Retention curves are largely determined in the first session. A user who reached something useful on day one may return on day thirty; one who left confused almost never does, regardless of what you ship afterwards.
Separate the two churn problems
Users who never activated and users who activated and later left are different problems with different fixes, and averaging them hides both. The first is an onboarding problem; the second is a value or competition problem. Split the cohorts before deciding what to work on.
Never-activated and stopped-returning are different diseases with the same symptom.
Find the retention-predicting action
In most products there is an early action that strongly correlates with sticking — inviting a colleague, connecting a data source, completing one real task. Identify it from your cohort data, then design the first session to reach exactly that. It is the most leveraged product work available.
Compress time to first value
Every minute between arriving and getting something useful loses people. Look at the first session as a stopwatch problem: what can be removed, deferred, defaulted or pre-filled. Sensible defaults beat configuration screens for almost every user.
Returning must not mean restarting
People get interrupted. If someone leaves halfway through setup and returns tomorrow to an empty state, they will not try a third time. Persist progress and resume where they stopped — a small engineering effort with a disproportionate effect on activation.
The second session is a separate design problem
Teams design the first session and leave the second to chance. What brings someone back — a notification about something that happened, a result that finished, a colleague's activity — deserves the same deliberate design, because retention is a sequence of returns rather than one good start.
Ask the people who left
A short message to users who signed up and never returned produces uncomfortable, specific and immediately actionable answers. Response rates are low and the replies are worth more per response than almost any other research, because these people have no reason to be polite.
