Most marketing measurement starts with a person. A cookie, a device ID, a hashed email. The apparatus of attribution exists to follow that person from an ad to a checkout and then argue about which touch deserves the credit. When the person becomes hard to follow, which is what happened across the last five years, the apparatus breaks.
Geo experiments throw out the person. They do not measure individuals, they measure places. You run your campaign in some cities, withhold it from comparable cities, and read the difference in sales between the two groups. That difference is the causal lift. No one was tracked, and no consent box was needed because there was nothing to consent to. A market cannot be de-anonymized.
This is not a workaround that limps along until something better arrives. The engine underneath the good versions of it, the synthetic control method, came out of academic econometrics, where the standard of proof is high. This post explains both: the geo experiment as a research design, and synthetic control as the technique that makes it trustworthy.
Origin: a question economists could not run an experiment for
The hard problem in geo testing is older than digital advertising, and it was solved by people who were not thinking about ads.
Economists constantly want the effect of something they cannot randomize. What did a tax change do to one state. What did a conflict do to one region. You cannot run a controlled trial on a state: there is one of it, it got the treatment, and there is no identical state in a lab as a control. The naive fix is to pick a similar state and compare, but no two states are similar enough, and the comparison is only as good as the match.
In 2003, Alberto Abadie and Javier Gardeazabal published a paper in the American Economic Review on the economic cost of terrorism in the Basque Country. They needed a counterfactual Basque Country, a version that never had the conflict, and no single Spanish region matched the real one well enough. Their move was the founding idea of the synthetic control method. Instead of choosing one comparison region, they built one: a pool of other regions, each with a weight, combined into a synthetic Basque Country whose pre-conflict economy tracked the real one closely. After the conflict began, the real region and its synthetic twin diverged, and that gap was the estimated cost: per capita GDP fell about 10 percentage points relative to the synthetic control.
Abadie returned to the method in 2010 with Alexis Diamond and Jens Hainmueller, in a Journal of the American Statistical Association paper on California's Proposition 99, the 1988 tobacco control program. A synthetic California, blended from other states, showed what cigarette sales would have done without it. The real state fell well below, by roughly 26 packs per capita per year by 2000. That paper, and the software released with it, put synthetic control into wide use. Susan Athey and Guido Imbens later called it, in a 2017 review of applied econometrics, arguably the most important innovation in policy evaluation in the previous fifteen years.
The crossover into marketing was natural once someone noticed the shapes matched. A region that got a campaign is a treated unit. Regions that did not are a donor pool. What you want is the counterfactual: sales in the treated region had the campaign never run. That is Abadie's problem with ad spend swapped in for a tobacco tax. Google's measurement researchers worked this ground from the mid 2010s, publishing on geo experiments analyzed with a time-based regression and on CausalImpact, a 2015 method by Kay Brodersen and colleagues that uses Bayesian structural time series to predict a market's counterfactual. The technique had left the economics department.
Present: why this survived the privacy collapse intact
Sort measurement methods by one question, do they need to identify a person, and the damage of the last five years stops looking random.
Multi-touch attribution needs the person. It is built on a cross-site record of one user's path, and Apple's App Tracking Transparency, the loss of third-party cookies in Safari and Firefox, and consent rules cut that signal hard. Methods built on individual identity took the damage.
Geo experiments need nothing about a person. The inputs are aggregate: total sales in a region, total spend in a region, by week or by day. There is no identifier to lose, no consent to obtain, no cross-device gap to patch, because the method never looked at a device. When the cookie collapsed, geo testing did not have to adapt. It was already on the other side of the problem. The market has followed: in a July 2025 survey by EMARKETER and TransUnion, 52 percent of US brand and agency marketers already run incrementality experiments, the family geo tests belong to.
There is a second reason the method holds up, about honesty rather than privacy. A geo test is a real experiment. You decide in advance who gets the treatment, change one thing, and read the result. That makes it causal in a way attribution can never be. Attribution watches the people who saw an ad and assigns credit; it cannot separate a customer the ad persuaded from one who was always going to buy and happened to click on the way to checkout. A geo test can, because the control regions show what buying looked like with no campaign at all. The deeper logic of that counterfactual is the subject of incrementality testing without the jargon; geo testing is the form incrementality takes when you cannot withhold ads from individuals.
The matched-market problem, and how synthetic control solves it
Here is where most geo tests quietly go wrong, and where the academic method earns its place.
The obvious way to run a geo test is the matched market test. Pick a test city, find a control city that looks like it, run the campaign in one and not the other, compare. The trouble is that a real city is never a clean twin of another real city. Detroit is not Milwaukee. As Haus argues in a case for synthetic control over matched markets, a single city carries a lot of noise, and one local event, a storm, a plant closure, a regional promotion, can contaminate the whole control and quietly wreck the test.
Synthetic control fixes this by refusing to pick one twin. It builds the control. From a donor pool of many untreated regions, it finds a weighted blend that tracks the test region's pre-campaign history as closely as possible. It might be, in Haus's illustration, 70 percent Milwaukee, 20 percent Philadelphia, 10 percent St. Louis, the specific mixture whose combined sales line lay almost on top of Los Angeles before the campaign. When the campaign launches in the real LA, the synthetic LA keeps predicting the no-campaign baseline, and the gap is the lift.
