Retention Analysis Skill

Diagnose why users leave, identify what keeps them, and recommend specific, testable interventions — not vague "improve onboarding" suggestions.

Retention Fundamentals

The retention curve has two components:

1. Steepness of initial drop (D1–D7) — onboarding problem
2. Long-term floor level — product-market fit indicator

A product with PMF has a retention curve that flattens. If it trends to zero, you have a PMF problem, not an onboarding problem. Name this distinction explicitly.

---

Retention Metrics Definitions

---

Retention Investigation Framework

Step 1: Segment the problem

Don't analyse "retention" — analyse retention for specific cohorts:

• New vs returning users
• Paid vs free
• Acquisition channel (organic vs paid vs referral)
• Onboarding path completed vs not
• Feature usage (power users vs lurkers)

Step 2: Find the inflection points

Where does the drop happen? D1? D7? Month 3?

• D1 drop → First session experience
• D7 drop → Habit loop not formed
• D30 drop → Value not delivered at depth
• Month 3+ drop → Boredom, competition, or lifecycle event

Step 3: Identify the "aha moment" correlation

Which early behaviour predicts long-term retention?

• Run correlation: users who did [X] in first 7 days vs 30-day retention
• Common patterns: connected an integration, invited a teammate, completed a core action N times

Step 4: Qualify the churn

Interview churned users — never skip this. Survey data alone is insufficient.

• "What was the trigger that led you to cancel/stop?"
• "What were you trying to accomplish that you couldn't?"
• "What would need to change for you to come back?"

---

Output Format

Retention Analysis — [Product/Segment] — [Date]

Question: [Specific retention question being answered] Period Analysed: [Date range] Segment: [Which users]

---

Current Retention Snapshot:

Retention Curve Shape: [Flattening / Still declining / Trending to zero] PMF Signal: [Strong / Weak / Absent — based on curve shape]

---

Root Cause Hypotheses:

"Aha Moment" Correlation: Users who [specific action] in first [N] days retain at [X%] vs [Y%] for those who don't.

---

Recommended Interventions:

Monitoring Plan:

• Metric to track: [X]
• Review cadence: [Weekly / Monthly]
• Alert threshold: [If X drops below Y, investigate immediately]

---

Required Inputs

Ask the user for these if not provided:

• Product and business model (SaaS / consumer app / marketplace / other)
• Current retention metrics (D1, D7, D30 if available)
• Segment to analyse (all users / paid / free / a specific cohort)
• Key question to answer (why is retention dropping? what drives retention?)
• Available data (analytics events, churn surveys, interview notes)

Deeper Materials

This skill ships with support files — use them when they are available:

• references/curve-reading.md — Reading Retention Curves Without Fooling Yourself. Apply it while producing the output; it carries the calibration and judgment calls the method summary above compresses.
• templates/retention-readout.md — a fill-in version of the deliverable with the quality gates inline. Offer it when the user wants to work the document themselves rather than have it generated.

Scoring Rubric (0–40)

Score any output of this skill before handing it over; 32+ is ship-quality.

Quality Checks

• Retention curve shape is diagnosed (flattening vs trending to zero = PMF vs onboarding)
• Cohorts are segmented before analysis (not all users lumped together)
• "Aha moment" correlation is identified or flagged as unknown
• Interventions are specific (not "improve onboarding")
• Churned user interviews are recommended (not just data analysis)
• Monitoring plan includes an alert threshold

Anti-Patterns

• Do not recommend "improve onboarding" without specifying what specific step to change and why
• Do not analyse retention without segmenting by cohort — aggregate retention curves hide cohort-specific patterns
• Do not treat DAU/MAU below 5% as a retention problem — at that level, it is a product-market fit problem
• Do not skip qualitative research — churned user interviews reveal reasons that quantitative data cannot
• Do not set a monitoring alert without specifying the threshold that triggers it

Guidelines

• Never recommend "improve onboarding" without specifying what to change and why
• Benchmark against industry — consumer apps, SaaS, and marketplaces have very different retention norms
• If DAU/MAU is below 5%, that's a PMF conversation, not a retention tactics conversation
• Always recommend talking to churned users — no amount of data replaces understanding the reason

---

Supporting file: references/curve-reading.md

Reading Retention Curves Without Fooling Yourself

A retention curve is the most information-dense chart in product analytics — and the easiest to misread. These are the reading rules.

