Retention cohort analysis without a spreadsheet

Group customers by when they arrived and watch how each group holds up over time.

A blended retention number averages over every change you have ever made. Cohorts pull those changes apart and show you which month improved things.

Run this with your AI assistant

meelu-analytics-mcp is a free, open source MCP server that runs on your machine. One command installs it and configures your assistant.

meelu-analytics-mcp View on GitHub All open source

The problem

Why this is hard to answer today

One retention figure for the whole business mixes customers who arrived under completely different products, prices and onboarding flows, and reports the average as though it described something.

A blended number moves for reasons you cannot separate — it might be retention improving or just a big recent intake.
Building the cohort grid by hand is a morning of spreadsheet work that is stale within a week.
Without cohorts you cannot tell whether the onboarding change you shipped in March did anything at all.

The solution

What the engine does instead

Cohort analysis is built in, not assembled. Group by arrival period, follow each group forward, and compare the curves.

A retention matrix built from raw customer rows — no pivot tables.
Curves per cohort, so an improvement shows up as a group that departs from the ones before it.
Changepoint detection to find when the shift happened rather than eyeballing it.
A trust rating that accounts for thin cohorts, because a cohort of eleven people is not evidence.

Built on Customer analytics, one of nine tool categories in the server.

How it works

Step by step

Step 1
Export customers with dates
You need when each customer arrived and either when they left or when they were last active. Raw rows are fine — you do not need to build the grid yourself.
customers.csv
```
customer_id,signup_date,last_active,churned_date,plan
CUST_001,2024-01-15,2026-09-01,,team
CUST_002,2024-01-22,2024-06-30,2024-07-01,starter
CUST_003,2024-02-03,2026-08-28,,pro
```
Step 2
Install the MCP server
Install meelu-analytics-mcp once, then restart your assistant. The install page has the one command and the README covers the rest.

Step 3

Ask for the cohorts

Ask for the grid and the comparison together — the grid alone rarely tells you what to do.

In your assistant

Using meelu, load ~/Downloads/customers.csv and show me
retention by signup month. Which cohorts held up best,
when did retention change, and is the difference between
the good and bad cohorts real or just noise?

Example output

What comes back

Shape of an answer — illustrative numbers, real structure.

Retention by signup month (% still active)

  Cohort     n     M1    M3    M6    M12
  2024-01   112   88%   71%   58%   44%
  2024-02    98   87%   69%   55%   43%
  2024-03   134   91%   80%   72%   61%
  2024-04   127   92%   81%   74%    —
  2024-05   141   90%   82%   75%    —

Changepoint: 2024-03. Cohorts from March on retain
materially better at month 6 and beyond.

Is it real? Yes — the March-onward difference at month 6
holds up under test, not explained by cohort size.

Trust: high — 612 customers across 5 cohorts.
Caveat: M12 is missing for recent cohorts by definition,
not because they churned.

That last caveat catches the most common cohort mistake — reading an empty cell in the bottom-right of the grid as a bad result rather than as time that has not passed yet.
Knowing March is the changepoint turns a chart into a question worth asking: what shipped in March, and can you do more of it?

In practice

Who asks this

Product lead

You reworked onboarding in March and want to know whether it landed.

Compare cohorts either side of the change. If the curve moved, you have something; if not, better to know before you build the follow-up.

Founder writing an update

Retention is “improving” but you cannot show it convincingly.

A cohort grid where recent rows sit above older ones is the clearest way to show it, and it is much harder to argue with than a blended average.

Growth analyst

Retention dropped and the obvious suspect is a channel that scaled up.

Cut cohorts by channel as well as month. A drop that is really a mix shift looks exactly like a retention problem until you split it.

FAQ

Common questions

What is cohort analysis?

Cohort analysis groups customers by when they arrived, then follows each group forward over time. Because everyone in a cohort started at the same point and under the same conditions, comparing cohorts shows whether things are getting better or worse. A blended average cannot do this: strong growth in new customers will mask falling retention indefinitely.

What is a retention cohort?

A retention cohort is a group of customers who signed up in the same period, measured by how many of them are still active in each later period. Plotted as a grid, each row is a signup month and each column a month since signup. Reading down a column compares the same point in the customer lifecycle across cohorts, which is the comparison that tells you whether a product change worked.

How do you read a cohort table?

Read across a row to see one group decaying over time, and down a column to compare groups at the same age. Healthy retention curves flatten out rather than falling to zero, and the level where they flatten is the share of customers who stay indefinitely. The empty bottom-right of the grid is not churn: those months simply have not happened yet, and reading them as loss is the most common way cohort tables get misused.

Should cohorts be weekly or monthly?

Monthly suits most subscription businesses, because billing is monthly and cohorts stay large enough to mean something. Weekly works if you have high volume and a short usage cycle, such as a consumer app. The binding constraint is size: small cohorts move around for reasons that have nothing to do with retention, so a few dozen customers per cohort is the practical floor.

What is a good retention rate?

For business software sold on monthly subscriptions, 90% or better of customers retained each month is healthy, and the revenue figure should be higher still once expansion is counted. Consumer apps sit far lower. More important than the level is the shape: a curve that flattens means you have found a group the product genuinely serves, and a curve that keeps falling means you have not.

Keep going

Related use cases

Churn prediction

Work out what predicts a customer leaving — and who is at risk now

Customer segmentation

RFM segments and natural groupings found in your data

Revenue forecasting

Project revenue forward with trend, seasonality and honest error bars

Try cohort retention on your own data

Install meelu-analytics-mcp once and every use case on this site is available to your assistant — it is one server with forty-five tools, not six separate products.