cohort analysis

A cohort is a group of fans who arrived at the same time, followed afterwards as a block kept apart from everyone else. That separation is the whole idea. Every other agency metric pools people who joined at different moments, under different pricing, handled by different writers, and reports one average over the pile. Cohort analysis refuses the pooling, and in doing so it answers the one question a monthly dashboard cannot: is the work getting better, or is there simply more of it?

Why can’t an overall average tell you that?

Because recruitment moves the average as much as performance does.

Take an agency that genuinely improves its onboarding in March. Its March cohort holds better than every cohort before it. But March was also a big recruiting month, and new subscribers leave faster than long-standing ones, so the overall retention figure for March falls. Same month, same dashboard: the metric says worse, the reality says better.

It works in reverse just as reliably. An agency that stops recruiting watches churn improve for months while its revenue drains, because the only people left are the ones who were never going to leave. Aggregate metrics on a moving base tell you about the base, not about the work.

What does a cohort grid look like?

Rows are the month fans arrived. Columns are how long they have been around, not calendar months. Each cell is the same measurement taken at the same age.

Part of the grid	What goes in it	What it answers
Row	One arrival month, never edited afterwards	Who these people are
Column	Age of the group: month 0, month 1, month 2	Where the drop happens
Cell	Share still subscribed, or cumulative revenue per fan	The measurement itself
Reading down a column	The same age, across arrival months	Whether you are improving
Reading across a row	One group over its life	The shape of a normal life cycle

Reading down a column is the point of the exercise; everything else is preparation for it.

What do you track per cohort?

Three things, and no more at first:

• Share still subscribed at each month of age. This is where you see whether the first renewal (the big drop in any subscription business) is moving.
• Cumulative revenue per fan. Not per month: the running total, which climbs for one group and flattens for another. Two LTV shapes compare; two single numbers do not.
• Share who ever bought. A cohort that stays subscribed and never purchases is a different problem from one that buys immediately and leaves.

Keep the cohorts big enough to mean something. On a small account, group by quarter rather than by month. A grid built on a handful of people per row is noise arranged in a table.

What breaks a cohort analysis?

Four habits, each of which turns the grid into decoration:

1. Reassigning fans between rows. A fan who lapses and returns belongs to his original cohort. Move him and the grid stops being a record of anything.
2. Reading immature cohorts. The most recent row has only lived one month. It will look spectacular, every time.
3. Mixing traffic sources in one row. Fans from a free page and fans who paid to subscribe behave differently enough to cancel each other out inside a cell.
4. Building it by hand, once. A cohort grid is only useful when it updates itself, which is why it belongs in a creator dashboard and not in a spreadsheet somebody rebuilds each quarter.