The three curve shapes and what they mean

1. Flattening (the good one) — drops, then stabilises at a plateau. The plateau height IS your product-market fit for that segment; the flatten point is when habit formed. Strategy: raise the plateau (activation work) and move the flatten point earlier.
2. Declining forever — never flattens; every cohort eventually leaves. No amount of acquisition fixes this; the product isn't retaining anyone long-term. Strategy: stop pouring, find who (if anyone) does flatten, rebuild around them.
3. Smile — decline then recovery: resurrection is real (seasonal jobs, billing-cycle usage). Verify it's genuine reactivation, not a tracking artifact (auth changes and bot traffic both draw smiles).

The mistakes that generate wrong roadmaps

• Averaging across cohorts — a blended curve mixes your 2023 users with last week's; improvement in new cohorts is invisible. ALWAYS plot cohorts separately; the question is whether newer curves sit higher.
• Wrong activity definition — "opened the app" retention flatters; define retention on the value action (created/report-viewed/message-sent). If the curve changes shape when you switch definitions, the vanity definition was hiding the truth.
• Wrong period granularity — a monthly-use product (expense tools) measured weekly shows fake churn. Match the period to the natural job frequency; if you don't know it, that's the first finding.
• Survivorship in feature analysis — "users of feature X retain 2× better" is selection, not causation, until you compare similar users who did/didn't adopt (or run the experiment).

Benchmarks worth holding

The number to obsess over is the plateau, by segment. A product where enterprise flattens at 85% and self-serve at 15% isn't "average 40% retained" — it's two businesses, one working.

The one-sentence output

Every retention analysis should end in a sentence of this form: "Cohorts flatten at [x]% around week [n]; the biggest lever is [moving the plateau/flatten point] for [segment], via [the behaviour that separates flatteners from leavers]."

---

Supporting file: references/worked-example.md

Worked Example — Retention Analysis Skill

A gold-standard run of this skill: a "retention is fine-ish" blended curve pulled apart into one healthy segment and one segment with no product-market fit at all.

The brief

Repwell — a fictional consumer strength-training app ($9.99/mo premium, freemium base) — has a board question after D30 retention drifted down two quarters in a row: "Is this an onboarding problem or a product problem?" New signups run ~40k/month, roughly 55% organic and 45% paid social (broad "fitness" lookalike audiences). PM Sofia Reyes has event data for the Feb–Apr 2026 signup cohorts (n = 118,400: 65,100 organic, 53,300 paid social), onboarding goal-selection answers, and no churned-user interviews yet. Growth's OKR is 50k signups/month by Q4.

The output

Retention Analysis — Repwell New Users (Feb–Apr 2026 Cohorts) — 8 June 2026

Question: Is Repwell's declining D30 retention an onboarding problem or a product-market fit problem? Period Analysed: Signups 1 Feb – 30 Apr 2026 (n = 118,400), activity observed through 7 June Segment: All new users, split organic (55%) vs paid social (45%)

---

Current Retention Snapshot:

Retention Curve Shape: The blended curve appears to flatten near D45 at ~6% — but this is an artefact of averaging. Segmented: organic flattens between D30 and D60 at 10.5–11%; paid social is still declining at D60 (0.9%) and is trending to zero.

PMF Signal: Strong for the organic segment (curve flattens; DAU/MAU 19%). Absent for the paid-social segment (curve trends to zero; DAU/MAU 4% — below the 5% line where this stops being a retention-tactics conversation and becomes a product-market fit conversation for that acquisition source).

