Reading retention without a vanity headline

A retention percentage without a window is a slogan. “D7 is 23%” tells a board nothing about who was eligible, how identity was joined, or whether iOS backgrounding truncated the session before the event you call “active.” App Analytics teams still lead with that number because it fits in a slide. The curve is the artefact; the headline is a compression that usually lies.

Cohorts are a table, not a vibe

Start with a first-open date you trust. If first-open is assigned at SDK init on one platform and at first login on the other, you do not have a cohort. You have two products glued together with a hopeful join. Write the rule at the top of the memo. Then show D1, D3, D7, D14, D30 as a table for each weekly cohort, not as a single smoothed line that hides a bad week of instrumentation.

Small apps should not average the last ninety days into one curve and call it a trend. A release that broke app_opened for four days will look like a retention crisis. It is a tracking hole. The intensive spends an unglamorous hour comparing event volume against App Store download estimates, not because downloads are truth, but because a silent tracker is a common UK-side failure after a consent prompt change.

Truncation is not churn

Mobile sessions end for reasons that have nothing to do with affection for the product: a phone call, a low-power mode, an OS that freezes the process. If your “active” definition requires a session_end event that iOS often never sends, D1 will look philosophical. Prefer an action the person took — item_saved, play_started — over a session container the SDK invented.

Week three dips deserve a special suspicion. In several catalogues we have marked, the dip lined up with a push permission prompt that was not instrumented, or with a paywall that fired paywall_viewed on Android and nothing on iOS. Before you brief a designer, count the events. The curve is downstream of the catalogue. A pretty retention chart on a broken verb list is interior decoration.

What to write under the chart

Every readout we mark must include three sentences: the identity rule, the window, and one thing the data cannot see. Sampled web traffic, missing background events, users who declined analytics. If you cannot write the third sentence, you are not finished. You are protecting the chart from a question that will arrive in a steering meeting whether you like it or not.

This is the work of the Retention Curves Intensive, which assumes you can already join a user to a first-open date. If you cannot, sit the Event Taxonomy Studio first. A later note on session attribution covers the related mess of deep links and cold starts.