AI customer segmentation for e-commerce

Predictive Segments: What 'Likely to Purchase' Actually Predicts

Likely to purchase is a prediction about a group, not a promise about a person. Here is what the model is actually saying.

Klaviyo's predictive segments score every profile on four things: predicted lifetime value, probability of purchase in the next 30 days, expected date of the next order, and churn risk. Likely to purchase means the model thinks the profile will buy in the next 30 days, based on how similar profiles behaved. It is a prediction about a group, not a promise about a person, and it works best when the store has enough history to train it.

The four scores, and what each one says

Klaviyo builds predictive segments from four scores. Predicted lifetime value estimates what a profile is worth over the whole relationship. Probability of purchase estimates whether they will buy in the next 30 days. Expected date of next order estimates when. Churn risk estimates whether they will stop buying. Together, the four scores describe a buyer in a way a single segment never could.

Likely to purchase is the one most stores use first, and it is also the one most stores misunderstand. It does not mean the profile will buy. It means the model, looking at how similar profiles behaved, thinks this profile is more likely than average to buy in the next 30 days. The prediction is about the group the profile belongs to, not about the individual, and the distinction matters when you decide what to do with it.

What the model is actually looking at

The model trains on the store's own history: who bought, when they bought, what they bought, and how often. A profile that looks like past buyers, same first order, same time between orders, same product category, gets a high score. A profile that looks like past one-time buyers gets a low one. The model is pattern matching, not magic, and it is only as good as the history it was trained on.

That is why a new store gets weak predictions. With a few hundred orders, the model has almost nothing to learn from, and the scores are guesses. With thousands of orders, the patterns are real, and the scores start to mean something. A store that turns on predictive segments on day one and acts on the results is acting on noise.

How to use a prediction without trusting it too much

The right way to use likely to purchase is as a filter, not a fact. Mail the likely-to-purchase segment a little more often, with a little more urgency, and watch whether they actually buy at a higher rate than the rest of the list. If they do, the model is working, and you can lean on it more. If they do not, the model is wrong for your store, and you should stop using it.

The test is the same as any other segment test: compare the predicted group against a control. A prediction that does not change the outcome is not a prediction. It is a label, and labels do not sell anything.

When the model gets it wrong

Predictive segments fail when the buying pattern changes. A store that adds a new product line, changes its pricing, or moves into a new market will confuse the model, because the history it was trained on describes a different store. The scores keep coming, and they look just as confident, but they are predicting a past that no longer exists.

The fix is to retrain, and to check the predictions against reality on a schedule. A quarterly review, comparing the predicted buyers against the actual buyers, tells you whether the model is still working. A model that is never checked is a model that is never wrong, which is the same as a model that is never right.

Keep reading

Frequently asked questions

What does likely to purchase actually mean?

A profile the model expects to buy in the next 30 days, based on how similar profiles behaved. It is a probability, not a certainty, and it works best on stores with enough order history to train the model.

What are Klaviyo's four predictive scores?

Four scores: predicted lifetime value, probability of purchase in the next 30 days, expected date of the next order, and churn risk. Together they describe who will buy, when, and how much they are worth.

Are predictive segments available on free plans?

No. Predictive analytics is a paid-plan feature in Klaviyo, and the model needs order history to train on. A new store with few orders will not get useful predictions.

When do predictive segments fail?

When the model is wrong, usually because the store's history is thin or the buying pattern changed. A prediction is only as good as the data behind it, and a store that changed its product mix will confuse the model.