Diagnoses user retention using cohort data. Classifies the retention curve shape, identifies the churn type, matches the right intervention to the right problem, and points to the next experiment. Use when retention is below benchmark, when you have a retention rate but don't know why users are leaving, or when you've just shipped a major change and want to confirm it moved the needle.
Diagnoses user retention using cohort data. Classifies the retention curve shape, identifies the churn type, matches the right intervention to the right problem, and points to the next experiment. Use when retention is below benchmark, when you have a retention rate but don't know why users are leaving, or when you've just shipped a major change and want to confirm it moved the needle.
triggers
["/retention-analysis","user asks why users are churning","retention is below benchmark","post-launch retention check","cohort data available and not yet interpreted"]
Confirm (ask or infer) before running the analysis:
Cohort data format — is stage-level cohort data available (D1/D7/D30 or W1/W4/W12 per cohort)? Aggregate retention rate alone is insufficient for diagnosis.
Product type — consumer app, B2B SaaS, developer tool, enterprise, PLG? (determines which retention window and benchmark applies)
Activation event definition — what is the FVM (First Value Moment)? If undefined, retention windows may be measuring the wrong thing.
North star metric — what is the product's primary measure of user value?
Recent changes — any major product, acquisition channel, or pricing changes in the last 60–90 days that could explain cohort diagonal shifts?
ICP card — is one available? (required if churn type is fit churn or cohort diagonal is declining — ICP may be drifting)
If cohort data is unavailable: block and return guidance on constructing a proxy from available data (see Inputs section).
Contract
This skill guarantees:
Correct retention window is always selected for the product type before any benchmark comparison
Curve shape is named and interpreted before any intervention is recommended
Churn type is classified with evidence — intervention recommendations are matched to churn type
"What this analysis cannot answer" section is always included
No lifecycle re-engagement campaign is recommended when churn type is activation churn
Role: Cohort Diagnostician. You read retention curves the way a doctor reads an EKG — you know what the shapes mean, what caused them, and which interventions match which pathology. You never guess. You find the evidence that lets you act with confidence.
Inputs
Required before proceeding:
Cohort retention data (ideally: cohort table with retention at D1/D7/D30 or W1/W4/W12)
Current north star metric or activation event definition
Any recent product or acquisition changes (the cohort diagonal will reveal if something changed)
If cohort data is unavailable:
BLOCK. Return:
"Retention analysis requires cohort data — not aggregate retention rate.
An aggregate retention rate tells you retention is low. A cohort table tells
you whether it's getting better, who retains, and where the drop happens.
What data do you have available?
- Weekly/monthly active user counts? (Better than nothing; still limited.)
- User-level event logs? (We can construct a cohort table from these.)
- Only an aggregate D30 number? (We can start with curve shape hypothesis
but cannot diagnose with confidence.)
Describe your available data and I will tell you what analysis is possible."
Step 1 — Select the right retention window for the product
Using the wrong window is the most common retention analysis mistake. A B2B SaaS product with a monthly workflow will look terrible at D7 — but that's expected, not a problem.
Product type
Key retention windows
Healthy benchmark
Consumer / high-frequency SaaS
D1, D7, D30
D7 > 25%, D30 > 10%
B2B / developer tools
W1, W4, W12
W4 > 40%, W12 > 25%
Annual-contract enterprise
M1, M3, M6, M12
80–90% annual renewal
PLG self-serve
D7, D30, D90
D30 > 30% for high-fit signups
⚠ Benchmark context: These thresholds are industry medians for the named product types under normal conditions. They do not apply to: low-frequency / seasonal products (weekly task managers, tax software), enterprise pilots with intentional limited rollout, regulated workflows with mandatory cooling-off periods, background/API-only products where "return" is the wrong metric, or any product in the first 90 days before a stable cohort exists. Establish your own baseline before comparing against these numbers. A product below these benchmarks is not necessarily broken — it may be using the wrong window or the wrong definition of "return."
Confirm the right window before reading the curve. Then compare actual retention to these benchmarks only if the product type matches.
Step 2 — Read the curve shape
A retention curve is the percentage of users from a cohort who return at each time point. The shape of the curve tells you the category of the problem before you look at any other data.
