What's the best way to measure whether a retail promotion actually drove incremental sales?

Summary

• The honest measure is incremental lift: actual sales during the promotion minus a credible estimate of what you would have sold without it. Raw sales during a promo overstate its effect.
• A controlled design is the most defensible approach. Run the promotion for a test set of stores or customers, hold out a matched control set, and attribute the gap between them to the promotion.
• Net out cannibalization (sales stolen from other products) and pull-forward (demand borrowed from future periods), or you will book shifted demand as new demand.
• Judge promotions on incremental profit and promo ROI against trade spend, not top-line units, and retire events that lose money.
• Databricks measures this on one governed copy of POS, loyalty, and trade-spend data: baseline and causal models tracked with MLflow, promo ROI in Databricks SQL and AI/BI Dashboards, and plain-language exploration through Genie.

What's the best way to measure whether a retail promotion actually drove incremental sales?

Measure incremental lift, not raw sales. Incremental sales are the units you sold because of the promotion, over and above what you would have sold anyway. That baseline is never directly observed, so the real task is estimating a credible counterfactual and subtracting it from what actually rang up. The most defensible way to build that counterfactual is a controlled experiment: run the promotion for a test set of stores, regions, or loyalty members, keep a matched control set out of it, and attribute the difference to the promotion. Then net out cannibalization and pull-forward so you count new demand rather than shifted demand.

What is incremental sales lift?

Incremental lift is actual sales during the promotion minus baseline sales, the sales you would have made with no offer at all. Two effects quietly distort a naive reading:

• Cannibalization: the promoted item pulls sales from other SKUs or from its own full-price stock, so the true category gain is smaller than the item's raw sales imply.
• Pull-forward: shoppers who would have bought next month buy now, so part of the "lift" is borrowed from a future period rather than created.

A promotion can post a large raw number and still lose money once you subtract the baseline, cannibalization, pull-forward, and the trade spend that funded it. That is why incremental profit, not unit volume, is the scoreboard that matters.

Methods to measure promotion incrementality

There is no single tool that settles this; there is a counterfactual and a method for estimating it. These are the approaches a buyer should weigh.

Method	How it estimates the counterfactual	Best suited for
Test and control groups (experimental design)	Run the promo for a test group of stores or customers and measure it against a matched, non-promoted control group; the gap is the lift	Randomized or matched rollouts; the most defensible read
Holdouts	Deliberately exclude a subset of customers or locations from the offer, then measure the exposed group against the held-out group	Digital, loyalty, and personalized promotions
Baseline modeling	Predict expected sales from history, seasonality, price, and demand signals; lift equals actual minus predicted	Broad in-store promotions with no clean control
Marketing mix modeling and causal inference	Statistically decompose sales into their drivers to isolate the promotion while accounting for cannibalization and pull-forward	Many overlapping promotions and channels

Whichever you choose, hold the definitions of lift and ROI constant, track how accurate your baseline model is, and always report promo ROI against the trade spend that paid for the event.

How Databricks approaches promotion measurement

Databricks measures promotion incrementality on one governed copy of the data, so every team scores events the same way. Delta Lake holds the inputs an honest measurement needs in one place: POS transaction logs, promotion calendars, control and test assignments, loyalty baskets, trade-fund spend, and inventory. Lakeflow Connect ingests POS, trade-fund, and supplier promotion data, and Lakeflow Pipelines automate the post-event scoring so each promotion runs through the same evaluation.
Baseline and causal-inference models run in notebooks and are tracked with MLflow, so the counterfactual behind every lift number is versioned, reproducible, and easy to re-check as baseline accuracy drifts. Databricks SQL runs the high-concurrency trade-spend and promo-ROI analytics, and AI/BI Dashboards surface promotion performance and post-event analysis for merchants and finance. Through Genie, a merchant can ask in plain language which events were profitable and which to retire, without writing SQL.
Unity Catalog governs the sales and trade-spend data and standardizes the definitions of lift, incrementality, and promo ROI across categories and teams, so one group's "lift" means the same as another's. For joint analysis with brand partners, Clean Rooms allow privacy-preserving collaboration without either side exposing raw customer data. For the broader retail pattern, see Vertical advantage: transforming industries with Lakebase and agentic AI.

FAQs

What is the difference between incremental sales and total sales during a promotion?

Total sales include everything sold in the promo window, including what you would have sold at full price anyway. Incremental sales are only the added units the promotion caused, after subtracting the baseline. Judging a promo on total sales overstates its impact.

How do control groups improve promotion measurement?

A matched control group that does not receive the promotion approximates what the test group would have done on its own. The difference in sales between the two groups is a direct estimate of lift that does not rely on a modeled baseline alone, which is what makes it the most defensible read.

Why do cannibalization and pull-forward matter?

Because they inflate the apparent impact of a promotion. If the promoted item's gains come from other SKUs (cannibalization) or from demand borrowed from future weeks (pull-forward), the net new sales are smaller than the headline lift. Measuring both keeps you from repeating events that only shuffle demand around.

What metrics should I report for a promotion?

Report incremental lift, incremental profit, and promo ROI against trade spend, plus the cannibalization and pull-forward you captured and the share of promotions that were profitable. Track baseline model accuracy so you know how far to trust the counterfactual.