What are the geo test periods?

A geo experiment has a timeline rather than just a duration, and each phase does a distinct job. The pre-period establishes how the markets relate to each other before anything changes. The test period is when spend differs between arms. The cooldown follows the test and captures effects that arrive late. Getting any of the three wrong produces a number that looks fine and answers the wrong question.

The pre-period does double duty and is the phase that decides whether the design is sound at all. It supplies the data for matching test and control markets, and it establishes the baseline relationship the estimate depends on — the correlation between the two sets over that window is the single best indicator of whether the control set is a credible counterfactual. It needs to be long enough to cover a full seasonal cycle, which for most businesses means six months to a year rather than the few weeks teams often allocate.

The test period has to be long enough for the media to work and short enough to afford. The constraint people underestimate is the conversion lag: if purchases typically happen two weeks after exposure, a four-week test spends its first fortnight accumulating conversions from pre-test exposure. That dilutes the measured effect, and the usual remedy is to discard an initial window from the analysis rather than to extend the test indefinitely.

The cooldown is the phase most often skipped and it is what separates a real effect from a timing artefact. A campaign that pulls purchases forward looks identical to one that creates them during the test period; only the weeks afterwards distinguish them. If the treated markets dip below control after the campaign ends, demand was borrowed rather than made, and the test-period lift overstates the true contribution — sometimes to the point of reversing the conclusion.

The cooldown also matters for the next test on the same markets. Adstock means advertising keeps working after it stops, so running a new experiment immediately means the previous campaign's residual effect is still present in the markets that received it. A washout of several weeks between tests is what stops one experiment contaminating the next, and it is a real constraint on how many geo tests a year a business can actually run.

A four-week geo campaign is measured with a two-week cooldown that the team almost skipped. Weekly conversions in treated markets are compared against control across all three phases.

Two phases are doing corrective work here and both would have been missed by a bare test-period readout. The first two weeks were diluted by conversions arriving from pre-test exposure, which is why weeks 3-4 show a much larger effect — the honest test-period estimate discards that initial window rather than averaging it in. More importantly, the cooldown shows treated markets running 3.1% below control after the campaign ended, which is the signature of borrowed demand: customers who would have bought in weeks 5-6 bought in weeks 3-4 instead. The campaign created some genuine incremental demand, and less than half what the test period suggested. A team reporting +4.3% would have overstated the return by more than double, and nothing inside the test period could have revealed it.