Segment detail (blended figures above reconcile: e.g. D30 = 0.55 × 11% + 0.45 × 2.1% ≈ 7.0%):

---

Root Cause Hypotheses:

"Aha Moment" Correlation: Users who log 3 workouts in their first 7 days (12% of signups) retain at 31% D30 vs 3.7% for those who don't — an 8.4× gap (weighted: 0.12 × 31% + 0.88 × 3.7% ≈ 7.0%, matching the blended figure). The gap is upstream, not in-product: 18.1% of organic signups hit the threshold vs 4.6% of paid — mismatched intent never starts workout 2, no matter what the UI does.

---

Recommended Interventions:

Honest tension to surface for the board: cutting bad paid traffic will lift the retention percentage without retaining a single additional human — and Growth's 50k-signups OKR currently depends on that channel. The right frame is cost per D30-retained user by channel, not signups or blended retention.

Qualitative gap: no churned-user interviews exist yet. Before intervention 1 is locked, run 10 interviews with churned paid-social users and 5 with churned organic users — the goal-selection data says who mismatches, only interviews say why they thought Repwell was for them.

Monitoring Plan:

• Metric to track: weekly-cohort D7 retention by acquisition channel; training-plan creation rate in first session
• Review cadence: Weekly
• Alert threshold: paid-social cohort D7 < 7% or organic cohort D7 < 20% → investigate immediately; plan-creation rate target ≥ 35% post-experiment

---

Why it's shaped this way

• The curve is segmented before it is diagnosed — the anti-patterns forbid analysing retention without cohort splits because "aggregate retention curves hide cohort-specific patterns"; here the blended 7% D30 literally averaged a healthy 11% and a doomed 2.1% into a misleading "fine-ish."
• Flattening vs trending-to-zero is named explicitly, per segment — the fundamentals section demands the PMF-vs-onboarding distinction be made explicitly; the answer to the board's question is "both, in different segments," which only the shape diagnosis can show.
• Paid social's 4% DAU/MAU is escalated to a PMF conversation, not a tactics list — the skill draws a hard line at 5% DAU/MAU, so the output refuses to propose onboarding tweaks for a segment that was never going to retain.
• The aha-moment correlation is quantified on both sides of the split (31% vs 3.7%) and reconciled back to the blended number — the output format requires "users who [X] retain at [X%] vs [Y%]," and the reconciliation makes the arithmetic auditable rather than decorative.
• No intervention says "improve onboarding" — each names the specific step (plan selection at end of first workout), the target drop point, and the basis for the lift estimate, per the guideline banning vague onboarding advice and the quality check requiring specificity.
• Churned-user interviews are recommended before committing spend, not after — Step 4 of the investigation framework says "never skip this"; the goal-selection data is treated as who-evidence, not why-evidence.
• The monitoring plan carries numeric alert thresholds per channel — the anti-pattern bans alerts without a stated trigger level; "watch retention weekly" alone would fail the checklist.
• The intervention that flatters the metric is flagged as revenue-relevant — killing paid traffic improves the ratio while cannibalising the growth target, so the output reframes to cost per retained user instead of letting the board celebrate a denominator trick.

---

Supporting file: templates/retention-readout.md

Retention readout: [product/segment] — [date]

Retention defined as: [the value action, not "opened"] per [period matching natural job frequency: daily/weekly/monthly]

Curve summary (cohorts plotted separately — never blended)

Shape: flattening / declining / smile → [what that means here — references/curve-reading.md] Trend across cohorts: newer curves sitting [higher/same/lower] → [the verdict on recent product work]

Segment split (the plateau is the story)

What separates flatteners from leavers

[behavioural differences in the first N days — correlation until tested, labelled]

The one sentence

Cohorts flatten at __% around [period n]; the biggest lever is [plateau/flatten-point] for [segment] via [behaviour].

Actions