Understanding your metrics.

Customer health score

0–100 composite of how a customer is trending

A single 0–100 number that rolls up recency, frequency, spend, and momentum into one read on how a customer is trending. Higher is healthier.

It is a composite of the signals below — not a separate measurement. Use it to sort a list quickly; open the underlying signals before you act on any one customer.

Return probability

Modeled chance the customer orders again in the window

The modeled likelihood that a customer places another order within their expected window, scored on your store's own purchase history.

Read it as a ranking instrument, not a promise. A customer at 0.72 is more likely to come back than one at 0.31; neither number is a guarantee about any single person.

Natural-return floor (~2.7%)

Baseline return rate with no intervention

In our own backtests, a meaningful share of lapsed customers come back on their own — ~2.7% in a given window — with no message sent at all.

This floor is why we hold out a control group where an audience is large enough to carry one. Where there is a control, the campaign is read against that control's own observed return rate; the fixed floor is the baseline only for uncontrolled, observational results. Either way, a campaign only counts as working if it beats the return rate of customers who got nothing. See recovered revenue below — the figure itself is not published yet, while the comparison is rebuilt to count only the customers a message actually reached.

Predicted next-order window

The date range a customer is expected to reorder in

The date range, not a single day, in which a customer is expected to order again — derived from their own cadence and the product's replenishment pattern.

It widens for irregular buyers and tightens for consistent ones. When a customer passes the far edge of their window without ordering, they move toward overdue.

Cadence

The customer's typical gap between orders

The typical interval between a customer's orders — for example, every 34 days. It is the backbone of the next-order window and of overdue detection.

Cadence source

Where a customer's cadence estimate comes from

Whether a customer's cadence is drawn from their own repeat history, borrowed from the product's typical replenishment interval, or from a blend of both.

We label the source so you know how much personal history sits behind an estimate. A customer on their second order leans on the product's cadence; a customer on their tenth leans on their own.

Replenishment

How fast a product is typically re-purchased

The typical interval at which a consumable product is bought again across all customers — a 60-day coffee bag, a 90-day supplement. It anchors cadence for customers who do not yet have enough personal history.

Second-order window

The critical period after a first purchase

The stretch after a customer's first purchase when the second order is most likely — and most fragile. Converting a first-time buyer into a repeat buyer is the single largest lever on retention.

In our backtests, customers flagged inside this window returned at roughly 3.2× the rate of unflagged customers in the same window. Read that as a lift in the flag's precision — where to spend your attention — not a promise that a message triples anyone's likelihood of returning. It is why Retrics watches this window closely.

Value tier

Where a customer ranks on lifetime spend

A ranking of customers by their spend with you — top, middle, and emerging tiers. It answers “how much is this relationship worth so far?” and is measured, not modeled.

Lifecycle stage

Where a customer sits in their journey

Where a customer sits in their journey with you: new, active, at risk, lapsed, or recovered. Stage is driven by recency against the customer's own expected window, so an “active” cadence-buyer and a dormant one-timer are treated differently.

RFM

Recency, Frequency, Monetary scoring

The classic segmentation trio: how recently a customer bought (Recency), how often they buy (Frequency), and how much they spend (Monetary). Each customer gets a score on all three.

RFM is transparent and battle-tested. Retrics uses it as a readable foundation and layers the predictive models on top, rather than replacing it with a black box.

LTV forecast (modeled)

Projected lifetime value — a forecast, not a fact

A projection of how much a customer is likely to spend with you over their lifetime, modeled from their behavior so far and the patterns of similar customers.

It is explicitly a forecast. We label it modeled so it is never confused with realized spend. Treat it as a planning input, not a booked number.

Cohort retention

How each start-month group retains over time

Customers grouped by the month of their first order, then tracked to see what share keeps buying in each following month. It shows whether the customers you are acquiring now retain better or worse than those from a year ago.

Repeat rate

Share of customers who order more than once

The percentage of customers who place a second order — the most direct measure of whether your product earns a habit. Retrics tracks it overall and by cohort so improvements are attributable.

Revenue at risk

Expected value of relationships trending away

The expected value of customers who are drifting toward lapse — their forecast value weighted by how likely they are to leave. It sizes the opportunity in front of you so you can prioritize.

It is a modeled figure, useful for ranking where to spend attention. It is not money you have lost, and not money you are guaranteed to keep.

Recovered revenue — observational

Revenue from customers who returned after a message

Revenue from customers who came back after a campaign Retrics prepared for them. It is easy to compute but it credits the campaign with every return — including the ~2.7% who would have come back anyway.

Retrics labels it observational — and today withholds the amount alongside the causal one: until a send carries recipient-level evidence, no row in the ledger prints a dollar. What you see is the return count, the assignment, and the control kept back — never a figure that could be mistaken for proof the message caused the return.

Recovered revenue — causal

Return lift measured against a held-out control

The revenue that a campaign returned above and beyond a held-out control group who received nothing. This is the honest number: it takes the holdout's own return rate as the baseline — not the store-wide floor — and credits only the difference.

Causal measurement requires a holdout at send time. The figure itself is not published yet. The recovered-revenue figure is withheld today. The holdout is still carved and never messaged on every eligible send, so the evidence accumulates — but the comparison is being rebuilt to count only the customers a message actually reached, and Retrics won’t publish a dollar measured over anything looser.

Concretely: the treated side is currently derived from the frozen audience rather than from who a message actually reached — so it counts anyone suppressed, unreachable or never sent to, which biases the treated return rate downward by an unknown amount, and every dollar derived from it with it. Retrics reports the ledger, the assignment counts and the withheld control, and holds the amount back until the comparison is rebuilt from recipient-level send evidence.

Still unsure how a number is derived? Write to support@retrics.ai and we will walk you through the methodology behind it.