Here's the version of CLV you'll find in most analytics tools. Take the average order value across your past customers. Multiply it by an average repeat rate. Multiply that by an average gross margin. Multiply that by the lifespan you assume your relationship lasts. You get a single number, applied to every customer the same way, computed from data you only have for people who already bought from you.
For the merchant, the resulting number is fine for board slides. For the visitor in front of the store right now, evaluating their first purchase, it's structurally useless. They don't appear in the data the average was built from.
This is the gap we wanted to close.
What we actually want to know
The merchant's real question, at the moment a visitor lands on the storefront, is: how much should I invest in keeping this person here. A bargain hunter who'll convert on a discount and never return is worth less than a researcher who'll convert at full price and come back four times. The dollar amount in each case is different. The right intervention is different. The right ad attribution is different.
You can't separate those visitors using historical AOV. You can only separate them by reading the behaviour that's actually happening in the session.
So that's where we started.
The gap we wanted to close A bargain hunter who'll convert on a discount and never return is worth less than a researcher who'll convert at full price and come back four times. Historical AOV can't tell you the difference.
The architecture, in plain terms
The pipeline runs in four stages. We'll describe each at a level that's useful without revealing implementation details we'd rather not advertise to anyone who wants to game them.
As a visitor moves through the storefront, the tracker collects observations of their interaction. Time spent on specific elements. Patterns of scroll. Whether they consult specs or reviews or comparison tools. Whether they revisit the price. Signals are normalised to remove device-class and locale variation. Raw observations stay on the visitor's device. What leaves the device is a vector of normalised features, not a recording of their session.
The feature vector is classified against five behavioural archetypes from the behavioural economics literature, not from k-means clustering of past customers. Researcher. Comparison shopper. Hesitant buyer. Bargain hunter. Loyalist. Each has documented purchasing patterns: how long they evaluate, what they respond to, when they return. The classifier outputs probabilities across all five, because most real shoppers are a mix.
Given an archetype mix and the merchant's catalogue, we estimate when the visitor is likely to return and what category of purchase they're likely to make. Survival models are the right tool here because the relevant question isn't "will they return" but "when, with what probability." Both of those drive different operational decisions.
A point estimate of CLV is misleading. A range with explicit uncertainty is useful. The Monte Carlo step samples from the upstream distributions and produces a confidence interval, not a number. The merchant sees both the centre estimate and how confident the model is in it.
The forecast horizons are 30 days, 1 year, and 3 years. We don't go further. Past three years, the model's uncertainty exceeds its signal.
What the model knows versus what it predicts
This distinction matters more than the architecture does.
The model knows what behavioural signals the visitor has produced in the current session, plus any prior sessions for visitors who've returned. It knows the merchant's archetype-conditional patterns at the catalogue level. It does not know things it can't measure: the visitor's emotional state, their household budget, whether they're shopping for themselves.
The predictions widen their confidence intervals when the underlying data is thin. A first-time visitor with twelve seconds on a single page gets a CLV estimate with a wide band. A returning visitor with three completed purchases gets a narrow one. The downstream recommendations weight by that uncertainty rather than pretending it's not there.
The honest framing If we don't know, we say we don't know. If we know with low confidence, the action we recommend reflects that. We don't believe in models that pretend to know things they can't.
Why per-merchant, from day one
We made one architectural choice that's worth defending: each merchant gets a model trained only on their own store's data. There is no pooled cross-merchant training set. There is no shared centroid we move toward.
The reasoning is structural, not philosophical. A supplement brand's bargain hunter behaves differently from a luxury watch brand's bargain hunter. Pooling them produces a model that's bad at both. The signal that matters is the merchant's own data, and the only way to use it cleanly is to keep it isolated.
The trade-off is that new merchants take a few weeks of traffic to reach calibrated confidence intervals. We accept that trade-off because the alternative, importing a centroid from someone else's store, is worse for the merchants we care most about. A specialist store gets a specialist model.
What the merchant actually does with this
Three operational decisions improve with session-one CLV.
Ad spend prioritisation. When a visitor lands from a paid channel, the cost of acquisition is real. If session-one CLV is high, the merchant can extend the customer-acquisition window. If it's low, the merchant can cap spend on the cohort that channel produces. Most retargeting decisions in 2026 are still made by people relying on conversion rate, which is the wrong metric. Predicted CLV is the right one.
Lookalike audience construction. Knowing which behavioural archetypes convert at the highest predicted CLV lets the merchant build lookalike audiences from the right seed. A merchant whose loyalists buy four times a year produces a different lookalike than a merchant whose bargain hunters buy once. The seed selection is the leverage point. Predicted CLV identifies the seed.
Cart recovery decisions. Not every abandoned cart deserves a recovery flow. A high-CLV cart abandonment is worth a personalised follow-up. A low-CLV cart abandonment is worth either nothing or a single, generic prompt. Spending recovery effort uniformly is wasteful. Predicted CLV makes the spending decision rational.
The point of forecasting isn't to be right about every shopper. It's to be useful at the moment a decision has to be made, which is almost always before the data is in.