Churn rate
The share of customers or revenue you lost in a period. Retention's evil twin: if 95% stayed, 5% churned.
See it
What it is
Churn rate is the share of what you had at the start of a period that was gone by the end. Count it two ways and they will disagree. Logo churn counts customers: 5 of 100 accounts cancelled, 5%. Revenue churn counts money, which matters more when one enterprise account is worth two hundred hobbyists. Net revenue churn nets expansion from your existing customers (upgrades, seats added, usage overages) against the revenue you lost, which is cancellations plus contraction from downgrades. It goes negative when the expansion is bigger than churn and contraction combined, meaning the customers you already had are worth more this month than last even though some of them left. New customers never count toward it.
For subscriptions the churn event is obvious: a cancellation. For everything else you have to invent it, usually as an inactivity rule ('no session in 30 days'). That window is a business decision, not a data one, and a weekly tool and a tax tool need wildly different numbers.
Gotcha: monthly churn does not multiply by 12. Five percent monthly is not 60% a year, it is 1 minus 0.95 to the twelfth power, about 46%. Watch the denominator too: customers at the start of the period, customers at the average of the period, and customers who were even eligible to cancel give three different answers, so pick one and write it down. And with a small base, churn is mostly noise. Two cancellations out of forty is a 5% swing that means nothing.
Ask AI for it
Build me a churn rate model for my business (described below). Define logo churn, gross revenue churn, and net revenue churn with the exact formula and the exact denominator for each, plus the churn event definition (cancellation, or an inactivity window I should choose). Show a worked monthly example with real-looking numbers where logo churn and revenue churn point in opposite directions, and explain why. Convert monthly to annual correctly using compounding, not multiplication. Then work out how big a population I need before the number is worth acting on: take my baseline churn rate and my target confidence-interval width (or the smallest change I need to detect) and calculate it. If I have not given you those inputs, do not invent a round threshold; show the formula, say what it depends on, and leave the number unresolved until I supply them. Finish with 3 segment breakdowns most likely to show me where the churn is actually coming from.