Automation and AI

Churn Prediction

Also called Attrition prediction, churn modelling

Working out which existing customers are drifting towards leaving, early enough to do something about it.

Quick facts: Churn Prediction

Category
Automation and AI
Also called
Attrition prediction, churn modelling
Level
Advanced
Affects
Retention, revenue forecasting, support priorities
Where to see it
CRM reports, subscription billing dashboards, GA4 audiences, custom models built on order or usage history
In this article4
  1. How churn prediction works
  2. Why churn prediction matters
  3. Common mistakes with churn prediction
  4. How to act on it

How churn prediction works

First you have to say what churn means for your business. For a subscription it is straightforward: the cancellation. For a shop, a clinic or an agency there is no cancellation event, so churn has to be defined as an absence, meaning no order, visit or login for longer than a customer of that type normally leaves it. Getting that definition right is most of the work.

Then you look for what precedes the absence. Falling frequency, shorter visits, unopened email, a support complaint, a failed payment, a change of contact person at a business customer. A model weighs these together and produces a risk score for each customer. As with any propensity model, the score is far more useful as a ranked list to work through than as a verdict on any one person.

Why churn prediction matters

Keeping a customer is nearly always cheaper than replacing one, and the gap widens in markets where advertising is competitive. More importantly, the window to intervene is short and silent. Most customers do not complain before they leave; they simply stop coming, and by the time the revenue figure moves, the decision was taken long before.

It also improves how you spend elsewhere. Knowing which customers are likely to stay lets you estimate customer lifetime value with more confidence, and that number is what tells you how much a new customer is worth acquiring.

Common mistakes with churn prediction

The reflex response is a discount, sent to everyone flagged. That trains loyal customers to wait for offers and hands money to people who had no intention of leaving. The score tells you who to look at; it does not tell you what to do, and those are different questions.

The second mistake is modelling without acting. Plenty of businesses build the score, admire it in a dashboard and never build the intervention, which is where all the value sits.

The third is seasonality mistaken for risk. A customer who buys once a year is not lapsing halfway through the year. A model has to know the natural rhythm of your business, or it will flag most of the list every quiet season and be ignored within a month.

How to act on it

Split the flagged customers by likely reason rather than treating them as one group. Someone who hit a service problem needs an apology and a fix. Someone who has simply forgotten you needs a reminder of what they got from you. Someone whose circumstances changed may not be saveable at all. The message should match the cause, and that means looking at a sample by hand before designing the campaign.

Test the intervention properly by holding back a comparable group who receive nothing, so you can tell whether contact made the difference. Feed what you learn back into the service itself, because the durable fix for churn is usually operational rather than promotional. Where the follow-up is automated, an automated email programme triggered by the risk score keeps the response consistent without anyone watching the list daily.

Do and do not

Do

  • Define churn precisely for your business first
  • Act on the reason, not only the score
  • Hold back a control group to prove it worked

Do not

  • Send discounts to everyone the model flags
  • Treat one quiet month as certain departure
  • Ignore customers who left with no warning signs

Questions people ask about this

What counts as churn if I do not sell subscriptions?

You define it as an absence measured against normal behaviour. Look at how long your customers usually leave between purchases or visits, and set the churn point comfortably beyond that. A restaurant and a car dealership will land on very different definitions. The important thing is to choose one, write it down, and measure it the same way every period.

Can I predict churn without a data scientist?

Often yes, at a useful level. A simple recency rule, flagging customers who have not bought within their normal cycle, catches a large share of the risk and needs only a spreadsheet or a CRM report. Formal modelling adds precision when you have many customers, several behaviour signals, and enough history for the patterns to be dependable.

Should I offer a discount to customers flagged as at risk?

Only sometimes, and rarely as the first move. A discount answers a price objection, but most people leave because of a service problem, a change in their own situation, or simply forgetting. Find the reason before choosing the response, and remember that predictable win-back offers teach customers to lapse on purpose.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.