating synthetic personas: individual AI-generated audience members built by sampling from the attributes you have switched on, things like age band, income level, or a particular attitude or trait. Each persona then answers your campaign questions as if it were that kind of person, and the platform aggregates their responses to describe how the whole segment tends to think.

Why more attributes isn't better

The temptation is to switch on every attribute you have available, on the theory that more detail makes a more realistic persona. In practice, the opposite tends to happen. When a persona carries many active attributes at once, its answers start to read as if they are justified by that exact combination of traits rather than reflecting how the broader group actually behaves. You end up with a very specific, very confident-sounding voice that may not generalize to real people who share only some of those traits.

Concretely: a persona built with age band, income level, and three attitude traits switched on at once answers as a very specific kind of person — someone simultaneously older, affluent, brand-loyal, and an early adopter. Real audience members rarely carry every one of those traits together, so the persona's opinions end up describing a corner case rather than the age group or single trait you actually meant to study.

A safer approach

The safer approach is to keep only the attributes that actually matter for the question you are asking, and switch the rest off. If you are testing how an age group responds, turn on just the age band and leave everything else neutral. If you are testing a single yes or no trait, such as whether someone already uses a competitor, isolate that one trait on its own. This keeps the persona close to the dimension you are actually trying to learn about.

It also means a segment's label is only a convenience name for a particular attribute combination, not a fixed archetype — if a suggested segment name doesn't translate well to your market, redefine the underlying attributes rather than treating the label as fixed.

It also helps to increase the number of individuals sampled per segment, particularly whenever more than one attribute stays active. A larger sample smooths out the quirks of any one persona's answer and gives you a more statistically stable picture of how the segment as a whole responds, rather than leaning on a handful of individual voices.

Related reading

Segments are usually worth deciding alongside campaign intent rather than as an afterthought, since intent shapes how the report talks about the data and segments decide whose simulated responses make up that data. And once a wave has run, the same over-fitting discipline applies to reading it: unrelated or duplicate entities can distort a segment's numbers just as easily as an overloaded persona can.