Churn: Defining It Before Predicting It
By the end of this lesson, you should be able to: tell the two kinds of churn apart, pick an inactivity threshold that fits the decision, turn a monthly churn rate into a yearly one correctly, and explain why churn has to be defined before anyone tries to predict it.
Churn counts the people who left. For most products, nobody tells you they've left, so churn has to be inferred.
Two kinds of products, two kinds of churn
The first question is whether your product has a moment where users tell you they're leaving.
Products with a contract. Subscriptions like a music or video streaming plan, or a software license. A user who cancels has clearly churned, and you know the exact date. This is called contractual churn.
Products without a contract. Most apps, games, social networks, shops and marketplaces. Nobody tells a food delivery app they're done ordering. They just stop opening it. This is called non-contractual churn, and here churn has to be inferred from silence.
| Contractual | Non-contractual | |
|---|---|---|
| Examples | Paid subscriptions, software licenses | Free apps, games, shops, marketplaces |
| How you know they left | They cancel | They stop showing up |
| Main question | When did they cancel, and why? | How long a silence counts as leaving? |
Even subscriptions have a quieter version. Involuntary churn is when a subscription ends because a payment failed, like an expired card, not because the user chose to leave. It's often worth measuring on its own, because it's fixed by better billing, not a better product.
Defining churn when nobody cancels
For a product without a contract, you need a rule: how many days of inactivity count as churned?
There's no single right answer, and the choice changes the number a lot. Count someone as churned after 7 quiet days, and you'll report a lot of churn, including many people who were just on holiday. Wait 90 days, and you'll report much less churn, but you'll find out very late.
Real companies pick a threshold and write it down. Duolingo's growth model, for example, defines dormant users as learners who have been inactive for at least 30 days. Anyone who comes back after that counts as resurrected, which is the subject of the next lesson.