Reading a retention curve without lying to yourself

A retention curve is a story with the boring parts removed. The tool draws a line; the line implies a personality. “Our users come back.” “This cohort is healthy.” Both sentences can be false while the line looks athletic. The work is the exclusions you applied, and the ones you did not.

First, name the return event. Opening the app is a weak return if your product is a reader, a bank, or a marketplace. Billing success is a dishonest return if the user never did the thing they pay for. In Retention Modelling Desk we force a choice: the same event you used for activation, or a stricter one. Mixing them mid-year is how teams discover a “miracle” that was a definition change.

Second, split resurrection. Users who left for ten weeks and came back because you sent a twenty-percent code are not the same as users who never left. Some tools fold them into D30 by default. If you present that number beside a retention narrative, you are mixing a campaign result with a product habit. Finance will eventually notice. Product should notice first.

Third, write the falsifier. “If D30 among users who completed onboarding in under four minutes falls while D30 among the slow group rises, we will stop claiming onboarding length is the lever.” Without that sentence, every wiggle becomes a win. With it, you can spend week five of the atelier looking slightly disappointed, which is a professional emotion.

Calendar retention and bounded retention answer different questions. Use both if you must; label them. Unbounded “any activity in the last 30 days” is a popularity contest with a moving window. It has a place in an ops dashboard. It does not belong in a strategy memo unless you explain the window in the same paragraph as the number.