Course outline

Geo Experiments and Incrementality

By the end of this lesson, you should be able to: explain why attribution and incrementality are different quantities, design a geo holdout, analyse it at the right sample size, and know why the assignment step deserves more care than the analysis.

Two numbers that sound like the same number

Alder Ads reports on the campaign:

Orders in markets running ads1,685,037
Of those, last touch was an ad214,000
"Advertising drove"12.70% of orders

Nothing there is wrong. Those 214,000 people really did click an ad, and the ad really was the last thing they touched before buying.

The question a budget needs answered is a different one: how many of those orders would not have happened?

Turn the ads off and find out

Half of Alder's 31 markets went dark from day 84. The other half kept running. Which markets went dark was decided in advance, not by how they were performing.

BeforeAfterChange
Ads stayed on3,8354,121+7.46%
Ads switched off3,0203,067+1.57%

Markets that lost advertising entirely still grew 1.57%, because most of the demand was never coming from ads. The gap between the arms is what advertising added:

Incremental lift+5.05%
95% interval[+2.36%, +7.82%]
Planted truth+4.10%
Three panels. On the left, two indexed order series that track together for twelve weeks and separate after the switch, with the ads-off arm falling behind. In the middle, two bars: last-click attribution at 12.70% and the geo holdout at 5.05%, with a dashed line at the true 4.10%. On the right, two overlapping histograms of the estimate across 600 re-assignments, the coin-flip one wide and the matched-pair one much narrower, both centred on the truth.
Left, the holdout. Middle, credit against causation. Right, what the assignment step buys you.