Cohort Analysis

Comparing groups who started together, tracked over their own lifetimes.

Data Analyst · Mid-level · core Tech Entrepreneur · Product-Market Fit · core

Also known as: cohort analysis, cohort table, cohort retention

A cohort is a group sharing a starting event — signed up in March, installed v2.4, joined from a campaign — tracked forward along its own lifetime. Cohort analysis lays those lifetimes side by side: March signups at month 0, 1, 2 against April signups at month 0, 1, 2. Calendar date disappears; age does the comparing.

         month0  month1  month2  month3
Mar cohort 100%    62%     48%     41%
Apr cohort 100%    65%     51%      —
May cohort 100%    58%      —       —
(read down columns: same age, different cohorts — the fair comparison)

Columns compare fairly (same age, different eras); rows show a cohort ageing. A rising column means newer cohorts behave better — the product improving. A falling one means the opposite, even if headline totals still grow on the back of acquisition.

The classic mistakes:

Calendar-aligned comparison. March signups in June against June signups in June compares 3-month-olds with newborns. Align by cohort age, always.
Reading rows as improvement. A cohort’s own row declines by construction (users leave over time) — that slope is the product’s retention shape, not a verdict on recent work. Verdicts live in the columns.
Tiny recent cohorts. Last week’s cohort has two data points and enormous noise. Do not brief the all-hands on a cohort younger than the behaviour needs to stabilise.
Survivorship in disguise. Cohorts defined by a later event (“users who upgraded”) select for the engaged. Define cohorts by the starting event, before outcomes are known.
One cohort table for every question. Retention cohorts, revenue cohorts and feature-adoption cohorts answer different things. Build the cohort around the behaviour being studied (see retention analysis, funnel analysis).

The discipline: define by start, align by age, compare down columns, and distrust any cohort too young to have lived. See survival analysis for the formal machinery underneath.