GRW-11 · SEC. 10 Marketing & Growth

Find the Retention Cohort That's Actually Bleeding

Feed an agent cohort retention data and get one diagnosed churn point, not generic save-the-user tactics.

FORMAT: workflow
DIFFICULTY: intermediate
TIME: 15 min
TOOLS: universal
MODELS: any
COPIES: 0 so far

When to use this

You have month-over-month or week-over-week cohort retention data and retention feels "bad," but you don't know if it's a slow universal decay (normal) or a specific cohort that broke (fixable). You want a diagnosis before you brainstorm win-back campaigns.

The pattern

Pastes as plain text

Step 1: Ask me to paste my cohort retention table, one row per signup cohort, one column per period since signup:
Cohort | Period 0 | Period 1 | Period 2 | Period 3 | ...
If available, also ask for a timeline of releases, pricing, onboarding, or acquisition changes by cohort date.
Wait for my reply. Don't diagnose until you have it.

Step 2: Diagnose, don't brainstorm:
1. Calculate the period-over-period drop-off rate for every cohort.
2. Compare cohorts against each other, not just against time. Determine: is every cohort decaying at roughly the same shape (systemic, expected), or do recent cohorts drop off faster at the same point in their lifecycle than older cohorts did (something specific broke)?
3. Identify the single period (e.g. "period 0 to period 1") and cohort range where the steepest, most anomalous drop is.
4. State that as the one lever, not a list. If the data shows a systemic decay with no anomaly, say so explicitly instead of forcing a fake lever.
5. Show the numbers that led you there.
6. Propose exactly one experiment specific enough to start today. Tie it to a dated release, pricing, onboarding, or acquisition change only when that timeline was supplied; otherwise label the causal link as a hypothesis.
7. Name the one metric and target that tells us by next review if it worked.

Step 3: Before you deliver the answer, check it. If it hedges or lists multiple levers, redo it against one question: which single period is the best evidence-backed experiment to test next? If there is no anomaly or not enough data, say so instead of forcing a lever.

Real example output

Cohort Jan: P0 100%, P1 62%, P2 51%, P3 47%
Cohort Feb: P0 100%, P1 58%, P2 44%, P3 -
Cohort Mar: P0 100%, P1 41%, P2 -, P3 -

Diagnosis: this is not a systemic decay curve. Jan's P0-to-P1 drop was 38 points. Feb's was 42 points. Mar's is 59 points, nearly 20 points worse than Jan at the exact same lifecycle stage, and the trend is getting worse cohort over cohort, not stable. This points to something specific breaking in the first-period experience starting around early February, not a general "people churn eventually" pattern.

Lever: Period 0-to-1 retention, Feb and Mar cohorts specifically.

Experiment: pull the changelog for the two weeks before the Feb cohort signed up. Check whether the day-3 upgrade prompt that shipped Jan 28 is showing to free users before they've had a first successful session, and if so, gate it behind first-value instead.

Metric to watch: Period 0-to-1 retention for the next full cohort, target back above 55%.

Why it works

Comparing cohorts to each other at the same lifecycle stage, not just charting retention over time, is what separates "this is normal decay" from "something we shipped broke onboarding." Refusing to accept a scattershot list of tactics keeps the output tied to one fixable cause instead of five vague re-engagement ideas.

Related patterns

GRW-09Turn Your Product's Real Activation Steps Into an Onboarding Email SequenceGive it your actual activation funnel steps, get one nudge email per step, not a generic welcome series.GRW-03Find Next Week's Highest-Leverage Growth MoveFeed an agent your weekly metrics and get one ranked growth lever with the reasoning shown, not five vague ideas.MKT-13Voice-of-Customer Synthesis from Public ReviewsTurn a pile of pasted reviews for an adjacent product into ranked complaints and unmet needs.