Customer Churn Prediction: Signals, Models and Tools

By the time a customer cancels, the warning signs are often weeks old. Logins thinned out. Orders came later than usual. A support ticket went unanswered for too long. Churn prediction is the practice of catching that pattern while the customer is still around to hear from you.

This guide takes the marketing and retention view, not the data science one: which signals matter, how the models think, what to do with a score, and which tools already do the math. For the wider family of models, see our guide to predictive analytics in marketing.

Define churn before you predict it

A model can only predict an outcome you define precisely, and "churn" means different things in different businesses. A subscription product has a clear event: the customer cancels or does not renew. A store or a free app does not, so churn has to be defined as a gap, such as no order in 120 days or no session in 30.

Braze makes this explicit. Its Predictive Churn asks you to write the churn definition as something a user does (uninstalls, unsubscribes, disables an account) or stops doing (starts a session) within a churn window of up to 60 days. Adobe's Customer AI likewise lets you define custom churn goals. Separate voluntary churn (the customer chose to leave) from involuntary churn (a card failed), too. The second is a billing problem with its own fixes and should not be mixed into the same model.

Signals that predict churn

Change matters more than level. A customer who has always logged in weekly is fine at weekly; a customer who went from daily to weekly is not. Good models, and good health scores, compare each customer with their own history as well as with everyone else. The SaaS churn reduction strategies guide covers the product side of acting on these signals.

How churn prediction models work

There are three common approaches, and they are not mutually exclusive.

• Rules. "No login in 14 days" triggers a journey. Transparent and quick to set up, but blind to anything you did not think of.
• Health scores. You choose and weight metrics such as usage, support history and survey results into one number per account. Common in B2B customer success.
• Machine learning models. The model looks at customers who churned and customers who stayed, learns which behaviors separated them, and scores active customers. Braze, for example, documents gradient boosted decision trees for Predictive Churn.

Two settings shape every model. The prediction window is how far ahead it looks: GA4's churn probability, for example, predicts whether a user active in the last 7 days will be inactive in the next 7. The audience is who gets scored: Braze only scores users who are still active under your definition, so a customer who already churned is not in the list. Short windows suit apps with daily use; subscriptions with annual contracts need a much longer view.

What to do with a churn score

A churn score is a sorting tool. It tells you who to look at first, not what to say. Decide the playbook by score band before you see individual names:

• Low risk. Leave them alone, and keep them out of discount campaigns. They were staying anyway.
• Rising risk. Send a product nudge: a feature they have not tried, a reminder of what they set up, a tip tied to their use case.
• High risk, high value. Escalate to a human. A CSM call or a personal email from the account owner beats another automated sequence.
• High risk, lower value. Run an automated retention journey and, where it fits, a win-back offer. Our list of retention emails that win back lost customers has templates.
• Already lapsed. Move them out of retention and into a win-back program with its own goals.

Then measure honestly. Hold back a random share of each high-risk band from the campaign and compare churn between the two groups after the prediction window closes. If the treated group does not churn less, the score may be fine but the action is not working. Small SaaS teams without a lifecycle marketer can hand the sending to a service: Questera, an AI marketing team for SaaS, lists welcome, onboarding and win-back email among the work it runs (Growth plan $49 a month as of October 2026). It is not a churn-scoring tool, so pair it with one of the predictions below.

How to start, step by step

1. Write your churn definition. An event or an inactivity gap, plus a time window that matches how often customers normally buy or use the product.
2. Split voluntary from involuntary churn. Handle failed payments with billing tools and dunning, not the churn model.
3. Check your signals are tracked. Sessions, key feature events, orders and support data need consistent names and a shared customer ID.
4. Use the prediction your tools already have. Turn on the churn feature in your engagement, analytics or CS platform before building anything.
5. Set score bands and actions. Decide in advance what each band receives and who owns it.
6. Hold out a control group. Keep a random slice of each band out of the campaign.
7. Review after each window. Compare churn in treated and held-out groups, and check that high scores still match real churn.

