Intelligence Models

Churn Detection model

How the churn score is calculated, what the risk bands mean, and how to act on them.

What is the churn score?

The churn score is a number from 0 to 100 that represents the probability that a customer will not make another purchase. A score of 91 means the model estimates a 91% chance this customer will churn if no action is taken.

How is it calculated?

The churn model is a gradient-boosted ensemble trained on your store's historical data. The primary signals, in order of typical importance:

  1. Purchase recency — How many days since the last order, relative to that customer's personal average
  2. Purchase frequency change — Whether the gap between recent orders is widening
  3. Average order value trend — Whether AOV is declining
  4. Email engagement — Open and click rates on Klaviyo campaigns (if Klaviyo is connected)
  5. Cohort comparison — How this customer's behaviour compares to similar customers at the same stage
  6. Product affinity — Whether the customer's favourite product category has become unavailable or went out of stock

Risk bands

BandScore rangeMeaningRecommended action
Safe0–30Customer is active and healthyNo action needed — focus on growth
Watch31–60Early warning signsLight nurture — include in repurchase sequences
At-Risk61–80Engagement is decliningRetention offer — discount or personal outreach within 48h
Critical81–100High probability of permanent churnPersonal outreach today — win-back before window closes

Score updates

Churn scores update every 2 hours. A score that jumps significantly in a short period (e.g., +18 points in 7 days) is a strong signal that something has changed in the customer's behaviour — this is flagged on the customer's profile.

Viewing scores

  • Protect — Shows all at-risk customers sorted by score, filterable by band
  • Explore → Churn Detection — Full score distribution, trend charts, band breakdown
  • Customer Profile — Individual score with driving signals breakdown

Score overrides

If you have context the model doesn't (e.g., a customer told you they're taking a seasonal break), you can manually override their score from the Customer Profile page. Overrides are logged and do not affect model training. The model resumes control at the next scoring cycle unless you lock the override.

💡
Best practice: Focus your retention budget on the At-Risk band (61–80), not just Critical. At-Risk customers are still saveable with a lighter touch — Critical customers often need more expensive win-back campaigns.
Was this article helpful?