Segment
A saved filter that names a subset of your users, like 'paid accounts on mobile', so any chart can be re-run for just them.
See it
What it is
A segment is usually a rule, not a list. 'Plan equals pro, on mobile, fired project_created at least twice in the last 30 days' is a segment, and membership recomputes every time you look, so people join and fall out on their own. Two flavors: property-based (traits sitting on the user record) and behavior-based (did or did not do event X, N times, within a window). Behavior-based ones almost always tell you more.
Reach for a segment when an average is hiding two different populations. Flat overall retention is usually excellent retention among team accounts sitting on top of miserable retention among solo users, and the fix for each is different. Dropping a segment onto an existing funnel or retention curve is the cheapest real analysis anyone on the team can run without a data request.
Gotcha: 'segment' and 'cohort' get used loosely, and the distinction that actually matters is fixed versus recomputed. An acquisition cohort (everyone who signed up in March) is fixed by a past event and its membership can never change. A behavioral audience ('fired project_created twice in the last 30 days') recomputes, so people join and fall out on their own. Confusingly, most tools file both under the word 'cohort', and a saved segment can also be frozen into a static list for an export or a campaign. So read the definition, not the label: a chart you saved last quarter may quietly be measuring a different set of humans today. Slice far enough and every segment shows a difference, most of them noise, so watch the sample size. One more: lowercase 'segment' and Segment the CDP get confused constantly, so say 'saved filter' or 'audience' when the room stalls.
Ask AI for it
Define 5 candidate segments for my product, each written as an explicit rule I can paste into an analytics tool. For each one give: a short name, the exact conditions (user properties plus behaviors with event names, counts, and time windows), whether membership is fixed or recomputed, and the one question it answers that the overall number hides. Make at least three behavior-based rather than property-based. For size, write the count query for each rule (SQL or the tool's own query syntax) and run it if you have access to the data, reporting the measured membership. Where you cannot query, label the size 'unknown' rather than estimating a share of the user base. Once the counts exist, flag any segment too small to analyze and say what the count threshold should be for the analysis I want to run on it.