GeoLift is Meta's open-source R package for planning and analysing geo experiments. It uses synthetic control to build the counterfactual and, unusually, includes power simulation that tells you which markets to treat before any money is spent.
GeoLift is an open-source R package from Meta for running geo experiments. Analytically it is built on synthetic control: the counterfactual for the treated markets is a weighted combination of untreated ones, with the weights fitted so the blend tracks the treated set closely over the pre-period. That part is well-established methodology rather than anything proprietary.
What distinguishes it in practice is the planning half. Before committing budget, it simulates the experiment across candidate market sets and reports the minimum detectable effect each would achieve for a given duration and spend. That answers the question most geo tests fail to ask — whether the design can detect the effect you expect — and it frequently returns an uncomfortable answer, which is precisely its value. A great many geo tests are run without the power to resolve anything, and the finding arrives after the money is spent.
The market-selection output is the practical payoff. Rather than treating the largest markets or the ones a media team prefers, it searches for the treatment set whose synthetic control fits best and whose size supports the effect you want to detect. Those two criteria pull against each other — the best-fitting markets are not necessarily the ones with enough volume — and having the trade-off computed rather than argued is a real improvement on how these decisions are usually made.
It inherits synthetic control's assumptions and they are the things to supervise. The donor pool must be genuinely untreated, which for advertising means checking the media plan rather than the correlation matrix, since a market receiving spillover fits especially well. The pre-period must be long enough and stable enough for the weights to mean something. And the fitted weights should be inspected: a synthetic control resting almost entirely on one donor market is fragile, because anything happening in that market propagates straight into the estimate.
The honest framing is that it is a well-built implementation of a known method with unusually good planning tooling, not a different kind of evidence. A randomised geo test where assignment creates the counterfactual is still stronger. GeoLift's contribution is making a careful synthetic-control analysis accessible to teams who would otherwise run a before-and-after comparison — which is a substantial improvement on the realistic alternative rather than on the ideal one.
The estimator it implements, and the planning quantity that is its distinguishing feature.
ŷ_treated,t = Σ wⱼ · y_donor j,t, wⱼ ≥ 0, Σ wⱼ = 1Non-negative weights summing to one, fitted on the pre-period. The constraint prevents extrapolation beyond the donors.
lift = Σ ( y_treated,t − ŷ_treated,t ) over the test windowActual minus synthetic counterfactual, accumulated — see counterfactual forecast.
simulate known effects over candidate market sets; report detectable MDERun before spending. The feature that most distinguishes it from a bare synthetic control implementation.
inspect the fitted weightsA synthetic control dominated by one donor is fragile — see the correlation calculator for the pre-period fit.
An advertiser plans a geo test with a £400,000 budget and expects roughly a 5% lift. GeoLift's power simulation is run across candidate market sets before anything is committed.
The media team's preferred ten large markets could only detect an 11.2% effect. A different set of 24 markets detects 4.6%, below the effect expected.
The first two rows are the case for running the simulation at all. Treating the ten largest markets is the intuitive plan — the most spend, the most volume — and it would have produced a test incapable of detecting the effect the campaign was expected to deliver. The likely outcome would have been an inconclusive result read as "the campaign did not work", which is the worst possible return on £400,000 of media and several weeks. The selected set of 24 markets is a mix of sizes chosen because their synthetic control fits well, and it brings the detectable effect below the expected one with room to spare. The final row is the fragility check passing: the largest donor weight is 0.19, so no single market dominates the counterfactual. Had one donor carried 0.6 of the weight, a store closure or local promotion there would have propagated straight into the estimate, and the design would have been worth revisiting despite a good headline fit.
When your advertising efforts are up and sales are increasing, that’s great news, but it doesn't necessarily mean the ads are the reason. Here are three popular tools to measure the lift of a campaign.
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.