Two design choices make this trustworthy. The weights are constrained: in the classic method, each weight is non-negative and the weights sum to one. That keeps the synthetic control inside the range of real regions, an interpolation rather than an extrapolation, so it cannot invent a counterfactual no real place could produce. And the fit is judged on a long pre-campaign window. If the synthetic version tracks the real region for a year before the test, that history is the evidence it will keep tracking absent the campaign. Blending also cancels noise: a shock in one donor region is diluted by all the others. Haus reports its synthetic control approach produced lift estimates four times more precise than traditional matched market tests, measured by the width of the confidence interval.
Open source carried this from journals into marketing teams. Meta's GeoLift is the best known: an MIT-licensed R library that runs the whole pipeline, market selection, power analysis, and inference, on synthetic control methods. Its documentation describes it as combining augmented and generalized synthetic control techniques for sturdier estimation on the small samples geo tests usually have. Google open-sourced matched_markets for its time-based regression approach, and the CausalImpact package implements the Bayesian route. A marketer no longer needs an econometrician on staff to run the method an econometrician invented.
How to run a geo test well, in plain language
The method is sound. A test still fails if the design is sloppy. Five decisions carry most of the outcome.
Choose the right geographic unit. In the US, the common choice is the DMA, the designated market area. Triple Whale recommends DMAs as granular enough to give you many possible test and control groups, but large enough that the sales signal stays stable. States give too few units to match well; zip codes are too small and too noisy.
Give it enough pre-period. The synthetic control is only as good as the history it learns from. Triple Whale advises up to a year of geo-level data to capture seasonality, with a pre-period four to five times longer than the test window and free of major promotions or tracking breaks.
Run it long enough. A test that is too short never separates the campaign's effect from ordinary weekly variation. Lifesight suggests four to eight weeks for most tests, longer for slow consideration cycles.
Do the power analysis before you launch, not after. This is the step teams skip and regret. A power calculation simulates the test in advance and asks whether you could realistically detect a lift of the size you expect, given the natural variability of your sales. If weekly sales already swing by 5 percent, a test designed to detect a smaller lift than that will not find a signal even if one exists. Design for at least 80 percent power. GeoLift's power calculators make this a calculation, not a guess.
Pick test markets that represent your business. Synthetic control matches the control to the test; it does not make the test itself representative. Measured makes this point well: synthetic controls alone are not enough, because a test built only from your three strongest cities measures lift in strong cities, not lift you can extrapolate nationally.
Future and impact: a sharp instrument, not the whole toolbox
Geo testing is becoming a default rather than a specialism, and the honest case for it has to include where it stops.
It needs enough distinct geographies. The method works by blending many donor regions, and some markets do not have enough. Recast notes that France has only around a dozen regions, which makes a well-powered geo test there genuinely hard. Below a certain number of units, the donor pool is too thin to build a good synthetic control, and no statistics fix that.
Spillover blurs the line. Geo testing assumes treatment stays in the treated regions and control regions stay clean, and people do not cooperate. They commute across DMA lines, and a video ad gets shared far outside the region that paid for it. Spillover leaks the campaign into the control, inflates the baseline, and pushes the measured lift toward zero. Good design fights this by separating test and control markets geographically, but it never disappears entirely.
It measures one campaign, not the mix. A geo test answers a specific question: did this channel, or this spend change, in this window, cause incremental sales. It does not value your whole media plan. Recast even suggests the honest name for synthetic control might be geo modeling rather than pure experimentation, because a constructed counterfactual still rests on modeling assumptions, even if far weaker ones than attribution makes.
None of that is a reason to skip it. It is a reason to place it correctly. Geo testing is the causal check in a measurement system, not the system itself. Marketing mix modeling values the whole budget over time, attribution gives the fast daily signal, and geo experiments settle the causal questions the other two can only estimate, becoming a ground-truth input that calibrates an MMM. That division of labor is the argument of the measurement triangulation series: each method covers a hole the others leave.
The core idea is the part worth keeping. For two decades, measuring an ad meant following a person, and that era is closing for reasons that will not reverse. Geo experiments measure the thing that is both untrackable and undeniable: whether sales in the places that saw your campaign rose above what comparable places did without it. The person was never the unit that mattered. The lift was.
Council summary
The post argues that geo experiments survived the privacy collapse because they never depended on identifying a person: the unit is a market, the inputs are aggregate sales and spend, and there is no identifier to lose. Its sharpest teaching is the matched-market problem and the synthetic-control fix. A single real city is a noisy, contaminable twin, so rather than picking one comparison market you build a weighted blend of many, with non-negative weights summing to one and a long pre-period to prove the fit. The takeaway is practical: a geo test is a true experiment that yields causal lift, but it needs enough distinct geographies, leaks under spillover, and answers one campaign rather than the whole media plan. Place it as the causal check that calibrates an MMM, not the measurement system itself.
Comments