Quick answer
A geo-lift experiment changes advertising or another marketing treatment in selected non-overlapping geographic markets and estimates the resulting difference from control markets or a synthetic counterfactual. Randomized matched-market designs assign comparable geographies to treatment and control before the campaign and provide the strongest basis when feasible. Quasi-experimental synthetic-control methods construct a weighted combination of untreated markets that tracks the treated market before intervention, adding assumptions about comparability and unobserved shocks. A sound study defines the intervention, aggregate KPI, market unit, test period and causal estimand; uses sufficient stable pre-period data; plans power and minimum detectable lift; limits spillover and concurrent regional changes; verifies pre-period fit and actual media delivery; and reports lift, incremental cost or return, uncertainty and sensitivity. Geo tests reduce dependence on person-level identity but can be expensive, underpowered and vulnerable to contamination, seasonality, stock, pricing and local shocks.
What is a geo-lift experiment?
A geo-lift experiment estimates causal marketing impact using geographic units such as cities, regions, postal areas or designated markets. Treatment changes advertising availability, spend or strategy in selected markets while aggregate outcomes are compared with a counterfactual.
The method is useful when user-level holdouts are unavailable, offline sales matter or media affects whole markets. It can evaluate television, out-of-home, search, social, retail media and multi-channel interventions if geographic delivery and outcome data are sufficiently controlled.
Geo lift is a family of designs, not one estimator. Randomized market experiments, matched-market tests and synthetic-control quasi-experiments make different assumptions and should be labeled accurately.
Randomized, matched and synthetic-control designs
In a randomized geo experiment, non-overlapping markets are assigned to treatment and control conditions. Matching or stratification before randomization can improve precision by grouping markets with similar pre-period outcomes and characteristics.
Operational constraints sometimes prevent random assignment. A matched-market quasi-experiment selects comparison geographies with similar history. Synthetic-control methods construct a weighted combination of untreated markets that reproduces the treated unit's pre-intervention path.
Synthetic control can create a stronger comparison than one chosen market, but it relies on good pre-treatment fit, a suitable donor pool and no unaccounted post-period shock that affects treatment differently. It does not turn a nonrandomized design into an RCT.
Estimand and KPI
Define the treatment, comparison, markets, aggregate outcome, lag and decision threshold.
- What exactly changes by geography?
- Which outcome can be measured consistently everywhere?
Data and feasibility
Assemble complete stable pre-period panels and assess targeting, spillover and number of independent markets.
- Can media be controlled by market?
- Are outcome data complete at the same granularity?
Power and selection
Simulate decision-relevant effects and choose balanced or well-fitting test and control markets prospectively.
- What lift can the design detect?
- Does the counterfactual fit before treatment?
Execution
Deliver the assigned media difference and log every regional shock or contamination risk.
- Did spend and reach separate as planned?
- Did price, stock or promotion differ?
Estimate and act
Compare observed and counterfactual outcomes, quantify uncertainty and connect lift to bounded economics.
- How robust is lift to reasonable specifications?
- Which next decision follows at this scale?
Build a complete geographic panel
At minimum, data need a consistent time variable, geographic identifier and numeric KPI for every market-period combination. Daily data often provide more time points than weekly aggregates, provided day-level noise and operational timestamps are reliable.
Use a stable pre-period long enough to learn seasonality and market relationships. Meta's GeoLift documentation recommends complete panels and substantial history, while its best-practice guidance stresses daily granularity and enough pre-treatment periods and geographies for planning.
Include pre-treatment covariates only when they improve prediction and are available consistently. Record store changes, price, stock, distribution, promotions and measurement revisions. Never fill missing outcomes silently or drop markets after seeing treatment results.
Plan power before choosing markets
Geo tests often have few independent units and noisy sales outcomes. Millions of transactions inside ten cities do not create millions of randomized units. Power depends on market count, pre-period predictability, outcome variance, treatment intensity, test length and analysis method.
Define a minimum lift that would change the decision and simulate the full design under plausible effects and nulls. Select markets, duration and spend prospectively. A design that can detect only an implausibly large effect should not be launched for reassurance.
Holdout cost matters. A larger treatment or control share may improve precision but sacrifice short-term media opportunity. Choose the smallest design that answers the business question honestly, not the smallest budget that can produce a chart.
Select markets without looking at the answer
Market selection should use pre-intervention data and operational constraints only. Evaluate balance or synthetic pre-fit, donor weights, residual patterns and placebo performance before treatment begins. Preserve the candidate process and exclusions.
Avoid a donor pool contaminated by similar campaign changes or markets tightly connected through commuting, media spill, fulfilment or tourism. A synthetic control built from exposed regions understates the contrast.
Do not choose the comparison after observing which untreated market happened to decline. This outcome-driven selection manufactures lift. If a design must change after launch, classify the analysis as exploratory and explain the resulting bias risk.
Protect treatment separation
Geo targeting is imperfect. Media crosses borders, people travel, national campaigns reach every market and platform delivery can under-spend small regions. Measure actual incremental spend, reach or impressions and report whether assignment created the planned contrast.
