Most companies do not have a data problem when it comes to churn. They have an integration problem. The reasons customers leave are already sitting in three places inside the organisation. Nobody has put them together.
This post explains what each source can and cannot tell you, and how to combine them into one diagnosis that says who is at risk, what predicts it, and when to act.
What churn rate analysis means
Churn rate analysis has three parts. The first is measurement: how many customers, or how much revenue, you lost in a period. The second is prediction: which of your current customers show the same pattern as the ones who left. The third is explanation: why they left, in their words rather than yours. Most dashboards stop after the first part. The value is in the other two.
The basic churn rate formula
Customer churn rate for a period is the number of customers lost during the period divided by the number of customers at the start of it.
Churn rate = customers lost in the period ÷ customers at the start of the period
Worked example: you start January with 1,000 customers and 40 cancel during the month. Monthly churn is 40 ÷ 1,000 = 4%. Over a year, 4% monthly churn compounds to roughly 39% of the starting base, which is why a small monthly figure deserves attention.
Two definitions to settle before you calculate anything: what counts as a customer (an account, a subscription, a buying unit), and what counts as lost (cancelled, downgraded to a free tier, or inactive for a set number of days). Different teams often use different definitions, and the reconciliation usually takes the first week of a churn project.
Revenue churn
Customer counts treat a small account and a large one the same. Revenue churn does not.
Revenue churn = recurring revenue lost from churned or downgraded customers ÷ recurring revenue at the start of the period
Worked example: 1,000 customers pay SEK 500,000 a month in total. The 40 who cancelled paid SEK 12,000 between them, and another 30 customers downgraded by a combined SEK 8,000. Gross revenue churn is (12,000 + 8,000) ÷ 500,000 = 4%. If expansion from the remaining customers added SEK 15,000, net revenue churn is (20,000 − 15,000) ÷ 500,000 = 1%. When customer churn and revenue churn disagree, the difference tells you which segment is leaving.
Cohort and segment churn
An average churn rate hides the pattern. Cutting the same calculation by cohort (customers who started in the same month) and by segment (size, plan, channel, region) shows where the loss is concentrated. A common result is that churn is low overall but very high in the first ninety days for customers acquired through one channel, which is a different problem from churn at renewal among long-standing accounts, and needs a different fix.
From rate to risk
Once the definitions and the segment view are in place, the analysis moves from counting to predicting. The question changes from "how many did we lose" to "which current customers look like the ones we lost". That is where the three data sources below come in.
Churn is rarely one thing
The first mistake is to look for a single cause. Churn is usually a mix of segments leaving for different reasons at different points in the lifecycle. A price-driven customer who leaves at renewal is not the same problem as a customer who stops using the product in month three and drifts away.
If you treat them as one number, you will design one fix, and it will work for a fraction of the customers you are losing.
A useful diagnostic therefore answers three questions separately: who leaves, what predicts it, and when in the relationship the risk peaks.
Source one: your own customer data
CRM records, transactions, usage logs and support tickets show what actually happened. They are the only source that is not filtered through memory or politeness.
What this data is good at:
- Finding behavioural signals that precede churn, such as falling usage frequency, a drop in order value, a spike in support contacts, or a contract that was downgraded before it was cancelled.
- Timing. The data will show whether risk peaks at onboarding, at first renewal, after a price change, or after a specific service failure.
- Building a predictive model that scores every current customer on likelihood to leave, so the retention effort goes to the right accounts.
What it is bad at: explaining why. The data will tell you that customers who contacted support twice in a month are three times more likely to leave. It will not tell you what they were unhappy about, or whether the problem was the product, the pricing, or the way the complaint was handled.
A practical note: in most companies the first two weeks of a churn project go on agreeing a churn definition and reconciling customer IDs across billing, CRM and product systems. Budget for that.
Source two: what your own people know
Key account managers, sales reps and customer support staff know things that never reach a database. They know which customers have been complaining for six months, which competitor is calling their accounts, which feature was promised and never delivered, and which renewal conversations are going badly.
This knowledge is usually held as anecdotes. It is rarely written down, and when it is, it sits in free-text CRM notes nobody analyses.
Structured interviews with fifteen to twenty-five front-line staff, run by someone outside the reporting line, surface this quickly. The questions are simple: which customers are you worried about, why, and what would have kept the ones we lost?
Two cautions. First, staff explanations are partial. Sales will say price, support will say product quality, product will say sales oversold. Each is partly right. Second, this source needs to be handled without blame, or people stop talking.
Source three: what customers say
Sometimes you simply have to ask. Short, targeted research with lost customers and at-risk customers fills the gap the first two sources leave: the customer's own account of what went wrong and what would have changed the outcome.
Keep it short and specific. Twenty-minute interviews with twenty to thirty lost customers, plus a brief survey to a larger sample of current customers, is usually enough. The aim is not a representative satisfaction study. It is to test the hypotheses generated by the data and the internal interviews.
The main limitation is that customers do not always do what they say. A customer who says they left over price may have been disengaged for months before the price came up. That is why this source should never stand alone, and why it works best when it is fielded after the first two sources have narrowed the questions.
Combining the three into one diagnosis
The value comes from the overlap. In practice the sequence looks like this:
- Model the customer data to find the behavioural signals and timing of churn, and to segment leavers.
- Interview internal staff to explain the signals and add what the data cannot see.
- Run short customer research to confirm or reject the explanations, segment by segment.
- Build a risk score for current customers and a short list of actions per segment, each with a trigger and an owner.
The output is not a report. It is a list of customers to contact this month, a reason for contacting each group, and a change or two to product, pricing or service that removes a recurring cause.
For a Nordic digital-services company this approach reduced annual churn by 18 per cent within one year. Most of the gain came from two segments and three actions, none of which were visible in the aggregate churn number.