Product Analytics pack · Stage 2. analytics-question-framing → metric-root-cause / cohort-retention → opportunity-sizing → experiment-design → insight-storyteller

Cohort Retention

Averages hide retention problems: a growing product can look healthy while every new cohort churns faster. Cohorts show whether the product is getting better at keeping people.

Step 1 — Define it precisely

Agree on four things before computing anything:

Step 2 — Build the table

If the data is a CSV with one row per activity event (user_id, event_date, optionally signup_date and a segment column), run scripts/cohort_table.py (in this skill's folder):

python scripts/cohort_table.py events.csv --user user_id --date event_date --period month
python scripts/cohort_table.py events.csv --user user_id --date event_date --signup signup_date --period week --segment channel

It prints a retention triangle (cohort × periods since start, as % of cohort size) and can save it as CSV with --out. If there's no signup column, each user's first event counts as their start.

Leave cells blank where a cohort hasn't had time to reach that period. Don't treat them as zero.

Step 3 — Read the curves

Look for these patterns and say which ones apply:

• Flattening (good): the curve drops early, then levels off. The level it settles at is the core retained audience. A curve that keeps declining toward zero suggests weak product-market fit for that cohort.
• Early drop: most loss happens between period 0 and 1. That points to onboarding and first value, usually the cheapest place to improve.
• Newer cohorts better or worse: compare the same period across cohorts (read down a column). Improving columns mean product changes are working.
• Smile curve: retention rising again in later periods (reactivation, seasonality, or a new feature bringing people back).
• Segment gaps: large differences by channel or plan often matter more than the overall trend.

Flag small cohorts (fewer than ~100 users) as noisy.

Step 4 — Quantify

• Retention at key milestones: e.g. week 1, week 4, month 3.
• Change across cohorts at the same milestone, in percentage points.
• Optional lifetime value: if revenue per active user per period is known, LTV ≈ Σ (retention in period t × revenue per active user) over the horizon. State the horizon and whether it's discounted.

Output: retention summary

# Retention: <product / population>, <period range>
**Definition:** cohort = <…>, return event = <…>, period = <…>
**Headline:** <one sentence, e.g. "Month-1 retention improved from 31% to 38% for cohorts after the June onboarding change; curves flatten around 22%.">

## Retention table
<triangle, or a link to the CSV>

## What it shows
- <pattern 1, with numbers>
- <pattern 2>

## Segments
| Segment | M1 | M3 | Note |
|---|---|---|---|

## Recommended actions
1. <action tied to the biggest drop-off>

When done

Suggest metric-root-cause if a cohort or segment suddenly got worse, or opportunity-sizing to value a retention improvement.