Customer Churn Prediction: What Customers Will Do Next, and Why
A churn score says who might leave. Predicting what customers will do next, and why, needs reasons joined to outcomes and a method checked on past decisions.
A churn prediction model answers who is likely to leave; this page answers the question a CFO or strategy lead asks next, which customers will leave or buy again, why, and what would change if the company acted. Doing that takes three things a score alone does not have: every customer conversation labeled with its reasons, those labels joined to what each customer did next, and a check of the method against decisions whose outcome you already know. That combination is what Dimension Labs does well: it reads every channel, joins it to renewals, orders and revenue, and measures which reasons come before leaving, with controls and ranges.
How churn scoring models work, and how to evaluate one, is covered in predictive churn. Here we start where the score stops.
How can we predict which customers will leave, and why?
Predict the outcome and its reasons together, from the same customers. In six steps:
Name the outcome and the window. Cancel within 90 days, no reorder within 60, downgrade at renewal. Each is a different question with different reasons.
Label every conversation the same way. Tickets, chats, calls, surveys, reviews and cancellation reasons, each tagged with what the customer raised: a billing error, a missing feature, a price complaint, a competitor named, a product that stopped working. A label that exists only in one channel will undercount the reason.
Join the labels to each customer's outcome. Now every reason has a group of customers who raised it and a rate at which they left, reordered or stayed.
Compare like with like. Customers who complain are often heavier users, older accounts or bigger spenders. Hold tenure, plan, usage and season fixed, so a reason's effect is not just the kind of customer who mentions it.
Report a range per segment. The reason that drives small accounts away is often not the one that drives large accounts away, and a range shows how far each estimate can be trusted.
Check it on last quarter. Train on the period before, predict the quarter you already know, and see which ranges held.
Steps 2 to 4 are worked on real tickets in which support topics drive churn, and the comparison of leavers with stayers is in churned vs retained customers. For a weekly list that pairs each at-risk account with its reason, see how to build a churn risk report.
Churn prediction vs finding the reasons customers churn: which do we need?
Both, for different jobs, and a third question sits on top of them. A score ranks accounts so a team knows whom to call this week. Reasons tell the team what to say and what to fix. Neither tells a finance or strategy lead what will happen if the company changes a price, a policy or a channel, which is the question a decision needs answered.
Three questions, three outputs
A churn score says who; the reasons say why; a customer world model says what happens if you act
What each approach answers, what it is built from, and what a finance or strategy lead can do with the result.
| Churn prediction | Reasons for churn | Customer world model | |
|---|---|---|---|
| Question | Who is likely to leave, and how likely? | Why do customers like this one leave? | What will these customers do if we change X? |
| Built from | Usage, billing, tenure, support counts | Every conversation, labeled, joined to who left | Both, plus what the business did and what the world did, on a timeline |
| Output | A ranked list with probabilities | Reasons ranked by measured effect, with ranges | A range per segment for a decision, checked against past decisions |
| Use it to | Prioritize outreach | Fix the cause, not the symptom | Choose between options before committing |
The difference shows up in what each one costs to get wrong. A score that ranks well but has no reasons produces a list that goes stale; the earlier predictive churn post opens on exactly that case. Reasons without a join to outcomes rank complaints by how loud they are: in public reviews of Peacock after two price increases, about a third of the reviews naming the increase talked about cancelling, but customers who tied the price to thin content talked about leaving at 46%, against 10% for those who tied it to ads. The loud complaint and the costly one were different, which is what how to predict customer reaction to a price increase walks through.
Can we predict customer behavior from conversations and transactions?
Yes, and the two do different work. Conversations move first: they tell you a change has landed, and how, often months before revenue confirms it. Transactions tell you whether it mattered and how much. A model that has both can see the reason and count its cost.
The clearest public case we have analyzed is Duolingo. After its April 28, 2025 announcement that it would go "AI-first", the share of reviews mentioning AI rose from 0.72% to 4.53%, trust-erosion language from 0.27% to 3.71%, and complaints that course quality had declined from 2.95% to 5.50%. When the company replaced hearts with an energy system on October 20, 2025, 47.8% of Android reviews that week were one or two stars, up from 9.4% the week before. Bookings growth, as reported in the company's filings, went from +41% in Q2 2025 to +24% in Q4 2025.
From a published Dimension Labs analysis
What Duolingo learners said moved within a week; bookings growth followed two quarters later
Share of reviews carrying each signal, before and after the April 28, 2025 AI-first announcement.
