Cohort analysis is the antidote to misleading aggregate metrics. Instead of "our overall retention was 40% last month" — which blends new and old users and can't be compared period-over-period — you look at "users who installed in week X retained at 45% by day 30". That's a number you can compare to the same-age cohort from a month ago, a quarter ago, a year ago. Aggregate metrics hide trends; cohort metrics expose them.
That 4.6-point D1 drop over 18 months is what cohort analysis exists to surface. Aggregated catalog stats would mix old "good" cohorts with new "worse" ones and report a midpoint number that hides the trend entirely. By holding each cohort separate and comparing the SAME slice (D1) across cohorts, the secular decline becomes obvious — likely a combination of ATT-era attribution loss, lower-intent paid traffic mixing in, and market saturation.
Three common cohort definitions
- Install cohorts — group users by install date. Default for retention curves, the broadest view of product health.
- Purchase cohorts — group by first-purchase or first-subscription date. The right view for monetization: track cumulative revenue per user from the moment they paid, comparing across acquisition periods.
- Feature-adoption cohorts — group by when users hit a defined product milestone (completed onboarding, used feature X, hit habit-loop trigger). Useful for product-led retention analysis.
Many consumer apps track all three in parallel; they answer different questions.
The "revenue triangle": mature subscription apps track a cohort revenue table where each row is a cohort (by month or week) and columns are cumulative revenue per user at D30, D60, D90, D180, D365. The structure looks like a triangle because newer cohorts have fewer observation points. Reading the triangle:
- Compare same-age columns across rows to see whether monetization is improving (later rows hit higher cumulative revenue at D30 than earlier rows) or worsening.
- Watch the long-tail asymptote — by D180-D365, cumulative revenue should be plateauing for healthy retention shapes.
- Off-trend cohorts indicate either a data / attribution bug OR a real compositional shift in acquisition mix.
Mean-reversion in cohort curves: cohort tables typically show that revenue and retention from recent cohorts converge to a stable steady-state shape. If one cohort is wildly off-curve, two explanations:
- Data / attribution bug — most common. Check that the cohort definition is consistent, that revenue events are firing correctly, that no MMP / analytics changes coincided with the cohort start.
- Real compositional shift — the acquisition mix changed (new ad creative, new geo, new ASA campaign), pulling in a different user type. Investigate by segmenting the cohort by acquisition source.
Never assume an off-curve cohort is just noise — investigate every time.
Tools and implementation: every major analytics platform (Amplitude, Mixpanel, Heap, GA4, Singular, Adjust, AppsFlyer) exposes cohort views as core functionality. For deeper analysis (custom cohort definitions, multi-event cohorts, revenue-cohort triangles), most teams build SQL views in their warehouse (BigQuery, Snowflake, Redshift). Spreadsheet-based cohort tracking still works for small apps — but doesn't scale past a few thousand users per cohort.
Cohort types — what each answers
| Cohort type | Grouping basis | Key question it answers | Best for |
|---|---|---|---|
| Install cohort | By install date | How does retention curve compare across acquisition periods? | Broad product-health monitoring (default) |
| Purchase cohort | By first-purchase date | How does cumulative revenue per paying user compound over time? | Subscription monetization (LTV modeling) |
| Feature-adoption cohort | By date user hit defined product milestone | Does the feature drive retention vs non-adopters? | Product-led growth, feature impact analysis |
| Channel cohort | By acquisition channel (Meta vs TikTok vs organic) | Which paid channel produces highest-LTV users? | UA channel-mix optimization |
| Geo cohort | By install country | How does retention / LTV vary across markets? | International expansion + localization investment |
| Creative cohort | By ad creative the user converted from | Does this creative produce better users beyond install volume? | Creative quality vs install-only optimization |
Mature analytics programs maintain 3-5 cohort dimensions in parallel. Install cohorts are the default; layering channel + geo cohorts on top reveals which paid acquisition is genuinely producing high-quality users vs just inflating top-of-funnel.
The single-metric line above is the simplest cohort comparison — D1 retention across consecutive install cohorts. The flat plateau through 2024 followed by a step-down in 2025-Q1 is the kind of pattern that demands investigation. If you saw this on your own product you'd dig into acquisition-mix changes, OS update timing, or measurement methodology shifts around that exact quarter.
Median D1 / D7 / D30 retention across 7 quarterly install cohorts (US, MWM)
| Cohort | D1 | D7 | D30 | Apps measured |
|---|---|---|---|---|
| 2024-Q1 | 31.1% | 10.1% | 4.4% | 200K+ |
| 2024-Q2 | 31.2% | 10.1% | 4.4% | 200K+ |
| 2024-Q3 | 31.2% | 10.1% | 4.3% | 200K+ |
| 2024-Q4 | 29.1% | 9.8% | 4.6% | 300K+ |
| 2025-Q1 | 26.2% | 8.6% | 3.7% | 300K+ |
| 2025-Q2 | 26.3% | 8.7% | 3.8% | 300K+ |
| 2025-Q3 | 26.5% | 8.6% | 3.7% | 300K+ |
The table version of the same data is the cohort triangle in its simplest form. Each row is a cohort, each column an age. Reading horizontally tells you what one cohort experienced over time; reading vertically (D1 column, top to bottom) tells you how the SAME-AGE behavior changed across cohorts. That vertical read is the real power — and where the −4.6-point story is hiding.