Predicted run-out: how reorder timing actually works
“Predicted run-out” sounds like magic and is mostly arithmetic. If a customer bought the same coffee on March 1st and again on April 4th, they have told you something precise: that product lasts them about 34 days. The prediction is that simple — and the craft is in everything around it: what to do with one purchase, with ten, with a double order, and with a customer who says “you’re early.”
The core signal: their own interval
The best predictor of when someone runs out is not an industry average or a product setting — it’s the median gap between that customer’s own repeat purchases of that product. The median matters: one weird gap (a holiday, a gift order) shouldn’t bend the whole rhythm the way an average would. Per customer, per product, the model keeps updating as orders arrive.
The fallback ladder
- Two or more purchases → the customer’s personal median interval. The strongest signal there is.
- One purchase → the merchant’s default interval for that product, if one is set. A 90-count vitamin bottle has an obvious candidate.
- No signal at all → silence. This rung is the important one: a durable product with no rhythm should never generate a reminder. Guessing trains customers to ignore you.
What adjusts the estimate
- Quantity — two bags last about twice as long; the prediction stretches proportionally.
- “Still have plenty” — an honest button on the reminder that pushes the date out and biases future predictions longer, within bounds.
- An early reorder — decays that bias back. The customer’s actions always outrank the model’s memory.
- Snooze — “ask me in two weeks” defers the reminder without touching the underlying rhythm. Later is not never.
From run-out date to reminder date
The reminder must land before the run-out by enough margin for two things: the customer taking a day or two to act, and the parcel shipping. Try it with your own numbers:
Reorder timing calculator
That’s 6 days before the predicted run-out: enough for the customer to act and the parcel to arrive before the jar is empty — the window where a reminder feels like service, not marketing. This is the arithmetic Special Delivery runs per customer and per product, continuously.
This is why run-out prediction beats every “30 days since last purchase” email automation: the generic version is wrong for the daily double-espresso household and for the one-cup-on-Sundays household, in opposite directions. Personal rhythm is the difference between a reminder that feels like service and one that feels like spam.
Where the reminders go
Prediction sets the clock; the channel ladder decides the messenger — the QR on the box always, email at the predicted date, a printed postcard for high-value customers who let the email pass. And every response feeds the next cycle’s timing, which is the part that compounds: the program gets quieter and more accurate the longer it runs.
Common questions
›How many purchases does a prediction need?
Two purchases of the same product establish a personal interval — the gap between them. Before that, a sensible default set by the merchant (say, 30 days for a 30-serving supplement) can stand in. One purchase with no default should produce silence, not a guess.
›What happens when the prediction is wrong?
The customer tells you, and the model listens. A "still have plenty" tap stretches the next prediction; an early reorder pulls it back in. Wrong-but-correctable beats silent-and-stubborn.
›Does buying a bigger size change the prediction?
It should. Two bags last roughly twice as long as one, so quantity scales the predicted interval. A model that ignores quantity nags the person who stocked up — the exact customer you least want to annoy.