What Is Cohort Analysis?

What does cohort analysis mean?

Cohort analysis groups users by a shared characteristic, usually the week or month they signed up, then tracks how each group behaves over the following periods instead of treating all users as one blended average. The Corporate Finance Institute puts the purpose simply: "Companies use cohort analysis to analyze customer behavior across the life cycle of each customer." A total active user count mixes last week's signups with last year's, so a growing top of funnel can hide a shrinking base for months; a cohort table keeps them apart.

The reason it matters is how fast early drop-off happens. Andrew Chen, publishing retention data from Quettra drawn from over 125 million Android devices, reported that "the average app loses 77% of its DAUs within the first 3 days after the install," and 90 percent within 30 days. That data covers Android apps in 2015, not B2B software, so do not use it as a benchmark for your product. What it still shows is that most of the story sits in the first few columns of a cohort table, which is exactly the part an aggregate number smooths over.

How do you read a cohort table?

A cohort table has one row per cohort (everyone who signed up in a given week), one column per period since signup (week 1, week 2 and so on), and in each cell the share of that cohort still active in that period. The table below is illustrative, for a fictional team notes app, not benchmark data.

Signup week	New accounts	Week 1	Week 2	Week 3	Week 4	Week 5
5 Jan	410	51%	40%	35%	27%	33%
12 Jan	395	50%	39%	28%	33%
19 Jan	430	59%	41%	44%
26 Jan	445	49%	50%
2 Feb	400	60%

The empty cells are not zeros: those cohorts have not reached that age yet. The bold cells all fall in the same calendar week, the week of 2 February. Read the same table three ways and it answers three different questions.

Across a row: where does one cohort drop off, and where does it flatten? The 5 January cohort loses about half its accounts in week 1, then declines more slowly and holds in the low 30s. The level where a row stops falling is the cohort's real retention. The steep first step is an activation question; the flat tail is the base of accounts that found lasting value.

Down a column: are newer cohorts doing better? Week 1 retention sits near 50 percent for the first two cohorts, then jumps to 59 and 60 percent for the 19 January and 2 February cohorts. If an onboarding change shipped on 19 January, this column is the evidence that it worked, because it compares cohorts at the same age.

Along a diagonal: did something hit everyone at once? The bold cells dip regardless of when each cohort signed up (27, 28, 41 and 49 percent against their neighbours), and the rows recover the week after. A drop that follows the calendar rather than cohort age points to an outage, a holiday, a pricing email or a release, not to product decay. The 26 January cohort's week 1 figure of 49 percent is a diagonal effect; read on its own it would wrongly suggest the onboarding change had stopped working.

Where should you look for the real retention number?

Look for where the curve flattens, not at any single cell. A cohort that falls from 100 to 51 to 40 and then holds at 33 has a retention floor of about a third; a cohort that keeps sliding a few points every week has no floor yet, which is a much bigger problem even if its week 4 number looks similar today. Plot each row as a line and the difference is obvious in a way a table of percentages is not; a curve that dips and then rises again is covered in the entry on the smiling retention curve.

Andrew Chen drew the practical conclusion from the same Quettra data: "the best way to bend the retention curve is to target the first few days of usage, and in particular the first visit." The flat tail is hard to move directly. The height at which it flattens is mostly decided by how many accounts reach real value in their first days, which is why cohort work and activation metric work belong together.

What should count as active in a cohort table?

Match the definition of active to how often the product should be used. A tool people open every working day should count someone as active only after several active days in a week; a weekly planning tool can count one meaningful session a week; a monthly reporting tool should use monthly periods. A daily threshold on a monthly product makes healthy accounts look churned, and a loose threshold on a daily product makes drifting accounts look healthy.

Whatever you choose, count a meaningful action, not a login. And change the definition deliberately: tightening it lowers every cell, so compare old and new tables only when both use the same rule.

How should you segment a cohort table?

Split the table before drawing conclusions. The three splits that most often change the answer are fit (accounts that match your ideal customer profile against those that do not), engagement tier (heavy, regular and light users), and unit of analysis (users against accounts). A cohort whose overall row looks mediocre can hide an ideal-customer segment that flattens high and a poor-fit segment that falls to zero, and those call for opposite actions: invest in onboarding for the first, change targeting for the second.

In B2B products, build the account-level table too. An account with five seats can lose two users and still renew, so the account row is the one that predicts revenue. The retention analysis guide covers how to turn these splits into decisions, and AI for cohort analysis covers drafting them faster.

What are the common mistakes when reading cohort data?

The most common is reading one cohort's dip as a product-wide problem, when a column or diagonal read would show it was one signup period or one calendar week. The second is comparing a young cohort's latest cell with an old cohort's final cell: compare cohorts only at the same age. The third is trusting small cohorts. A weekly cohort of 40 accounts can swing ten points on a handful of accounts, so group small weeks into months before you read trends.

The honest limitation of cohort analysis is that it shows when and for whom retention changed, not why. The why comes from pairing the table with events (what did retained accounts do in week 1 that churned ones did not?) and from talking to customers in the cohort that moved.

Cohort analysis example

The practical challenge in Builders Camp's Product Analytics bootcamp gives members three retention patterns from a B2B field service tool and asks them to interpret the cohort data using row, column and diagonal perspectives: which view reveals each pattern, what it means, and what to do about it. The write-up of that challenge shows the full brief.

Where does cohort analysis fit in Builders Camp's bootcamps?

The Product Analytics bootcamp is a one-week program with 2 live sessions and 9 microlessons, directed by Mário Araújo. Its published topics include cohort retention analysis, engagement segmentation into high, medium and low tiers, and account-level analysis, alongside activation rate and B2B instrumentation. Data for Product Managers is a two-week bootcamp with 4 live sessions and 8 microlessons, directed by Andre Albuquerque, whose published topics include retention and cohort analysis, funnels and activation analysis, and segmentation.

Builders Camp runs live and self-paced bootcamps in product management and AI product building. See the Product Analytics bootcamp for the next cohort dates.