Predictive Segments: When a Model Score Beats Your RFM Buckets

Your RFM buckets say a customer is a champion. The model says they will not buy again for a year. The model is right more often than the bucket, and the reason is that the bucket looks backwards while the model looks forwards.

A predictive segment is built from a model score, such as predicted customer lifetime value or churn risk, rather than from past behavior alone. Klaviyo's predictive analytics scores each customer on the likelihood of their next order and their expected value over the next year. The model needs data to work: at least 500 customers with orders, 180 days of order history, and orders in the last 30 days. The score updates as the model retrains, at least weekly.

RFM looks backwards. The model looks forwards.

RFM segments, recency, frequency, monetary value, sort customers by what they have already done. A customer who bought three times last year is a champion. A customer who bought once and never returned is at risk. The buckets are honest about the past and silent about the future. A champion who has stopped buying is still a champion in the RFM grid until enough time passes to demote them, and by then the window to win them back has closed.

A predictive segment sorts by what the customer is likely to do next. Klaviyo's predicted customer lifetime value estimates what a customer will spend over the next year, and its churn risk score estimates the likelihood they will not order again. The model retrains at least weekly, so the score moves as behavior changes. A champion whose predicted value is falling gets flagged while there is still time to act, not after the demotion.

Related reading on this site: SegmentMuse and Cohort Segments: What the Month a Buyer Arrived Predicts cover the neighbouring problems in more depth.

The thresholds Klaviyo actually uses

Klaviyo's churn risk is a three-band score. Low risk is 33 percent or below, medium is 33 to 66 percent, and high is above 66 percent. The bands are not guesses. They come from the model, and they are the same bands the platform uses for its own retention recommendations. A customer at 70 percent churn risk is not at risk in the abstract. They are more likely than not to stop buying.

The model has minimums, and they are worth knowing before you build on it. Klaviyo requires at least 500 customers with orders, 180 days of order history, and orders in the last 30 days before it will generate predicted values. A store below those thresholds gets no scores, and the honest move is to use RFM until the order history catches up. A store above them gets scores that update weekly, and the honest move is to use both.

Combine them: the segments RFM cannot see

The strongest use of predictive scores is the combination, not the replacement. Pair the model with RFM and the segments RFM cannot see appear. At-Risk Champions: high RFM, high churn risk, the customers worth saving. Sleeping VIPs: high lifetime value, no recent order, the customers worth waking. A customer who is a champion by RFM and high risk by the model is a different campaign from a customer who is low value by both.

The campaigns write themselves once the segments exist. At-Risk Champions get the win-back with the best offer, because their predicted value justifies it. Sleeping VIPs get the early access or the exclusive product, because they respond to status, not discounts. The model does not replace judgment. It tells you which customers judgment is worth spending on.

Frequently asked questions

What is a predictive segment?

A segment built from a model score, such as predicted customer lifetime value or churn risk, rather than from past behavior alone. Klaviyo's predictive analytics scores each customer on the likelihood of their next order and their expected value over the next year.

How does Klaviyo score churn risk?

In three bands: low risk is 33 percent or below, medium is 33 to 66 percent, and high is above 66 percent. The model retrains at least weekly, so the score moves as behavior changes.

What does Klaviyo need before it generates predicted values?

At least 500 customers with orders, 180 days of order history, and orders in the last 30 days. A store below those thresholds gets no scores and should use RFM until the order history catches up.

Should predictive segments replace RFM?

No. Combine them. Pairing the model with RFM reveals segments RFM cannot see, such as At-Risk Champions with high RFM and high churn risk, and Sleeping VIPs with high lifetime value and no recent order.