21 July 2025 · method
Cohort windows that lie
A window imported from a SaaS template can be perfectly drawn and still describe a week nobody in your product lives.
A cohort is a set of people who began something in the same window: installed in the week of 3 March, started a trial on the same Tuesday, survived the paywall in the same month. The window is the definition. Everything after is a story about time passing for that group.
A segment is a set of people who currently share a property: iOS, paid, “power user,” lives in Manchester. Membership can change tomorrow without anyone installing anything. That is allowed. It is not a cohort. Mixing them is the first lie; using D7 because the template offered it is the second.
Why the mix-up is expensive
Steering meetings use “the Android cohort” when they mean “people whose last device was Android.” Then a person switches phones and appears to churn from a group they never entered by date. Product responds with a feature. The packet never said that.
In Funnel Forensics we force a caption under every chart: entry event, entry window, return definition. If the caption cannot be written, the chart does not go to the steering page. It may still live in the lab.
A practical test
Ask: if this person lost the property tomorrow, would they leave the line? If yes, you have a segment. If they stay because their install week is fixed in history, you have a cohort. Teach the distinction once and several vanity arguments evaporate.
Cohort Craft spends week two on windows and week three on ghosts that fake an entry. See the syllabus and why event names rot.