Churn in eCommerce is quieter than churn in subscriptions. Nobody cancels — they just stop coming back, and by the time the dip shows up in your revenue chart, the customers are long gone and far harder to win. Predicting which customers are drifting away, while there’s still time to act, is one of the highest-return uses of your data. But a prediction is only half the job; the intervention is where most programmes fall down. This article covers both: how to model lapse risk in a non-contractual business, and how to act on it without burning margin.
Churn looks different without a subscription
In a contractual model — SaaS, telecoms — churn is an event you can timestamp: the cancellation. In retail there’s no cancellation. A customer who hasn’t bought in 90 days might be churned, might be between purchase cycles, or might be about to reorder. So the first task is to define churn for your business rather than borrow someone else’s.
The standard approach is to derive a lapse threshold from your own data. Look at the distribution of gaps between purchases; the point where the probability of a customer ever returning drops sharply is your working definition. A consumables brand might call churn at 60 days; a furniture retailer might need 18 months before “lapsed” means anything. Picking the wrong window invalidates everything downstream, so spend real time here.
What predicts lapse
The features that carry churn signal overlap heavily with segmentation and CLV work, which is why these capabilities are best built together:
- Recency relative to the customer’s own cadence — the single strongest signal. A 40-day gap is fine for a quarterly buyer and alarming for a weekly one.
- Engagement decay — falling email opens, fewer site visits, longer gaps between sessions.
- Order trajectory — shrinking basket size or downgrading to cheaper items.
- Negative experiences — a return, a delivery problem, a support contact that didn’t resolve cleanly.
- Discount dependency — customers acquired or sustained purely on promotions lapse faster when the offers stop.
Engagement and experience signals usually live outside your commerce platform, so a churn model is only as good as your joined data. If web, email and support events aren’t connected to the customer record, that’s the first build — see building a unified customer data model.
Modelling approaches
Start with a rules baseline
You can flag at-risk customers today with a rule: recency exceeds 1.5x their typical purchase interval and email engagement has dropped. It’s crude, explainable, and gives you something to intervene on while a model is built. Always ship this first — it also becomes the benchmark a model must beat.
Classification models
The common framing is binary classification: predict whether a customer will lapse within the next N days. Logistic regression gives you an interpretable, well-calibrated baseline; gradient-boosted trees usually lift accuracy when you have richer features. The output you want is a probability, not a yes/no, so you can prioritise by risk and combine it with value.
Survival models
If when a customer will lapse matters as much as whether, survival analysis (such as a Cox model) predicts time-to-event and handles customers who haven’t lapsed yet more gracefully. It’s a better fit when timing the intervention is the whole game.
Calibrate and validate
Validate on a held-out time window, the same way you would for predicting customer lifetime value. Check that predicted probabilities match observed lapse rates — a model that ranks well but is miscalibrated will misallocate your retention budget. And watch the false-positive cost: chasing customers who were never going to leave wastes money and can annoy them.
Risk alone isn’t enough — combine it with value
The most common mistake is treating every at-risk customer equally. Intervening costs money and attention, so you should spend where the payback is highest. Cross churn risk with predicted value:
- High value, high risk — your priority. Personal, generous intervention is justified.
- High value, low risk — protect, don’t pester. Light loyalty touches only.
- Low value, high risk — automated, low-cost nudges at most; some are fine to let go.
- Low value, low risk — leave alone.
This prioritisation is the difference between a retention programme that pays back and one that sprays discounts at everyone and erodes margin.
Designing the intervention
A prediction that doesn’t change what a customer experiences is worthless. The intervention should match why the customer is at risk — and that’s where most programmes are lazy, defaulting to a blanket discount.
- Diagnose the cause. A customer lapsing after a delivery problem needs a service gesture and reassurance, not 15% off. One drifting from disengagement might respond to a reminder of what they liked or a replenishment nudge.
- Try non-discount levers first. Replenishment reminders, new-in alerts for a favourite category, early access, a helpful check-in. These protect margin and often work better than money off.
- Reserve discounts for where they’re needed. Use them as a last resort for genuinely price-sensitive, high-value customers — not as the default for everyone.
- Fix the root cause. If returns or delivery issues are driving churn, the durable fix is operational, not a winback email. Resolving service friction quickly — see our work on helpdesk automation — prevents the lapse in the first place.
Measure incrementally, not by open rates
Judge a churn programme by incremental retention, proven with a control group. Hold back a random portion of the at-risk population, treat the rest, and compare retained revenue between the two. Vanity metrics — emails sent, coupons redeemed — tell you nothing about whether you changed anyone’s behaviour. Redemptions in particular can be pure cannibalisation: discounts handed to people who would have returned anyway.
Common pitfalls to avoid:
- Wrong churn window that mislabels healthy customers as lapsed.
- Acting on risk without value, spreading budget too thin.
- One-size-fits-all winbacks that ignore the cause of churn.
- Training customers to lapse by always rewarding them with a discount for going quiet.
- No holdout, so you can’t prove the programme works.
Conclusion
Effective churn prevention is a loop: define lapse correctly for your business, predict risk as a probability, prioritise by value, intervene on the actual cause, and measure incremental retention against a control. Done well, it’s quietly one of the most profitable things a retailer can build, because keeping a good customer almost always costs less than finding a new one. If you’d like help building this as part of a broader data insights capability, get in touch.