Tools that ship churn prediction

Feature names and requirements below were read on each vendor's own site or documentation in October 2026. For broader retention stacks, see our roundups of customer retention tools and predictive marketing platforms.

Braze Predictive Churn

Scores each user in a prediction audience from 0 to 100 on how likely they are to churn under your definition, with risk scores updated at a frequency you choose, and lets you target messages by risk. Limit: a purchased feature; the preview available before purchase builds one model at a time but cannot target messages by churn risk and does not update regularly.

Klaviyo churn risk prediction

Shows churn risk as a percentage on each qualifying profile, based on the number and frequency of orders: it drops when a customer orders and rises as time passes between orders. Limit: order-based only, and Klaviyo's profile predictions need real order history (its CLV predictions require at least 500 customers with orders and 180 days of order history).

Google Analytics 4 churn probability

Predicts whether a user active in the last 7 days will be inactive in the next 7, and offers suggested audiences such as Likely 7-day churning users (churn probability above the 80th percentile), which are shared with linked Google Ads accounts for re-engagement. Limit: a 7-day window, and at least 1,000 returning users who churned and 1,000 who did not in the qualifying period.

Adobe Customer AI

Generates churn and conversion propensity scores for individual profiles in Adobe Experience Platform, with the influential factors behind each score, against churn goals you define. Limit: it runs on experience event data in Adobe Experience Platform, and Adobe says it cannot determine whether an offer will make a customer more likely to purchase.

Amplitude Predictions

Scores users on the likelihood of a future action over 7, 30, 60 or 90 days, which you can frame as a retention or churn outcome, and syncs predictive cohorts to tools such as Braze and Iterable. Limit: Custom Predictive Cohorts are on Growth and Enterprise plans, and Amplitude says B2B companies are less likely to benefit.

ChurnZero ChurnScores

ChurnZero's customer health scores, called ChurnScores, rate an account's likelihood to renew or churn from the metrics you choose, such as product usage, support history and survey results, with alerts when scores change. Limit: built for B2B customer success teams rather than high-volume consumer messaging.

Pecan AI

Lists churn prediction and customer winback among its predictive use cases and builds models for business teams on data from warehouses and CRMs. Limit: it delivers the scores; the outreach runs in your other tools.

Mistakes that waste a good model

Discounting everyone with a high score is the classic one. Some of them would have stayed without the discount, and others leave for reasons a coupon will not fix, such as a missing feature or a champion who changed jobs. Look at the influential factors behind the score where your tool shows them, and match the action to the reason.

The other mistakes are quieter. Scoring customers who have already churned wastes budget. Judging the program on email opens instead of churn hides whether it works. And never revisiting the churn definition means the model keeps predicting last year's problem. Track the result in the same customer retention KPIs you report to the business.

Frequently Asked Questions

What is customer churn prediction?
Customer churn prediction estimates how likely each customer is to stop buying, using or paying within a set time window, based on patterns in past customers who did and did not churn. The result is usually a score or probability per customer that retention campaigns and customer success teams can act on.

What data is used to predict churn?
Behavioral data does most of the work: how often a customer logs in, opens the app or orders, which features they use, how long since their last purchase, and changes in those patterns. Support tickets, failed payments, survey scores and plan or contract details add context.

Which tools have churn prediction built in?
As of October 2026, Braze sells Predictive Churn, Klaviyo shows a churn risk prediction on qualifying profiles, GA4 offers churn probability and a Likely 7-day churning users audience, Adobe Experience Platform has Customer AI, Amplitude has Predictions, and ChurnZero offers ChurnScores for B2B customer success. Pecan AI builds custom churn models.

What should you do with a churn score?
Split customers into score bands and decide an action for each before you look at individual names: a product nudge for mild risk, a direct retention message or CSM call for high risk, and no discount for customers who were never at risk. Keep a holdout group to measure whether the actions reduced churn.

Is churn prediction the same as a customer health score?
They overlap. A health score combines metrics you choose, such as usage, support history and survey results, into one number per account. A churn prediction model learns from past churned and retained customers which signals matter. Many B2B teams use both.