For replenishment brands on Shopify
Every night, machine learning decides for each customer whether an email would lead to an order, what to offer and when. Everyone else is left alone, and a group that is never emailed shows what the emails really added.
What you get every night
Every night it reads your whole order history and retrains three machine learning models: who will reorder within 60 days, who will start a subscription within 90 days, and who will buy a product new to them within 90 days. Each is graded every night on customers it has never seen, against the rule it replaced, and the report records it when a model loses.
Then it decides, for every customer, which one message would help right now, hands that decision to your Klaviyo account, and paces the sends across the day. On a recent night the reorder model scored a ROC-AUC of 0.7835 on 302,671 customers it had never seen.
01
Exactly one campaign, with a calibrated probability that means what it says: a 0.30 is a 30 percent chance within about 90 days, comparable across campaigns. Nothing in Unlapse reads 99 percent.
02
One to three per customer, from a ranker over their whole purchase history, never a free sample, never a product they were only ever given.
03
Built only from the signals that made the decision, so your team can check any send. For the demo customer: “Active buyer, 6 orders about every 31 days, 34 days since the last one; model ranks them about 12 times likelier to reorder than average.”
04
Anyone who just bought, anyone emailed in the last 21 days, and anyone Klaviyo cannot deliver to. On the first brand that is about three quarters of the people it could contact, every morning.
05
An even slice every 20 minutes, never a burst.
Four jobs
of the reorder model's top thousand picks bought within 60 days, on a recent night.
Against a base rate of 0.84%.
See who should buy again and whether they are in rhythm, entering their window, or overdue. The model is retrained every night and graded on customers it has never seen. Stop guessing from a days-since-order column.
of recently lapsed customers actually came back, on a leak-free holdout.
Long lapsed 2.10%, dormant 0.81%.
Catch buyers who skipped a cycle before they go cold. The three tiers are treated differently because they behave differently. A customer who paused their subscription is never sent a win-back.
of the cross-sell model's top thousand bought a product new to them within 90 days.
Against 20.9% for the co-purchase rule it was built to replace. Loses at the top
Recommend what buyers of Product A usually add next, from the co-purchase patterns in your own store, with the products this customer already has taken out. The model currently loses to the rule at the top of the list, and the engine reports that every night.
of the incumbent retention tool's “upgrade to subscription” list were already active subscribers. It could not see the subscription platform.
Only about 23% were genuine candidates.
Spot one-time buyers who behave like subscribers. The engine reads your subscriptions in Loop, so it knows who is already subscribed, who paused, and who lapsed.
One night, in four loops
01
Five models retrained, 568,836 customers scored, one decision each.
02
Three products, one link, one reason. Your flows do the rest.
03
Flat within five percent, untuned.
04
That is how it knows what the rest were worth.
Demo data: an invented coffee brand, rendered through the engine's real payload shape. The charts use the first brand's anonymous aggregates.
The difference
| A reorder reminder | This engine | |
|---|---|---|
| When | A reminder fires when a customer passes their average gap between orders. | Models retrained every night predict who will order soon, for every customer. |
| What | The same product again. | The product they are due for, one they have not bought that people like them add, or a subscription. |
| The cart | A one-click cart for the product in the reminder. | Up to three products chosen for that customer, all in one cart link. |
| Who | Everyone past the interval. | One decision per customer per night, and most customers get nothing: recent buyers, people emailed in the last 21 days and people who cannot be mailed are left alone. |
| Proof | The dashboard counts every order after a reminder as its result. | 5 in 100 customers never get the engine's emails, so the difference they make is measured, not assumed. |
| Honesty | Each model is graded every night against the rule it replaced, and the report says so when a model loses. | |
Five percent of your customers are never emailed by Unlapse. Every report compares the emailed group with them, with a confidence interval, on your own orders. When the emails work, you see it. When a campaign is only taking credit for purchases that were coming anyway, you see that too, before you pay for another year of it.
How the measurement worksProof
We read your order history first and tell you what the engine would do with it. Then you decide.