Measurement and evidence
What is a geo-lift test?
Free, with no card. We read your homepage, draft who buys from you and show your monthly price before you join the waitlist. Nothing is bought until you've approved the plan and the ads.
what it means here
Here is what the term means for a UK buyer
The definition above holds anywhere. Buying media in the UK adds the detail below, including where AdBuyMCP does something different.
The most practical causal method available to a brand advertiser, and the one that suits out-of-home, cinema and television: channels where person-level measurement is impossible. It needs enough independent markets, enough spend and enough conversion volume, which is why a twenty-catchment retailer can usually power one and a single national campaign usually cannot. AdBuyMCP also refuses a contaminated design: if every candidate control market received delivery from the campaign, none is a holdout, and it says so and stops.
How are the control markets chosen?
AdBuyMCP matches candidates from twelve UK conurbations, from Greater London and Greater Manchester to Edinburgh and Cardiff. Each candidate is ranked on how closely its pre-campaign conversion series correlates with the test market's, weighted at 70%, and on population similarity, weighted at 30%, and the top three become the control group. Correlation does most of the work, because the design assumes parallel trends and a control market half the size is fine if it moves in step.
The two series are joined on date before they are correlated. A day with no conversions often produces no row at all, and aligning two gappy series by position can invert the correlation. Any market the campaign delivered into is excluded before matching starts.
How is a geo-lift result read?
The readout is a difference-in-differences with pre-period ratio scaling. The control series is scaled by the ratio the two groups held before the campaign, and the gap after exposure is the estimate, reported with a confidence interval from Student-t inference, which suits the short windows media tests run on.
A result may call itself causal only when every geographic outcome is observed, every delivery row maps to a live line and control exposure is verified. Sandbox delivery is labelled modelled, so no sandbox campaign reaches a causal grade, and AdBuyMCP has produced no causal result for anyone yet.
Type your company website or work email to see who we think buys from you and what reaching them 3, 5 or 12 times a week would cost.
Free, with no card.
related terms
These terms sit next to this one
Most confusion in media buying comes from two adjacent terms being used interchangeably, so these are the neighbours worth reading next.
Incrementality
The share of an outcome that happened because of the advertising, measured against what would have happened anyway. It is measured by comparing an exposed group with a comparable unexposed one, which counting the conversions that followed an advert cannot do.
Holdout
A group deliberately excluded from seeing a campaign and kept as a comparison against the group that was exposed to it. Without one, the counterfactual (what would have happened anyway) has to be assumed rather than observed.
Matched market
A control market chosen because its historical outcome pattern closely tracks the test market's, so that a later divergence can be attributed to the campaign.
Difference-in-differences
A method that compares the change in an exposed group against the change in a control group over the same period, so that anything moving both in step (seasonality, a competitor, the weather) cancels out.
Statistical power
The probability that a test will detect a real effect of a given size, if an effect of that size genuinely exists in the population being measured. A design with low power will usually miss a genuine effect, raises the share of the results it does produce that are noise rather than signal, and exaggerates the size of the effects it does detect.
same group
These terms are also in Measurement and evidence
What you can prove afterwards, and how to tell a number that means something from one that does not.