A geo unit is the geographic area a geo experiment randomises — a postcode, a city, a media market or a whole country. The choice sets how many units you have and how much advertising leaks between them, and those two pull in opposite directions.
A geo experiment randomises places rather than people, and the first design decision is how finely to cut the map. Postcodes, cities, designated market areas, regions and countries are all used. The choice looks administrative and it determines both how much statistical power the test can have and whether the estimate is biased, which makes it the most consequential decision in the design.
The tension is straightforward. Power comes from the number of units, because a geo test is cluster randomization with markets as clusters — twenty markets is twenty observations regardless of how many millions of people live in them. So finer units are better for precision. But advertising does not respect boundaries: television and radio buy at market level, outdoor reaches commuters from neighbouring areas, and people travel. Finer units mean more leakage between treated and untreated areas, which is interference and biases the effect towards zero.
The workable rule is to match the unit to how media is actually purchased and consumed. If your television is bought by DMA, randomising at postcode level guarantees contamination, because the buy itself crosses the boundaries you drew. If the campaign is digital and geo-targeted precisely, finer units become viable and the power gain is real. The unit should be the smallest area at which a treated and an untreated unit do not meaningfully share exposure.
The second constraint is data. A unit needs enough volume for its own outcome to be measurable — a market contributing three conversions a week adds noise rather than information, and a handful of large markets can end up dominating the estimate however many units you nominally have. Checking the distribution of volume across candidate units usually eliminates the finest option before any power calculation is done.
In practice most geo tests end up at DMA level in the US or at regional level elsewhere, with somewhere between twenty and a hundred usable units. That is a small sample by any other standard, which is why matching markets carefully on pre-period behaviour matters so much here and why the estimates carry wide intervals. Being honest about that width is part of reporting a geo test properly.
The two quantities that pull against each other, and the check that usually decides between candidate units.
effective n = number of geo unitsTwenty markets is twenty observations, whatever their population. This is the binding constraint in most geo designs.
observed ≈ true × ( 1 − spillover share )Finer units mean more cross-boundary exposure — see interference.
each unit needs enough conversions to be measurableCompute the distribution of weekly volume per candidate unit. Thin units add noise rather than information.
share of total volume in the largest 3 unitsIf a few markets carry most of the outcome, the effective sample is smaller than the unit count — see the paired t-test calculator.
A retailer plans a geo test for a television and digital campaign. Three granularities are compared on the same market: 210 DMAs, 51 states, and 1,400 usable postcode districts. Television is bought at DMA level.
DMAs are the only viable choice: postcodes are contaminated by the television buy, and states give too few units to detect anything useful.
The postcode option has by far the best power on paper and is unusable, because the television buy crosses those boundaries by construction — treated and untreated postcodes inside one DMA see the same advertising, so the comparison is measuring a diluted version of the effect and there is no way to know by how much. States are clean and there are only 51 of them, giving a detectable effect of 12.4% that is larger than most campaigns produce. DMAs sit where the design has to sit: the unit matches how the media is bought, so contamination is limited, and 210 units support a 6.1% detectable effect. Note the concentration check passing at 19% — had three markets carried half the national volume, the effective sample would have been far below 210 and the DMA option would have been weaker than it looks.
Our “Geo Testing: Unlocking True Incrementality” webinar explored how teams can measure real-world impact when A/B testing isn’t possible - using geographic experiments and synthetic controls to reveal true incremental lift in marketing and product initiatives.
Traditional methods often fall short in measuring the true impact of marketing strategies. Here's how Geo Tests can resolve common marketing challenges.
Applying it to a live measurement problem is the part that goes wrong. If you are designing an experiment, reading a result you do not trust, or trying to work out what your marketing actually caused, that is the work we do.