Analytics & Retention
Why averages lie and cohorts tell the truth. How to build cohort retention curves, what shapes mean what, and how to use the analysis to make real product, marketing, and lifecycle decisions.
Aggregate metrics — "average customer retention is 65%" — hide the patterns that matter. A 65% average can be 90% for engaged users and 30% for unengaged users, or 65% across the board. The actions you take in each case are completely different.
Cohort analysis splits users into groups (cohorts) defined by some shared characteristic — usually signup date, but also acquisition channel, plan tier, persona, or feature use — and tracks each group's behavior over time. The patterns inside the average become visible.
The most common form is a retention table: rows are signup cohorts (Jan 2026 users, Feb 2026 users, etc.), columns are weeks or months since signup, cells show the percentage of that cohort still active.
A healthy SaaS retention curve looks like this: high week 0 retention (everyone is active immediately after signup), sharp drop in weeks 1–4, then gradual flattening to a stable plateau. The plateau height is what you care about — if it's above 0, you have a retained user base; if it asymptotes to 0, the product doesn't stick.
Smile curve (asymptote > 0). Users churn early, but the ones who stay stay forever. Product has long-term value for the right user.
Slope curve (slow steady decline). No floor. Every cohort eventually decays to zero. Common in products with one-time use cases.
Cliff curve (drop and dead). Sharp early drop, then near-zero retention. Activation isn't happening; product isn't delivering value.
The real value of cohort analysis is comparison — comparing one cohort to another to see what works:
Defining "active" badly. If "active" means logged in once, almost everyone counts. If "active" means did the core valuable action, the picture sharpens. Pick a definition that captures real value delivery.
Not enough cohort age. A 2-month-old cohort can't tell you about 6-month retention. Look at older cohorts for long-horizon answers.
Tiny cohorts with high noise. A 50-user cohort's retention curve is mostly noise. Wait until cohorts are statistically meaningful.
Treating retention as the only metric. Cohort revenue per user (LTV cohorts) sometimes tells a different story than retention cohorts — high-retention low-value users are not the same as low-retention high-value users.
Mixpanel, Amplitude, Heap, PostHog, and most modern product analytics platforms have cohort retention as a default report. You can build it in SQL against your data warehouse for full custom control. Spreadsheet versions work for small datasets.