Most targeting models answer one question: who is likely to buy? Uplift modeling answers a better one for a loyalty budget. Who will buy because of the offer, and who would have bought anyway?

What is uplift modeling?

Uplift modeling is a set of methods that predict the effect of an action on each customer. Gutierrez and Gérardy define it as modeling “the incremental impact of an action or treatment on a customer outcome” (Gutierrez and Gérardy, 2017). The action can be an offer, an email or a retention call. The outcome can be a purchase, a renewal or a member staying active.

The hard part is that no one sees both outcomes for the same person. A member either got the offer or did not. Statisticians call this the fundamental problem of causal inference. So uplift modeling “crucially relies on randomized experiments,” in which a random control group stands in for what treated customers would have done (Gutierrez and Gérardy, 2017).

How is uplift modeling different from response or propensity modeling?

A response model ranks people by how likely they are to buy after an offer. Its top picks are often people who buy every week with or without the offer. Schoder and Rößler, writing in California Management Review, say response models often target these buyers. The models cannot tell them apart from the people an offer moves (Schoder and Rößler, 2025). An uplift model ranks people by the change the offer causes.

The gap matters most in retention. Eva Ascarza tested the habit of targeting the members most likely to leave. She used field data from a wireless provider and a membership organization. Targeting by lift, the chance that the effort works, cut churn more than targeting by risk of leaving (AMA, 2019). A predictive churn score tells you who is at risk. It does not tell you who can be saved.

What are the four customer types in uplift modeling?

Uplift work sorts customers into four groups by how they act with and without the offer (Schoder and Rößler, 2025). Only one group pays back the cost of the offer.

TypeWith the offerWithout the offerWhat a loyalty team does
PersuadablesBuysDoes not buyTarget them. This is the only group where the offer adds sales.
Sure ThingsBuysBuysSkip paid offers. The reward goes to a purchase that was coming anyway.
Lost CausesDoes not buyDoes not buySkip them where each contact has a cost.
Sleeping DogsDoes not buyBuysNever target. The contact itself drives them away.

Sleeping Dogs, also called Do-Not-Disturbs, react badly to the contact itself (scikit-uplift). They are real. Radcliffe and Surry describe a mobile churn campaign that raised churn from 9% to 10%. An uplift model found a 30% segment worth targeting. Treating only that segment cut churn from 9% to under 8%, with 70% less spend (Radcliffe and Surry, 2011).

How do you build an uplift model?

Start with a randomized pilot. Send the offer to a random treated group and hold out a random control group. Record the outcome and each member’s data for both. Then train a model on the gap between the two groups.

The simplest method is the two-model approach. Train one model on the treated group and one on the control group, then subtract their predictions. Gutierrez and Gérardy group the field into three approaches: the two-model approach, class transformation and direct uplift modeling (Gutierrez and Gérardy, 2017). Uplift trees, a direct method, pick splits that separate members with high uplift from members with low uplift (Radcliffe and Surry, 2011).

Newer “meta-learners” wrap any standard model. The S-learner fits one model with the treatment as a feature. The T-learner fits two. The X-learner suits tests where one group is much larger than the other, as a small holdout is (Künzel et al., 2019). Open-source libraries such as Uber’s CausalML and Booking.com’s UpliftML implement these methods (CausalML; UpliftML).

How do you know if an uplift model works?

You cannot check a single prediction, since no member was both treated and untreated. Instead, rank members by predicted uplift and compare treated and control results within each slice. The Qini curve plots the extra sales gained as you target more of the ranked list. The Qini measure is the area under that curve (Radcliffe and Surry, 2011).

Judge the model in profit as well as response. In one campaign, conversion was 15.07% in the control group, 18.04% when everyone was treated and 17.32% with uplift targeting. Profit per customer told a different story: $5.26, $5.18 and $5.46 (Schoder and Rößler, 2025). Treating everyone won on conversion and lost on profit.

How do loyalty programs use uplift modeling?

A loyalty program is a strong fit, because it already knows who got which offer. Members carry an ID, so every treated and held-out member can be traced to their next purchase. The program’s best members are often Sure Things. A discount sent to them costs the full reward and adds little. Uplift modeling shows where a softer touch, or no offer at all, keeps the same sales.

Some firms treat Sure Things on purpose. Schoder and Rößler note the argument that loyal customers need to be kept happy with a treat from time to time (Schoder and Rößler, 2025). If a program makes that call, it should budget the cost as recognition and measure it with its own holdout. Our guide to loyalty personalization covers how to vary the offer by member.

The more often a program runs campaigns, the more the targeting is worth. Schoder and Rößler work through a retailer with 50,000 customers, where an offer could add at most 1.17% more conversions. At $50 each, the best case is 585 extra conversions and $29,250 per campaign. Run monthly, that comes to $351,000 a year (Schoder and Rößler, 2025). A program that sends offers every week gains more again.

Use it where offers cost real money: win-back bonuses, tier-saving offers and points multipliers. The same control group testing that proves a campaign worked becomes the training data for the next one.

For illustration, a program has 100,000 lapsing members. 40,000 are Sure Things, 10,000 Persuadables, 45,000 Lost Causes and 5,000 Sleeping Dogs. Each purchase brings $25 of margin, and the bonus costs $4 for each member who buys. Offer it to everyone, and 50,000 members buy with the bonus, at a cost of $200,000. The 10,000 Persuadables add $250,000, and the 5,000 Sleeping Dogs take away $125,000. The campaign loses $75,000.

Offer it only to the Persuadables, and 10,000 bonuses cost $40,000 against $250,000 of new margin. The Sleeping Dogs keep buying. The campaign makes $210,000. A real model sorts members far less cleanly than this. The direction of the result holds.

How GRAVTY handles uplift modeling

Uplift work needs clean groups, full histories and fast changes to targeting. GRAVTY keeps every earn, burn, tier change and offer response on one member record through Member 360. On lji.io, AgenticTest attributes incremental lift at the segment level while a campaign is still running. Visual Rules, the patented no-code rules engine, puts a new targeting rule live without an IT ticket.

Frequently asked questions

Is uplift modeling the same as incrementality?

No. Incrementality is the effect a campaign caused across a whole group. Uplift modeling predicts that effect for each customer, so you can choose whom to target next time. Both rest on a random control group (Gutierrez and Gérardy, 2017).

How much data does an uplift model need?

Much more than a simple lift test. Radcliffe and Surry say the control group must be at least ten times larger for modeling than for measuring lift. Their second rule is that the overall uplift times each group’s size should be at least 500 (Radcliffe and Surry, 2011). At a 1% uplift, that means 50,000 members in each group.

Can you build an uplift model without a control group?

Not reliably. Without random assignment, the treated and untreated groups differ in ways the model cannot see. Gutierrez and Gérardy call it a wrong belief that you can always get the effect from observational data by comparing the two groups (Gutierrez and Gérardy, 2017). Run a random pilot first.

Which algorithm is best for uplift modeling?

None wins every time. In simulations by Künzel and colleagues, the X-learner did well, but no meta-learner was best in every case (Künzel et al., 2019). Compare several methods on held-out data with a Qini curve, and pick the one that earns the most profit at the volume you can afford to target.

When is uplift modeling not worth the effort?

When contacting extra customers costs next to nothing. Schoder and Rößler note that online treatment costs can be close to zero. Uplift modeling only beats simpler methods if treating Lost Causes has a cost (Schoder and Rößler, 2025). In that case, treating everyone or a simple response model can be enough.