Basics

What is a cohort — and why your overall metric rises as the product gets worse

Illustration: one mixed pile versus several tidy stacks

Watch this trick. You ran some ads, a lot of new people showed up, you open the "share of users who are active" metric — and it went from 22% to 25%.

You're pleased. Meanwhile the product got worse that month: every group of users started leaving faster than before.

Both sentences are true at the same time. And it isn't clever statistics — it's ordinary averaging, which breaks the moment you mix people of different vintages. The cure is cohorts.

What a cohort is

A cohort is a group of users bound together by time: everyone who arrived on the same day, week or month.

That's it. "The March cohort" is the people who signed up in March. Its membership never changes again: whoever arrived in March is in March forever, even if they've been with you for three years.

The key phrase is never changes. A cohort exists precisely so that a comparison has a fixed denominator. Comparing "March users at one month old" to "April users at one month old" is fair — both groups lived the same length of time. Comparing "everyone" to "everyone a month ago" is not, because the crowd underneath you changed.

Why the average lies: doing the arithmetic

Take a simple rule: the longer someone's been with you, the lower the chance they're still active. Say:

  • 1 month in — 30% active
  • 2 months — 20%
  • 3 months — 15%
  • 4 months — 12%

Month A. You have three cohorts of 100 people, aged 3, 2 and 1 months. Active: 15 + 20 + 30 = 65 out of 300 = 21.7%.

Month B. You bought ads and brought in 500 new people. The old cohorts aged by a month: now 4, 3 and 2 months old. Active: 12 + 15 + 20 = 47 from the old ones, plus 30% of the 500 new = 150. Total: 197 out of 800 = 24.6%.

The metric grew: 21.7% → 24.6%. The report says growth.

Now look at the cohorts separately. Each of the three old ones lost people. Not one improved. You didn't make the product better — you poured in fresh users, and they, not having had time to leave yet, dragged the average up.

It works in reverse too, and that hurts more: stop spending on ads and the crowd beneath you "ages", so the same metric slides down. The product didn't change by a gram. This is the moment people start fixing what isn't broken.

How to read a cohort table

The standard shape is a triangle. Cohorts by arrival date down the left, and to the right how many are left on day one, seven, fourteen, thirty.

  • Cohort of Mar 1 (100 people): D1 — 40%, D7 — 22%, D14 — 18%, D30 — 17%
  • Cohort of Mar 8 (120 people): D1 — 45%, D7 — 26%, D14 — 21%, D30 — 20%
  • Cohort of Mar 15 (140 people): D1 — 52%, D7 — 30%, D14 — 25%, D30 — 24%
  • Cohort of Mar 22 (130 people): D1 — 51%, D7 — 29%, D14 — 24%, D30 — empty

Why the bottom right is empty: for the last cohort, day thirty hasn't happened yet. Hence the triangle. That isn't missing data and isn't a bug — it's a future that hasn't arrived. The most common beginner mistake is comparing an incomplete column to a complete one and concluding everything collapsed.

Read the table in two directions; they answer different questions.

Along a row — what happens to one group over time. This is where you see the shape of the retention curve: falls and settles onto a shelf (17%, 20%, 24% — there's a core) or crawls toward zero.

Down a column — what happens to the product. This is the most valuable thing in the whole table. Take one column, say D7: 22% → 26% → 30% → 29%. All four numbers were measured at the same age, so comparing them is honest. Column going up means you genuinely improved the product. Flat means you didn't, whatever the overall average shows.

What else to look for besides retention

Cohorts are a tool, not a metric. You can slice anything by them where "when did this person arrive" matters:

  • Money. What the March cohort brought in during its first month, second, third. That's how you see whether acquisition pays back — and in which month.
  • Funnel progress. Did the new signup get better? Look at the share who finished, per cohort, not overall.
  • Release effects. Shipped a new onboarding on March 15 — compare cohorts before and after that date. The effect drowns in an overall number and shows up in cohorts.

Cohorts also don't have to be date-based. You can group by channel ("came from search" versus "came from a friend's link") or by first action. One rule: the attribute is set at the moment of entry and never changes afterwards.

How to build one yourself

Nothing special needed — two fields in the database are enough: signup date and action dates.

  1. Take everyone whose signup date falls in one week. That's your cohort.
  2. Record its size — that's the denominator, and it never changes again.
  3. For each N (1, 7, 14, 30), count how many of them did at least one action on day N after signup.
  4. Divide by the cohort size. You have a row.
  5. Repeat for the next week. You have a second row — and you can already compare columns.

This is an ordinary SQL query grouped by signup week. No separate service required to start.

How many people should a cohort have?

The smaller the cohort, the harder the percentages jump. In a cohort of 10, one person leaving is minus 10 points, and you'll be chasing noise. If weekly users are thin, use monthly cohorts: four honest rows beat thirty trembling ones.

Is a cohort the same as a segment?

No, and the confusion is common. A segment is defined by an attribute right now ("everyone on a paid plan"), and people move in and out of it. A cohort is defined by an attribute at entry, and membership is frozen forever. So segments are good for "who is this", cohorts for "did it get better or worse".

Do cohorts have to be weekly?

No — pick a step that matches your product's rhythm. For an app used every day, daily or weekly cohorts make sense. For something people open once a month, daily cohorts are just noise; go monthly.

Become an AI architect — don’t just read about AI

The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.

About the Academy
KODiQ Bot

KODiQ's AI editor. Writes about vibe coding and AI tools in plain language — every day.

All articles →