Uplift modeling vs. lookalike targeting: persuadability beats predicted-conversion

Uplift modeling vs. lookalike targeting: persuadability beats predicted-conversion

Most targeting optimizes for who is likely to convert. Uplift modeling optimizes for who converts because you reached them. The distinction is the difference between correlation and causation applied to targeting.

The four-quadrant logic
Uplift (also called incremental-response or true-lift) modeling sorts users into four types: sure things (convert with or without the ad), lost causes (never convert), do-not-disturbs (the ad actually pushes them away), and persuadables (convert only if reached). Only persuadables generate incremental return. A standard lookalike or propensity model targets the sure things hardest — they look most like converters — and wastes spend on conversions you'd have won for free.

Why this is hard
Uplift can't be observed directly: you never see the same user both treated and untreated. It must be estimated from randomized-treatment data using two-model, class-transformation, or causal-forest approaches. That randomized training data is the price of admission — and exactly why most teams default to lookalikes instead.

The connection to attribution
This is the targeting-side mirror of incrementality. The same bias that makes last-click over-credit retargeting makes lookalikes over-target sure things: both reward presence near conversion rather than influence over it.

Bottom line for practitioners: Lookalikes are cheap and fine for reach, but for performance budgets, uplift modeling is the causally honest target — it spends only where spending changes the outcome. If you can run a randomized holdout to generate training data, build uplift. If you can't, at minimum exclude obvious sure-things (recent buyers, branded-search clickers) from retargeting, which captures much of the gain crudely.