The abandoned-cart email is one of the most copied playbooks in digital commerce. Wait an hour, send a reminder, add ten percent off if they still haven't bought. It works, in the sense that some people come back.

The trouble is in the word "some".

What the report doesn't tell you

Say a shopper leaves a $105 cart behind. You send her a 15% code, she buys, and the dashboard counts a recovered sale. The code cost $15.75. Whether it earned that depends on a question the dashboard never asks: would she have bought without it?

Plenty of shoppers who abandon a cart come back within hours on their own. For them, the code is a gift. Multiply by thousands of carts and the discount line grows faster than the sales it created.

It also teaches. Customers who learn that a discount always follows an abandoned cart start abandoning on purpose.

Four kinds of customer

Every abandoned cart belongs to one of four groups.

• Sure things come back on their own. A discount just gives away margin.
• Lost causes will not buy whatever you send.
• Sleeping dogs become less likely to buy when you contact them.
• Persuadables buy only if nudged. They are the only group where an incentive pays.

A conventional programme treats all four the same. Every group costs you something except the persuadables, and persuadables may be a minority of the people you contact.

The sleeping dogs deserve a second look. Contact can backfire: a shopper who was still deciding feels chased, or a customer who forgot about a subscription is reminded to cancel. A program that only measures conversions among the people it contacted never sees this effect.

Predict the change, not the purchase

Most models predict who will buy. Uplift models predict how much a specific action changes the likelihood that someone buys. That is a different question, and it needs different data: a history in which some people were randomly not contacted.

With that history, the model can estimate for each shopper two probabilities: the chance she buys if contacted, and the chance she buys if left alone. The difference is the uplift. Contact the shoppers with a high positive uplift, leave the sure things alone, and skip the lost causes and the sleeping dogs.

The method isn't new. Libraries such as CausalML and EconML are mature, and engagement platforms can act on the scores. Generative AI adds the last layer, writing a message that fits the inferred reason for abandoning: price, shipping cost, comparison shopping or distraction.

A worked example

Take 1,000 abandoned carts averaging $105, with a 15% code that costs $15.75 when used.

Treat everyone the same. Suppose 300 shoppers buy after the email. The conventional report says the program recovered 300 sales, $31,500 of revenue, at a discount cost of $4,725.

Now suppose a control group shows that 220 of those 300 would have bought anyway. Only 80 sales were caused by the code, which is $8,400 of revenue won. The discount cost on all 300 is still $4,725, and $3,465 of it went to the 220 who didn't need it.

An uplift model aims to find the shoppers behind those 80 sales. If it can send the code to a smaller group that holds most of the 80, the revenue stays about the same and the discount bill falls sharply. The numbers in your business will differ. The method of finding out is the same.

What it takes

You need three things.

• A history in which a random slice of shoppers received no contact. Without it, you cannot tell sure things from persuadables.
• Enough carts to see patterns: a few thousand abandoned carts is a workable start, and more helps.
• Signals the model can use: what was in the cart, how the shopper arrived, past behavior, price sensitivity, and how long she lingered.

Then act on the output. Persuadables get an incentive. Sure things get nothing, or a plain reminder. Sleeping dogs are left alone.

Guardrails

Review the segments before running them at scale. Check that the model isn't quietly excluding a group of customers for reasons you would not defend. Keep a standing control group so you can see whether the lift persists. Give finance a view of the discount saved against the sales won, so the conversation isn't about conversion rate alone.

Where to start

Begin with a holdout. Without a randomized control group you cannot train an uplift model or prove results. Even two months of clean holdout data changes the conversation with finance. Our piece on the holdout rule covers how to set one up, and the cart recovery field guide shows the method in full.