The cohort table: how to read it

One cohort is a row. Many cohorts make a table

In the previous lesson we computed retention for one cohort. One cohort is one row of numbers. Stack every cohort of the product one under another and you get a cohort table — the main instrument through which an analyst looks at retention.

The layout is simple and always the same:

• a row = a cohort, everyone who arrived for the first time on the same day (or in the same week);
• a column = the cohort's age: D0, D1, D7, D14, D30;
• a cell = how many of that cohort were active at that age — in absolutes or as a percentage of the cohort size.

Here is Kotomarket's black box, weekly cohorts, today is March 14. First in absolutes:

Rows of absolutes cannot be compared: the cohorts differ in size. So the table is almost always converted to percentages — each cell divided by the size of ITS OWN cohort (the fixed denominator from the previous lesson):

The size column always stays: without it the percentages hang in the air — exactly as in the lesson about denominators.

Absolute time versus cohort age

The most important idea of this lesson: a cohort table contains two different times, and they are constantly confused.

Cohort age lives in the columns. Moving right along a row is moving through one group's life: day one, day seven, day thirty.

Absolute time (the calendar) lives in the rows: the lower the row, the later the cohort arrived. But a calendar date inside the table lives in neither a column nor a row — it runs along a diagonal.

Check it on March 1. For the March 1 cohort that is D0. For the February 22 cohort it is D7 (Feb 22 + 7 days). For February 15 it is D14. For February 1 it is D28. One calendar date has cut diagonally across the whole table:

            D0    D1    D7   D14   D30
Feb 1        ·     ·     ·     ·     ·        ↖
Feb 8        ·     ·     ·     ·     ·         ↖  a diagonal =
Feb 15       ·     ·     ·   [Mar 1] ·          ↖  one calendar
Feb 22       ·     ·  [Mar 1] ·      ·           ↖  date
Mar 1     [Mar 1]  ·     ·     ·     ·            ↖

Why this matters. Events of the outside world — a release, a tracking outage, a holiday, an ad launch — happen on a specific calendar date and therefore leave a trace along a diagonal, not down a column. If event collection broke on March 1, the D0 cell of the March cohort sags, the D7 cell of the February 22 cohort sags, the D14 cell of an older one sags — and in the D7 column it will look like "one weird cohort", although the problem was shared and lasted a day.

Changes to the product itself — a new onboarding, a different price — affect entire fresh cohorts, that is, they show up in the lower rows. Telling those two patterns apart is half the craft of reading cohort tables.

Why the table is triangular

Look at the shape: the filled part of the table is a triangle, and the lower-right corner is empty. That is not lost data or a broken export.

The reason is exactly the one from the previous lesson: young cohorts have not lived the older ages yet. Today is March 14, therefore:

• the March 8 cohort only has D1 (March 9). Its D7 arrives on March 15 — tomorrow;
• the March 1 cohort has D1 and D7, while its D14 falls on March 15;
• the February 15 cohort will reach D30 only on March 17.

The lower the row, the shorter its filled tail. Hence the triangle — and hence the rule:

A column is compared only across rows where the cell is filled. An empty cell means "not known yet", not "zero".

The mistake nearly everyone makes when first building such a table: computing "average D30 across all cohorts". The empty cells either turn into zeros (the average collapses and a nonexistent retention drop appears) or silently drop out — and then the average is computed over old cohorts only and presented as the overall figure.

There is an opposite bias too: looking only at mature cohorts hides fresh data entirely, and a problem gets noticed a month later than it could have been. So in practice: read the early marks (D1, D7) across all cohorts including the fresh ones, and the late ones (D30) only across mature cohorts. D1 and D7 mature quickly and are the first to show that new generations have got worse.

How to read it: a row across, a column down

A cohort table has exactly two reading directions, and they answer different questions.

A row across — the life of one cohort. 38% → 14% → 10% → 7%. You see the shape of the retention curve: a sharp fall between D1 and D7, then flattening. Nearly every product behaves this way: most newcomers drop off in the first days, while those who survive to the second week leave far more slowly. If the curve reaches a plateau, the product has a core that stays. If it keeps sliding towards zero, there is no core.

A column down — comparing generations at the same age. This is the most valuable movement, because it compares like with like: everyone at exactly seven days old.

column D1:  38% → 37% → 36% → 35% → 33% → 32%
column D7:  14% → 13% → 12% → 11% → 11%

Six consecutive generations — and both columns creep downwards. That is no longer noise: each new generation retains worse than the one before. Meanwhile the size of the last cohort is larger than that of the first (1,200 → 1,500): more people arrive and a smaller share of them stays.

Entry #7 in L.'s journal: "Marketing celebrates DAU growth. I look at the D7 column of the new cohorts: 14, 13, 12, 11. We are pouring water into a leaky bucket and praising the tap."

A first heatmap read in one minute

Cohort tables are often coloured — darker means more, lighter means less. Such colouring is called a heatmap, and the eye reads it faster than digits. The order of a first read:

1. Check the denominators — the first column with cohort sizes. A row of 40 people will jump by tens of percent because of one cat, and it cannot be compared with a row of a thousand.
2. Look at the shape of a row — where the cliff is and where the plateau begins.
3. Walk the columns down — are generations improving or degrading?
4. Find anomalies. One odd row means something happened to that particular set of people (a traffic source, a promo). An odd diagonal means something happened on a specific calendar day (a release, an outage). An odd column usually points at the metric itself, or at that age landing on a weekend.
5. Ignore the lower-right corner — it is empty not because things are bad there.