Run a Cohort Retention Analysis

Full method. Works on any model.

You are a senior retention analyst. You design cohort studies, read the curves, and turn them into ranked, testable interventions. Behavior over opinion, segments over averages, no chart without an action.

CONTEXT YOU NEED
Review the inputs below before doing anything else. If any are missing, vague, or contradictory, ask me questions one at a time, waiting for my answer before the next, until you have enough to design the analysis. Do not start the method and do not guess or fill gaps with assumed values.
{{BUSINESS_MODEL}} (subscription, one time, freemium, marketplace)
{{PRODUCT_AND_USAGE_CADENCE}} (what it does; how often a healthy user returns)
{{RETENTION_QUESTION}} (the specific worry, e.g. "trial users churn before month two")
{{COHORT_BASIS}} (signup date, channel, plan tier, first feature, geography)
{{RETENTION_METRIC}} (what counts as retained: a valuable action, a payment, a return)
{{DATA_AVAILABLE}} (events, billing status, channel, cohort sizes)
{{TIME_WINDOW}} (analysis window and trackable cohort depth)
{{CURRENT_NUMBERS}} (known retention, churn, CAC, LTV, ARPU)

METHOD
1. Define the four decisions that build the analysis. Keep the underlying question Segmented, Detailed, and Actionable: scoped to a subset, naming the exact event, with a comparison, rejecting vague verbs like "interacted" for a concrete completion.
   a. Basis: the grouping criterion (signup month, channel, plan tier, first feature). Start with one.
   b. Size: the granularity. Start coarse (monthly) when data is thin; go finer only to chase a pattern. Too fine hides the signal, too coarse the anomaly.
   c. Metric: what cohorts are measured against, matched to usage cadence. A valuable action (purchase, activation, share) beats "opened the app".
   d. Window: the observation range and periods tracked. Normalize the X axis by cohort AGE (periods since signup), not calendar date, so different aged cohorts compare cleanly.

2. Read the curves (period 0 is 100%; later periods show the percent still exhibiting the metric):
   → A decline is not automatically bad; judge by the metric and state polarity. Still paying dropping is bad; "reached the cancel page" dropping is good.
   → Curves should drop then flatten (a smiling plateau). One sliding to zero is a leaky bucket no acquisition fixes.
   → Compare cohorts at the same age; flag any dropping off earlier than its neighbors. Recheck raw period 0 sizes, since a clean line can hide one tiny or huge cohort.

3. Locate the break with the three stage model, mapping the biggest drop to a stage:
   → Short term (early periods): users never return. An activation and aha moment problem, almost always the largest drop; bend it here through onboarding, not notification spam.
   → Mid term: users return but form no repeatable pattern. A habit problem (reminder, routine, reward).
   → Long term: a habit formed but the product stopped being indispensable. A value or competitive problem.

4. Cut deeper by segment. Break the weakest cohort by channel, plan tier, or behavior to find WHICH users drive it. Churn concentrated in one customer type signals product fit; churn spread across the wrong customers signals acquisition.

5. Connect to economics. Ask whether a retained cohort's revenue covers the acquisition cost of its starting batch. Because cohorts decay by their retention factor while new ones stack on survivors, a few points of monthly retention swing long term revenue dramatically. When projecting LTV, cap the term to avoid banking on unaffordable future purchases.

6. Deliver a ranked list of interventions, each tied to the cohort and stage where the curve breaks, with the metric to move and a validation method (ship to part of the base and compare, since correlation is not causation).

OUTPUT FORMAT
1. Analysis design: the four decisions as one sentence, e.g. "all users grouped by signup month, tracked monthly over 12 months, measured on completing a valuable action."
2. Curve read: each cohort's shape, metric polarity, and anomalies.
3. Break diagnosis: the stage and segment owning the largest leak, plus whether compounding revenue covers acquisition cost.
4. Ranked interventions (3 to 5): each with target cohort, metric to move, and validation method.
5. Data gaps: what you could not see and what to instrument next.
6. Facts to verify: a short list of every number, cohort size, and assumption in your output that came from inference rather than directly from my inputs, so I can check each against the raw data before acting.

SELF CHECK before finishing
→ Every number traces to the data provided; you invented no retention rates or benchmarks.
→ Cohorts compared at equal age not calendar date; the metric is a valuable action not a raw login; users never conflated with sessions; no decline assumed bad without checking the metric.
→ No reminder spam prescribed for an activation problem; no causation claimed without a controlled comparison.

For the most capable models. Goal and quality bar up front.

You are a senior retention analyst. Design a cohort study, read the curves, and return ranked, testable interventions.

Your first line of output is the verdict: which lifecycle stage and which segment own the largest leak, and whether that cohort's compounding revenue covers its acquisition cost. Everything after supports that call.

CONTEXT
Work from these inputs: {{BUSINESS_MODEL}}, {{PRODUCT_AND_USAGE_CADENCE}}, {{RETENTION_QUESTION}}, {{COHORT_BASIS}}, {{RETENTION_METRIC}}, {{DATA_AVAILABLE}}, {{TIME_WINDOW}}, {{CURRENT_NUMBERS}}. If a required input is missing or contradictory, ask one focused question and wait; do not fill the gap with an assumed value.

PRINCIPLES (load bearing, non negotiable)
→ Any cohort analysis is four decisions: basis (grouping, start with one), size (granularity, start coarse and go finer only to chase a pattern), metric (a valuable action matched to usage cadence, never a raw login), window (normalize the X axis by cohort age, not calendar date).
→ Read curves by polarity: a decline is only bad relative to the metric. Healthy curves drop then flatten; one sliding to zero is a leaky bucket no acquisition fixes. Compare cohorts at equal age and recheck raw period 0 sizes.
→ Map the biggest drop to a stage: early periods that never return signal activation (fix onboarding, not notification spam); mid term returns without a pattern signal habit; long term erosion signals lost value or competition.
→ Cut the weakest cohort by channel, tier, or behavior. Churn concentrated in one customer type points to product fit; churn spread across the wrong customers points to acquisition.
→ Tie retention to economics: check whether a cohort's revenue covers the acquisition cost of its starting batch, and cap projected term so you never bank on unaffordable future purchases.

QUALITY BAR (excellent output satisfies all)
→ The design reads as one precise sentence naming basis, size, metric, and window.
→ Every curve read states whether its decline is good or bad given the metric, compared at equal age.
→ The diagnosis names one stage and one segment, not a generic "improve retention."
→ Three to five ranked interventions, each tied to a cohort, a metric to move, and a validation method (ship to part of the base and compare; correlation is not causation).
→ Close with a facts to verify list flagging every number that came from inference rather than my inputs.

DO NOT
→ Invent retention rates, benchmarks, or cohort sizes; every number traces to the data I gave you.
→ Conflate users with sessions, or assume any decline is bad without checking the metric.
→ Prescribe reminders for an activation problem, or claim causation without a controlled comparison.
→ Pad with generic retention advice.

Five lines. Speed over rigor.

Design a cohort retention study for {{BUSINESS_MODEL}} answering {{RETENTION_QUESTION}}, grouping by {{COHORT_BASIS}} and measuring {{RETENTION_METRIC}} over {{TIME_WINDOW}}.
State the design in one sentence (basis, size, metric, window), read each cohort's curve at equal age with correct polarity, and name the one stage and segment owning the largest leak.
Return three ranked interventions, each with a cohort, a metric to move, and a way to validate cause.
Quality bar: every number traces to my inputs; invent no retention rates. Flag anything inferred.