Reading a retention curve without lying
A retention curve is a story about time. Most of the lying is done to time. We shorten the window so the product has not had a chance to be abandoned. We smooth a weekly dip until it looks like weather. We start the clock at first open instead of first value, so the early drop is someone else’s problem.
None of this requires malice. Product people are rewarded for hope, and charts are polite. The Instance Kernelhub prejudice is that a curve should be a little uglier than the deck that contains it. If your tool offers “best fit,” decline it for anything that will be shown to people who allocate money.
Truncation is the most common kindness. A seven-day curve for a weekly habit will look like a cliff; a ninety-day curve for a tax app will look like disinterest until January. Match the window to the job. Then publish the window in the same sentence as the number. “Sixty-two percent retained” is not a fact until you have said from when, until when, and of whom.
Smoothing is the second kindness. A moving average can hide a release that broke login for a day. If you must smooth, keep the raw series in the appendix and glance at it before you speak. In Retention Architecture we ask each student to present one chart they no longer trust. The exercise is not humiliation. It is practice in withdrawing a claim.
There is also the lie of the missing cohort. If new users from a paid campaign are mixed with people who arrived last year, the curve becomes a cocktail. Split them. If the paid cohort looks worse, that is information, not a reason to reunite the lines before the meeting.
Read slowly. Name the window. Leave the wiggle in. The product will survive an honest dip; it may not survive a year of decorative ones.