A test that switches ads off, or on, in some regions and not others, then compares all sales between the two groups.
A geo experiment splits a country into regions, picks some at random to change, and leaves the rest as they were. For brand search, the change is usually switching brand ads off in the test regions. Because the regions are chosen at random and measured over the same weeks, a difference in total orders between the two groups comes from the ads, not from the season or a sale.
Google's researchers set out the design in Vaver and Koehler's 2011 paper: regions randomly assigned to test or control, ads targeted to match, a pre-period with nothing changed, then the test period. Ranking regions by size before randomising narrowed their result by about a tenth. Later Google work on time-based regression was built for tests with few regions, which suits a country the size of the UK.
- Measure orders and revenue from every channel, not paid clicks, which always fall when ads go off.
- Ads are kept out of test regions with location targeting, which works on where Google thinks the searcher is, so expect some spill at the edges.
- Record who appears on your name in the test regions. A new advertiser there is part of the result.
The guide to bidding on your own brand has the steps, and incrementality is what the test measures.