Geo-Experiment

What It Is

A geo-experiment measures causal ad effect by randomizing geographic regions rather than users — the practical answer when privacy walls, offline conversions, or whole-funnel channels make user-level randomization impossible. Non-overlapping geos are split into control and treatment; treatment geos get the spend change; the difference, modeled against a counterfactual, is the incremental effect.

Design & Method (geo-based regression)

  • Group formation: stratified randomization (strata of similar-size geos), or matching geos that are predictive of each other when geos are few/heterogeneous. Rule of thumb: ≥30 geos.
  • Phases: pretest (4–8 wks, identical campaigns) → test (3–5 wks, treatment gets the change) → optional cool-down for delayed conversions.
  • Model: y₁,ᵢ = β₀ + β₁·y₀,ᵢ + β₂·δᵢ + εᵢ, where y₀ is pretest response (controls trend/seasonality), δᵢ is incremental spend (observed − counterfactual), and β₂ = iROAS read directly. Weights 1/y₀,ᵢ handle size heteroscedasticity. Precision reported as a CI on iROAS.

(Method: Vaver & Koehler, Google Research; operational detail from Google's "Estimating causal effects using geo experiments.")

How It Applies to Marketing Factory

Geo-experiments are platform-agnostic incrementality the factory can run without a walled-garden lift product — agent-ownable in the mechanics (form matched geo groups, apply geo-targeted spend, fit the regression, report the iROAS CI). Because a test has ~30–200 units over a few weeks, it is a small-N experiment: apply the minimum-detectable-effect gate to check the spend change is detectable, and avoid peeking via sequential-testing or a Bayesian decision. It is one method of incrementality-testing and yields incremental-roas.

Referenced from: incrementality-and-geo-experiments