Why cohorts matter
A single retention number like “70% one-year retention” hides far more than it reveals. The 70% blends customers acquired five years ago (when product, brand, and audience were different) with customers acquired last month. Trends are invisible because every period averages with every other period.
Cohort analysis fixes this. Group customers by when they joined, track each group’s behavior separately over time, and trends jump out. The January cohort might be retaining better than the June cohort because the product improved. The Q3 cohort might be retaining worse because a paid-acquisition push brought in lower-quality customers.
Almost every interesting question about customer behavior is a cohort question.
How cohorting works
The classic shape is a cohort table:
Cohort M0 M1 M2 M3 M4
Jan 2026 100 62 48 41 37
Feb 2026 100 65 52 45 -
Mar 2026 100 68 55 - -
Apr 2026 100 71 - - -Each row is a cohort (acquisition month). Each column is months since acquisition. The values are typically % retained, % active, average spend, or some other tracked metric.
Reading the diagonal vs. reading down the column tells you different things:
- Across a row = how one cohort’s behavior evolves
- Down a column = how each successive cohort compares at the same lifetime stage
The “down a column” view is what reveals trends. If Jan-2026’s M3 retention is 41% and Apr-2026’s M3 retention is 50%, you’re improving.
Cohorting dimensions beyond acquisition month
The most common cohort is acquisition period, but cohorts can be defined on any attribute:
- Acquisition channel, Meta-acquired vs. organic-acquired cohorts often have wildly different retention
- First-purchase product, subscribers vs. one-off buyers
- Geography, DACH vs. US cohorts may differ structurally
- Cohort × cohort, Jan-2026 × Meta-acquired vs. Jan-2026 × organic
Multi-dimensional cohorting gets powerful fast. It also gets noisy fast, make sure each cell has enough customers to be meaningful.
Common cohort use cases
- Retention monitoring. Tracking M30, M60, M90 retention by cohort, looking for trend shifts.
- LTV forecasting. Use early cohorts (with full history) to extrapolate LTV for recent cohorts (still developing).
- Channel quality assessment. Cohorts split by acquisition channel reveal which channels bring in healthy customers vs. which bring in churners.
- Product change impact. Pre-launch vs. post-launch cohorts reveal whether a product change moved customer behavior.
Common mistakes
- Blended retention reporting. Hides exactly what you need to see.
- Cohorts too small to be meaningful. A weekly cohort with 30 customers can’t tell you much. Aggregate up.
- Comparing cohorts at different lifetime stages. A young cohort’s “60% retention” is meaningless compared to a mature cohort’s “60% retention” if one is M3 and the other is M24.
FAQ about Cohort
What is a customer cohort?
A cohort is a group of customers sharing a common attribute, typically the period they were acquired in. Cohort analysis tracks each cohort’s behavior separately over time to reveal trends that blended numbers hide.
Why is cohort analysis better than averaging?
Blended averages mix customers acquired years ago with customers acquired last month, trends become invisible. Cohorts reveal whether retention is improving, whether channel quality is shifting, and whether product changes moved customer behavior.
What size should my cohorts be?
Big enough to be statistically meaningful. Daily cohorts are usually too small. Weekly or monthly cohorts work for most e-commerce businesses. Multi-dimensional cohorts (e.g. month × channel) need extra care that each cell has enough customers.