Uplift

Uplift is the difference between the treated group and the control group in a controlled test, expressed as an absolute or percentage change in the chosen metric. If the control converts at 2.0% and the personalised variant at 2.3%, the uplift is 0.3 points, or 15% relative. It comes with a confidence interval, and a claimed uplift without one should be treated as an anecdote.

Uplift modelling goes a step further: rather than predicting who will buy, it predicts for whom an action changes the outcome. Customers fall into four groups: those who buy anyway, those who never buy, those who buy only if nudged and those who are put off by the nudge. Only the third group is worth the cost of an offer. Uplift models estimate each person's incremental response and are the engine behind well-run next best action systems.

It is not the same as a propensity score, which ranks likelihood without asking whether the action mattered, and it is not the raw metric movement after a launch, which is confounded by everything else that happened. Uplift is the number that survives an A/B test.

• CFO: it is the only figure that belongs in a personalisation ROI calculation; total revenue from recommended items is not incremental revenue.
• Founder / CEO: vendors quoting uplift should be asked how it was measured; the answer separates engineering from marketing.
• Product manager: uplift by segment tells you where personalisation works and where it does harm, which shapes the roadmap.
• Data lead: uplift modelling needs randomised treatment data to train on, so the experimentation programme feeds the model, not the other way round.

Uplift exists as a concept because attribution is seductive and wrong. It is easy to sum the revenue from every purchase that followed a recommendation and call it the engine's contribution, ignoring that most of those purchases would have happened anyway. Discounts sent to loyal customers are pure cost. Measuring incremental effect, and modelling who is actually persuadable, directs spend to where it changes behaviour. The trade-off is that uplift is harder to measure than volume, needs randomised data and produces smaller, more honest numbers than the alternatives, which makes it unpopular with anyone selling a result.

• Reporting a recommender's contribution to a D2C brand as incremental revenue per visitor from a holdout group, not attributed sales.
• Targeting retention discounts only at subscribers whose churn probability an offer actually reduces.
• Deciding which trial accounts a SaaS sales team should call, based on incremental conversion rather than raw likelihood.
• Measuring whether payment reminder calls change repayment behaviour for a lender, or merely reach people who would have paid.
• Evaluating whether a school's intervention programme improves outcomes for the students it selects.

MetricA metric, and the modelling discipline around it. Eazyware reports personalisation results as uplift with confidence intervals from a controlled test, and builds uplift models for personalisation engines when treatment data exists.

Eazyware service that covers it: Personalization Engines. Starting prices are on the pricing page.