Geo lifts and the weather

Geo-experiments are popular in GB because user-level holdouts leak and because brands still think in regions for TV and outdoor. The method is sound. The folklore is not. Pairing “similar” cities by last year’s revenue ignores weather, festivals, and a distribution centre that quietly failed on a Tuesday.

We ask students to put a weather covariate in the design document even if they later drop it. The act of naming rain, heat, and travel disruption stops the post-hoc story in which the treated region “just had a better week.” If the week is unusable, drop it. Do not massage it in a footnote the CFO will never read.

Synthetic controls can sit beside a geo lift as a check, not as a rescue. If the synthetic series only behaves once you add five obscure predictors, you are decorating. Conversion lift measurement can survive an ugly map. It cannot survive a map that was edited until it looked causal.

The four-session workshop on geo-experiments exists because national brands keep repeating the same matching mistakes. Bring last year’s media calendar, not a pretty GIS export.