New sign-ups rise while the active customer base barely moves. The team responds with more campaigns, a new onboarding email and a discount. Revenue delays the truth: customers are not reaching repeated value, and every new cohort is being poured into the same broken system.
Behavioural retention practice separates activation, engagement and resurrection; analytical discipline chooses one metric that reflects the business model; and continuous discovery brings customer evidence back to the broken behaviour. Together they tell a founder where to look before trying to “improve churn”, and stop a lagging percentage from becoming a substitute for understanding value.
Diagnose the missing return
Acquisition cannot repair a missing reason to return.
01 · Value
Name the action
Define the behaviour that delivers real value.
02 · Cadence
Set frequency
Match how often the customer naturally needs it.
03 · Input
Find the break
Activation, engagement or resurrection?
04 · Cohort
Watch return
Compare the same customers over enough time.
Define retention around customer value
Logging in is rarely the value. A founder should name the action or outcome that makes the product worth returning to, then choose a frequency that matches the job. Daily retention is nonsense for a quarterly board-reporting product. Monthly revenue can also hide a customer who paid annually and stopped using the product in week two.
- Activation: did the customer reach the first credible experience of value?
- Engagement: did the customer repeat and deepen the valuable behaviour?
- Resurrection: what event or unresolved job can bring a lapsed customer back?
Read cohorts before averages
Group customers by start date, segment and expected use. Compare the share completing the value action at the right interval. A growing top of funnel can make total active users look healthy while each new cohort performs worse. Pair the curve with interviews from customers who activated, stalled and left; behaviour tells you where, while conversation helps explain why.
Aggregate retention can also improve while the underlying customer base becomes less healthy. Strong customers survive and gradually dominate the average; weaker segments disappear from view. Keep the segment and acquisition source attached to the cohort so the team can see whether it is improving the product or merely changing the mix of people who remain.
Let AI connect the behaviour to the evidence
An agent can assemble cohort movement, support interactions and cancellation reasons around one retention hypothesis. It should keep usage facts separate from inferred motivation and avoid a composite “health score” that nobody can explain. The result should be one proposed intervention and the behaviour expected to change.
Teach the agent one customer-value loop
Do not ask the model to “analyse churn”. Give it one customer group, one value-producing behaviour and the cadence at which that behaviour should recur. Its job is to connect the break in behaviour to the closest credible explanation and run one intervention through to a decision.
- Evidence pack: the accepted event definition, source and freshness; cohort start dates; product usage; billing and cancellation facts; support and customer interactions; release history; and explicit gaps where the instrumentation cannot answer.
- Working brief: reproduce the cohort view, identify the largest break in the value loop and separate observed behaviour from inferred motivation. Frame one intervention with a target cohort, expected movement and stop rule.
- Permitted action: prepare interview briefs, customer messages, instrumentation checks and the experiment record. After review, it can coordinate the approved outreach and monitor the named behaviour over the agreed window.
- Review gate: people approve customer contact, product changes and the meaning of the metric. The agent must not repair missing data with estimates, infer emotion from inactivity or silently redefine an event so the result improves.
- Learning record: preserve the metric version, cohort, intervention, observation and continue/change/stop decision. Corrections to event meaning and customer explanations should supersede earlier assumptions rather than being blended into a vague health score.
The “training data” that matters is the reviewed chain from event to explanation to intervention to outcome. That gives the next run a tested model of customer value instead of another generic churn checklist.
A useful next move
Choose one customer segment and one value action. Plot whether each of the last six cohorts repeated it at the natural frequency. Interview three customers immediately before the largest drop. Fix that input before increasing acquisition spend.
Continue with how to turn interest into commercial evidence, or read how to choose the first growth channel.