21 Nov 2025 · Journal
Why push cadence breaks lifecycle models
A notification can open an app. Opening an app can increment a retention numerator. The model then reports that the lifecycle improved. In class we treat that as a contamination event unless the push was itself the product (alerts, banking, transit).
The break is conceptual. A lifecycle window is supposed to describe unaided return. Cadence describes the company’s willingness to poke. Mixing them is how Experiment Design for Retention Levers gets students who want a “win” by Thursday. We mark those experiments as invalid even when the p-value looks tidy.
There is a second break: permission. Users who allow pushes are not a random slice. Their curves were different before you scheduled the 6pm ping. Comparing them to muted users and calling the difference “lift” is a sampling error with a marketing budget.
What we allow: a push that reminds someone of a time-bound action they already chose (a booked class, a parcel). What we reject as a window definition: open-from-push counted as activation. Put open-from-push on the tolerated shelf. Let the core action remain trusted.
If your only remaining lever is cadence, the oral defence should say so. Honesty here is rarer than a clever send-time model. Schema Flowgrid would rather a small unaided cohort than a large poked one that collapses when Apple quietly throttles you next spring.