Give the effect time.
Plan the treatment and post-treatment windows before reading the result.
Use the pre-test period to assess the design.
Keep treatment and control conditions clear.
Include the lag before finalizing the result.
Run geo tests that change spend by region, compare against what would have happened without it, and read the lift in sales and profit.
Geo test, Meta, Nordics
Started 24 Aug Ends 18 Oct
Lift in net sales
+18%
Test regions vs without ads
Confidence
95%
Interval excludes zero lift
Incremental revenue
€65K
Over 8 weeks
Used to calibrate attribution
×0.60
Meta's reported sales
Used by the teams at
Start with the comparison
Use the pre-test data to build comparable treatment and control groups. Change spend in the treatment regions while preserving the comparison.
Estimate the outcome that would have occurred without the intervention, then assess the difference and its uncertainty.

Change spend
Keep the plan
Compare outcomes across comparable regions.
Design around the decision
Investigate an existing channel, try a new one or test a change in spend. Each design answers a different question.
Explore the examples to see how a holdout and a scaling test create different comparisons.
Pause ads in treatment regions
Estimates depend on test design, statistical power and uncertainty.
Go beyond the headline
Timing, business outcomes and uncertainty all matter. A single uplift number is only part of the evidence.
Plan the treatment and post-treatment windows before reading the result.
Use the pre-test period to assess the design.
Keep treatment and control conditions clear.
Include the lag before finalizing the result.
Read sales alongside contribution and new-customer outcomes.
Use the range around the estimate to judge what the test can support.
A point estimate is more useful when you can see its uncertainty.
Put the evidence to work

Incrementality tests revealed profitable headroom on Google.
Questions
Incrementality testing estimates the effect of a marketing intervention by comparing outcomes with what would likely have happened without that change. Dema uses geo-based experiments, with treatment and control regions and a pre-test period to inform the comparison.
A holdout can investigate what existing spend contributes. A new-channel test can investigate additional demand. Scaling tests can explore what happens when you increase or decrease spend. The test design should follow the question and the decision you need to make.
The comparison needs to account for the regions, their pre-test behavior and the experiment design. Raw sales totals alone are not a causal result. Read the estimated effect with its uncertainty, diagnostics and the conditions under which the experiment ran.
Dema supports gross sales, net sales, net gross profit, new-customer profit, returning-customer profit and new-customer count. These are different ways of reading the same experiment: a channel can increase sales without producing the contribution or new customers you expected.
Allow enough time for the intervention and for delayed outcomes to arrive. The duration depends on volume, expected effect and the design. Dema includes treatment and post-treatment windows; review statistical power and the appropriate timing before launching.
An inconclusive result does not establish that a channel has no effect. The data may not distinguish the effect from noise at the precision the decision needs. Review the interval, test power and execution before deciding whether to repeat, redesign or use other evidence.
Finalized test evidence can inform MMM calibration and the factors used for causal attribution. The scope matters: a result from one channel, market and spend level should be interpreted in that context when it informs a wider decision.