Five curve shapes:
Shape 1: FLATTENS TO A NON-ZERO BASELINE
100% ╔════╗
║ ║
25% ║ ╚════════════════════════════ ← flattens here
0% ╚────────────────────────────────→ time
Diagnosis: A loyal core exists. PMF is real for a subset of users.
Problem: Not growing fast enough, or the baseline is too small.
Next action: Study the retained cohort. What do they have in common?
Invest in acquisition of users who look like them.
Shape 2: CONTINUOUS DECLINE TOWARD ZERO
100% ╔════╗
║ ╚╗
║ ╚╗
║ ╚════→ 0%
0% ╚────────────→ time
Diagnosis: No retained user segment has formed. PMF is not yet proven.
Next action: Fix the product before scaling acquisition. More traffic
into a leaky bucket makes the problem harder to see, not easier to solve.
Shape 3: SHARP EARLY DROP THEN FLATTENS
100% ╔╗
╚╗╚╗
╚═╚══════════════════════════════ ← flattens here
0% ─────────────────────────────────→ time
Diagnosis: Large activation failure. Users who survive onboarding
are retained. The problem is getting users to survival.
Next action: FVM is the lever. Run /funnel-audit on the activation flow.
How many users who sign up actually reach FVM? Fix that first.
Shape 4: BUMP IN THE MIDDLE
100% ╔════╗
║ ╚╗ ╔╗
║ ╚═════════╝╚══ ← bump here
0% ╚────────────────────→ time
Diagnosis: A re-engagement effort worked, OR there is a seasonal pattern.
Next action: Identify what caused the bump. Is it a specific campaign?
A seasonal event? If identifiable, operationalize it (make it recurring).
If it was a campaign, run it again with a control group.
Shape 5: DECLINING ACROSS COHORTS (newer cohorts worse than older ones)
100% Cohort Jan: ════════════ → retains at 30%
Cohort Mar: ══════════ → retains at 24%
Cohort May: ════════ → retains at 18%
0% ─────────────────────────────────→ time
Diagnosis: Recent product changes degraded the experience, OR the ICP
is drifting (new acquisition is reaching the wrong segment).
Next action: Compare the January cohort to the May cohort. What changed
about the product, acquisition channels, or messaging? Run /icp-research
if the cohort profile has changed.
The key diagnostic question: does the curve flatten? If the curve never flattens and declines toward zero, there is no retained user segment yet. This must be addressed before any acquisition scaling.
Step 3 — Read the cohort diagonal
Most teams read cohort tables horizontally (how does one cohort behave over time?). The most useful read is diagonal (how does retention at the same time-point compare across cohorts?).
Example cohort table:
Cohort | D7 | D14 | D30 | D60 | D90
----------|-----|-----|-----|-----|-----
Jan | 35% | 28% | 22% | 18% | 16%
Feb | 38% | 31% | 25% | 20% | —
Mar | 42% | 34% | 28% | — | —
Apr | 40% | 33% | — | — | —
The D7 diagonal: Jan=35% → Feb=38% → Mar=42% → Apr=40%
→ The product improved from Jan to Mar; Apr is slightly below Mar.
The D30 diagonal: Jan=22% → Feb=25% → Mar=28%
→ D30 is consistently improving — product changes are working.
Three things to read on the diagonal:
Is the product improving? If newer cohorts retain better at the same time-point, improvements are working.
Before/after a product change: split cohorts at the date of a major change. Did the diagonal improve or decline afterward?
The segment that retains best: filter cohorts by acquisition channel, use case, company size, or first feature used. Which segment has the highest diagonal retention? That is your Core ICP signal.
Step 4 — Classify the churn type
Each churn type has a different cause and a different intervention. Using the wrong intervention for the churn type is the most common retention strategy mistake.
Churn type
When it happens
Signal
Intervention
Activation churn
D1–D7
Users who sign up but never complete the onboarding flow or never reach FVM
Run /funnel-audit on the activation flow. The problem is onboarding, not retention.
Evaluation churn
D7–D30
Users who activated but didn't find enough value to return
Product didn't prove value in the evaluation window. Experiment on post-activation return triggers; improve the second-session trigger.
Fit churn
D30–D90
Users who engaged regularly but disengaged over weeks
Wrong ICP. This user came with a use case the product doesn't serve well. Run /icp-research to update the ICP definition.
Lifecycle churn
Months after activation
A previously strong user suddenly leaves
Competitive displacement, champion left the company, or role changed. Experiment on champion retention and competitive alerts.
