What is a geo test?

Pick regions. Change the ads in some of them. Leave the others alone. Compare sales after the change.

The regions that kept their ads stand in for what would have happened without the change. That is the point of a control. The result is a lift: the difference versus the control. Always say lift against which control.

A geo test is a regional version of a holdout test. It answers the question behind incrementality: what would have happened without the ads?

No. Google's GeoX documentation also lists heavy-up designs, where spend in some regions goes up instead of down¹. And only part of the map takes part. A holdout test does not require losing a whole channel.

This is the hard part. Two common ways:

• Matched markets. Pair regions that look alike.
• Synthetic control. Build a control as a weighted mix of untreated regions that follows the test region's trajectory before the change.

They are ways of building a control, not synonyms for the test. Haus, a measurement vendor, claims its synthetic control gives more precise results than matched market tests². It sells tests, so treat that as one position in the debate.

• Spillover. Ads from a test region reach people in a control region, or the other way round. The result drifts.
• Too small an effect. If the design cannot detect it, the result is noise. See minimum detectable effect.

For a planning guide, read how to run a geo holdout test. For choosing between methods, read geo test vs conversion lift test.

For ZFD: A geo test needs a design someone has signed off before the result, not after it.