Freeze or balance other controllable marketing where possible. Record regional pricing, promotions, distribution, stock, weather, events, competitor actions and outages. A random design protects against expected balance, but a small set of markets can still experience consequential shocks.
Run through the purchase cycle and planned conversion lag. Avoid stopping when the cumulative effect looks favorable. Treatment carryover can require a post-period, but define it prospectively and distinguish persistence from delayed measurement.
Geo-lift experiment example
The refill-store network defines a fixed incremental CTV treatment and keeps other paid-media policy stable. Verified first purchases and contribution are available in every market, making the aggregate KPI more defensible than platform-attributed visits.
The incident log makes stock and local activity visible, while randomized matched groups preserve a clear design. The final decision uses lift uncertainty and incremental economics at the tested spend.
A hypothetical refill-store network wants to know whether connected-television prospecting creates new first purchases beyond existing local demand across its city footprint.
Treatment is a fixed incremental CTV plan while other paid media follows a frozen policy. The KPI is daily verified first purchases with contribution after expected returns, observed through the normal purchase lag.
The analyst assembles complete daily city histories, store openings, price, promotions and stock. Markets with boundary leakage or structural changes are excluded before outcome results are known.
Power simulations select market groups and treatment intensity capable of detecting a decision-relevant lift. Within matched sets, assignment is randomized when operational constraints permit.
Delivery, reach and spend confirm treatment separation. National creator activity, stockouts, weather, competitor events and local discounts enter a dated incident log.
The report shows treatment and counterfactual trends, absolute and relative incremental first purchases, uncertainty and incremental contribution. An inconclusive interval leads to redesign, not an attributed-sales victory claim.
This example is hypothetical. Design and inference must match whether markets were randomized, matched or evaluated with a quasi-experimental synthetic control.
Estimate lift with the design you ran
For randomized market designs, analyze the treatment-control contrast in a way that respects assignment, market size and clustering. Pre-period adjustment can improve precision when specified appropriately. For synthetic control, compare observed treatment outcomes with the weighted counterfactual and examine fit and placebo evidence.
Report treatment and control trends, absolute lift, relative lift and uncertainty. Translate the effect into incremental CPA, return or contribution using a clearly defined incremental cost. Economic ratios inherit uncertainty from the lift estimate.
Run prespecified sensitivity checks for market exclusions, pre-periods, model choices and shock periods. Sensitivity is not permission to select the preferred result. Show the range and explain why the primary specification was chosen.
Interpret a local, scale-specific effect
The estimate applies to the tested intervention, markets, dates and operating conditions. Increasing budget can change reach, frequency, auctions and saturation, so lift should not be extrapolated linearly to a national plan.
A null or inconclusive result may reflect low effect, low power, contamination or unstable counterfactual fit. Distinguish these possibilities. Non-significance is not proof that the channel never works, and a positive point estimate with a wide interval is not a win.
Replicate material findings across another set of markets or season when feasible. Combine geo evidence with user-level tests, MMM and operational knowledge, noting that each method can estimate a different scope of effect.
Limitations and common mistakes
Geo experiments can be costly, slow and underpowered. Geographic outcomes may be affected by local shocks, cross-border behaviour and a small donor pool. Privacy is improved by aggregate outcomes, but market data can still require governance and minimum aggregation.
Common mistakes include choosing markets without power analysis, using a short unstable pre-period, selecting controls after results, ignoring spillover, changing promotions unevenly, comparing before and after without control and reporting synthetic-control output as randomized evidence.
A geo tool cannot fix an intervention that fails to separate, incomplete sales data or missing markets. Design feasibility should be proven before campaign commitment, including a dry run of targeting, ingestion and reporting.
Geo lift is credible when the counterfactual and treatment separation are designed before the campaign outcome is visible.
Frequently asked questions
What data does a geo-lift test need?
It needs consistent outcome data by time and geography, sufficient stable pre-period history, treatment information and any prespecified covariates or operational controls.
Is every geo-lift study randomized?
No. Some randomize matched markets; others use quasi-experimental matching or synthetic controls. Label the design and assumptions accurately.
How long should a geo test run?
Duration follows power, data granularity, purchase cycle, seasonality and lag. Plan it through simulation rather than using one universal number of weeks.
What is a synthetic control?
It is a weighted combination of untreated markets designed to reproduce the treated market's pre-intervention outcome path and approximate its counterfactual.
Can geo lift measure offline sales?
Yes, if offline outcomes are consistently attributed to geographic units and available for treatment and comparison markets across the required time periods.
Sources and further reading
- Google Research: Measuring Ad Effectiveness Using Geo Experiments ↗Primary research on randomized non-overlapping geographic treatment and control designs
- Meta Open Source: GeoLift Methodology ↗Official methodology for incrementality, synthetic controls, inference and power planning
- Meta Open Source: GeoLift Best Practices ↗Official operational guidance on data history, granularity, duration and market design
- Google Research: Time-Based Regression Matched Markets ↗Primary research on systematic matched-market design under practical constraints