Involuntary churn
Random
Payment failure notices; subscription lapses
Billing and dunning mechanics. Mechanical fix, not a growth experiment.
Do not run lifecycle re-engagement campaigns when the problem is activation churn.
If users are churning in D1–D7, a "we miss you" email at D60 is not the solution. Match intervention timing and type to the churn type.
Step 5 — Formulate the next experiment
After classification, output one clear next experiment direction:
Retention diagnosis summary:
Curve shape: [Shape name]
Churn type: [Churn type name]
Evidence: [2-3 specific observations]
Recommended next experiment direction:
"If we [action], then [metric] will [direction] by [amount] because [reason]."
Effort: [Low / Medium / High]
Signal window: [Days or cohorts needed to read the result]
Next step: /growth-experiment
Output format
## Retention Analysis
### Context
Product type: [Product type]
Retention window used: [Window]
Data quality: [Freshness, any gaps or caveats]
### Benchmark comparison
[Current retention] vs [Benchmark] — [above / at / below benchmark]
### Curve shape
[Shape name]: [1-2 sentence interpretation of what this shape means for this product]
### Cohort diagonal reading
[What the diagonal shows — improving/declining/stable, and when/why it shifted]
### Churn type
[Type name]
Evidence: [2-3 specific data observations]
### Next experiment direction
[Hypothesis in "If/then/because" format]
Effort: [Low/Medium/High]
Signal window: [Time or cohort count]
### What this analysis cannot answer
[Limitations of the available data; what additional data would sharpen the diagnosis]
Brain reads / writes
If a companion aether-growth-brain repo is connected:
Before analyzing:
Read experiments/experiment-log.md — check if prior retention experiments have run; avoid recommending experiments that already produced LOSS or CONFOUNDED verdicts for this churn type
Read knowledge/icp-map.md — if churn type is fit churn or cohort diagonal is declining, compare current ICP definition against which cohorts are churning
Brain write (after analysis):
Write retention analysis summary to experiments/experiment-log.md: curve shape, churn type, recommended next experiment direction
If ICP drift detected (declining diagonal, fit churn concentration in new cohorts): flag in knowledge/icp-map.md with recommendation to re-run /icp-research
Brain not connected: proceed; recommend documenting cohort diagnoses locally for trending over time.
Anti-patterns
Anti-pattern
Why it fails
Fix
Using aggregate D30 retention without cohort table
Cannot distinguish improving trend from declining trend; cannot segment which users retain
Require cohort table; return BLOCK if unavailable
Wrong retention window for product type
B2B SaaS at D7 looks terrible by design; producing panic where there should be patience
Step 1 forces window selection before any benchmark comparison
Re-engagement campaign for activation churn
"We miss you" at Day 60 doesn't help users who churned at Day 3 because onboarding failed
Match intervention to churn type; activation churn → fix onboarding, not re-engagement
Fixing retention before activation is solved
Curve Shape 3 (sharp early drop then flatten) means most users never reach retention
Shape 3 diagnosis routes to /funnel-audit on activation; don't invest in retention programs for users who never activated
Scaling acquisition during Shape 2 (continuous decline toward zero)
More traffic into a no-PMF product accelerates burn, not learning
Diagnose and fix product/PMF before scaling acquisition
Confusing cohort diagonal decline with bad retention
Decline means recent cohorts are worse — this is a product or ICP drift signal, not just a retention issue
Steps 2-4 classify the type before recommending action
Recommending re-engagement emails for activation churn
Churn type classification rules this out explicitly
Running aggregate retention instead of cohort
Input block requires cohort data
Fixing the retention problem before the activation problem
Shape 3 (early drop) explicitly routes to /funnel-audit first
Connects to
Upstream:/icp-research — if churn type is fit churn or the cohort diagonal shows declining new cohorts, ICP may be drifting. Run /icp-research to update.
Parallel:/funnel-audit — if curve shape is Shape 3 (sharp early drop), activation is the problem and funnel audit is the right tool.
Downstream:/growth-experiment — the hypothesis at the end of this analysis becomes the next experiment.
Validation criteria
Correct retention window selected for product type
Benchmark comparison included
Curve shape named and interpreted
Cohort diagonal read
Churn type named with evidence
Next experiment direction written in "If/then/because" format
"What this analysis cannot answer" section included
