Journal · 19 January 2026
Reading a retention curve without the theatre
Retention charts are where App Analytics goes to costume parties. A rainbow of weekly cohorts, a logarithmic axis nobody mentioned, and a tooltip that appears only if you hover like a surgeon. Pipeline Nodecore’s house rule is simple: if you cannot narrate the plot in two minutes with the projector off, it is decoration.
Cohort Signal week six is where we ask students to retire at least one pretty line. The line is rarely “wrong.” It is unemployed. It does not answer the question the meeting thought it was asking.
Name the window before you praise the shape
N-day retention, unbounded, and range retention are different sentences. Mixing them in one screenshot is how a product looks stable on Monday and doomed on Thursday. Write the window in the title, not in a footnote. “Returned on day 7” is not “returned at least once by day 7.”
Then name who is in the cohort. Install timestamp alone will include people who never completed a qualifying action. First-session completers versus the rest was the split that changed the Leeds civic debate. The blended curve had been a lullaby.
Decay is not a moral failing
Utilities should decay. Subscription audio should worry about a different shape than a meter app. Resurrection — people who return after a long silence — must not be poured back into the same jug as new-user decay. Boards love resurrection because it looks like forgiveness. It is often a push notification and a sale.
Print the curve. Screens lie about scale, which is why the Signal Room still uses paper. If the y-axis starts at 40 percent to make a dip look like a cliff, that is theatre. Start at zero or say why you will not.
Platforms
If event names differ by operating system, you do not have a retention comparison. You have two stories overlaid. Split first, rename later, as the podcast company did. Conversion did not improve. The argument did, which is the actual product of measurement.