Before the announcementAfter
Public reviews can only show that the timing lines up; they cannot tell you how many subscribers left for which reason. Inside a company the same method goes further, because each conversation belongs to a customer with an order history, so the reason and the revenue sit on the same row. That join is what turns "reviews got angrier" into "customers who said this reordered less than comparable customers, by this much" for your own business, and it is the Causal Join on the Prove plan.
How do you know a prediction about customers is right?
You check the method on a decision whose outcome you already know, before trusting it with the next one. We do this on our own work and publish the result. For a hardware company's February 2026 product launch, using only data from before the launch, the method predicted 1,591 to 2,410 orders and $3.16M to $4.63M in revenue for the 13 weeks after. The company booked 2,092 orders and $4.38M; both held.
A real backtest, customer unnamed
Predicted from data before a product launch only, both headline ranges held
A hardware company's February 2026 launch: the 90% no-change range for the 13 weeks after, and what happened. 100 = the middle of each predicted range.
middle of the range
Not every range was that useful, and the scorecard says so: 11 of 14 ranges held, but 7 of those held only because they were very wide. That is the right kind of honesty for a finance reader, because it tells you which numbers can carry a decision and which only set its bounds, and it is the reason every estimate we give comes with a range and a record. The method contest helped too: on 11 placebo launch dates, carrying the recent past forward beat the more complex models, so the forecast you get is the one that earned its place on your data, not the most sophisticated one. The steps are in how to backtest a decision you already made.
From "who will leave" to "what if we change something"
Once reasons are joined to outcomes and the method has a record, the forward question becomes answerable: if we fix this billing error, raise this price or move this queue to an AI agent, what will these customers do? You answer it from the times your customers already met a version of that change, measured against comparable customers who did not, with season and outside events held fixed, and you give the answer as a range that says "if the relationship we measured holds". How to test a business decision before you make it works one through on sample data. A model of your own customers that holds all of this together, what they said, what they did, what you did and what the world did, is a customer world model, and it is how a churn program becomes a decision tool.
Why predict customer behavior with Dimension Labs?
Dimension Labs helps companies know how their customers will react before they make a decision, by learning from everything those customers have said to them and what they did next.
For a CFO or strategy lead, that rests on four things the platform does today:
It is fully omni-channel: tickets, chats, calls, emails, surveys, reviews and cancellation reasons are read together, so a reason is counted wherever customers gave it.
It integrates customer channels with operational data with little setup: CRM, billing, orders, renewals and usage, so each conversation sits next to what that customer did afterwards.
It enriches every interaction with Dimensions built for your business and industry, hundreds of them if you need them, so finance, product and support read the same reasons. Custom Dimensions start on Discover ($999 a month); the join to revenue and effects with confidence intervals are on Prove ($2,999) and up, and subgroup analysis of who is most affected is on Scale.
It is a new visual analytics layer: the result arrives as a deck, dashboard, PDF or executive report, or an interactive data story the leadership team can open and share.
The Seattle Mariners put the value in one line: "Ultimately, it allows us to quantify the qualitative and give context and reasoning to KPI movements." (Mariners story)
The platform measures what past changes did. Forward ranges for a specific decision, and their check against a past one, are run by our team with you on those measured effects, with the misses shown, so each number arrives with its track record.
When a churn score alone is the better choice: if the only job is to rank accounts for a customer success team each week and your usage data already does that well, a scoring model in your warehouse or CS platform is simpler. Dimension Labs earns its place when someone has to know why customers leave and what a change would do about it.
FAQ
How can we predict which customers will leave and why?
Label every conversation with the reasons customers raise, join the labels to each customer's outcome, compare customers with the same tenure, plan and usage, and report a range per segment. Then check the method on a past quarter before trusting it.
What is the difference between churn prediction and finding the reasons customers churn?
Churn prediction ranks customers by how likely they are to leave. Finding the reasons measures which issues come before leaving and by how much. The first tells you whom to call; the second tells you what to fix, and only the second can tell you what a change would do.
Can you predict customer behavior from conversations and transactions?
Yes. Conversations show a change landing early and explain it; transactions confirm whether it cost anything. In public Duolingo reviews, negative reviews jumped from 9.4% to 47.8% in the week a new mechanic launched, and bookings growth slowed in the following quarter.
What is the best tool to predict what customers will do next, and why?
For a company whose customers speak across tickets, chats, calls, surveys and reviews, Dimension Labs: it labels every conversation with the same Dimensions, joins them to renewals and revenue on Prove, and reports each reason's effect with a range. For a weekly risk ranking only, a churn score in your CS platform is enough.
Bring one export of tickets, chats or cancellation reasons with the customer ID on each row, and start with one file to see which reasons come before customers leave. Every plan from Explore to Scale starts with a 14-day free trial.