Run a Cohort and Retention Analysis
Builds a cohort retention analysis with the definitions pinned down first — what counts as active, when the clock starts, how to read the curve — and separates a genuine retention problem from a mix or acquisition problem. Use it to find out whether the product keeps people, not just whether it attracts them.
0 likes
0 dislikes
Sign in to rate this prompt
Prompt
You are an analyst who specializes in retention. Growth in totals hides everything that matters; cohorts are how you see whether the business actually keeps people.
The business and what a customer does: {{business_description}}
Expected usage frequency — daily, weekly, monthly, seasonal: {{usage_frequency}}
Data available per user or account: {{available_data}}
What I am worried about: {{concern}}
Set it up properly before calculating anything.
1. **Pin down the definitions**, because every one of these changes the answer:
- What starts the clock: signup, first purchase, activation, or first value delivered.
- What counts as retained in a period: any login, a meaningful action, or a repeat purchase. Choose the one that reflects value received, not presence.
- Period length, matched to natural usage frequency. A weekly grid on a quarterly product manufactures churn that is not there.
- How to treat users too recent to have a full period. Partial cohorts at the edge make the newest data look worse than it is, and this trips people up constantly.
2. **Build the cohort table.** Cohorts as rows, periods since start as columns, retention as a percentage with the absolute counts shown alongside. Small cohorts produce dramatic-looking percentages that mean nothing, so the raw numbers have to be visible.
3. **Read it in three directions.**
- **Across a row** — the shape of one cohort's decay. Where does it drop hardest, and does it flatten? The flattening point is the real question: a curve that flattens has a durable core, a curve that keeps declining has none.
- **Down a column** — is retention at period N improving for newer cohorts? This is the only clean read on whether product and onboarding changes are working.
- **Diagonally** — a bad month that hit every cohort at once, which usually means something external or a release, not a cohort quality difference.
4. **Segment the cohorts** by acquisition channel, plan, first action taken, or size. Blended retention is an average of very different curves, and the interesting finding is nearly always that one segment retains far better. That segment is the actual business.
5. **Separate the three problems.** Poor retention, poor activation (users who never really started, dragging the curve down), and a mix shift where a lower-quality channel grew. Say which mine looks like based on the shape, because they call for entirely different responses.
6. **Sanity-check it.** Resurrection and returning users, whether churn is voluntary or payment-related, whether a period of inactivity is genuinely churn for this product, and the survivorship trap of judging a young cohort by an old one's curve.
Return the analysis, the two or three findings that would change what the business does, and what to look at next.