The synthetic control method for advertising measurement answers a question every high-consideration brand eventually faces: what would revenue have looked like in the markets where we advertised if we had not advertised there at all? A clean randomized test is the gold standard, but luxury brands rarely get one. Media is bought by region, launches happen in a handful of markets, and the number of serious buyers in any one area is small. Synthetic control builds a weighted "twin" of the treated markets from untreated ones, then measures the gap after launch. Used well, it turns an awkward regional rollout into credible evidence of incremental revenue.
This guide explains how the method works, when it beats a standard geo test, what data it needs, what results look like in practice and the mistakes that quietly ruin a read. It is written for marketing executives at brands where a customer is worth more than $5,000 and the decision takes longer than a month.
What is the synthetic control method for advertising measurement?
Synthetic control is a statistical technique from economics. Alberto Abadie and colleagues developed it to estimate the effect of policies applied to one region, most famously in a 2010 study of a state tobacco control program published in the Journal of the American Statistical Association. The idea transfers neatly to media.
You start with a group of treated markets, the regions where a CTV, audio or out-of-home campaign ran. You then take a pool of untreated "donor" markets and search for weights that, combined, reproduce the treated markets' pre-campaign history as closely as possible: weekly inquiries, site sessions from qualified traffic, booked consultations or revenue. If 40 percent of one donor market plus 35 percent of a second plus 25 percent of a third tracked your treated markets almost perfectly for the previous 26 weeks, that blend is your synthetic control. After launch, the synthetic version shows the expected path without the campaign. The difference between actual and synthetic is the estimated lift.
The appeal is transparency. Unlike a black-box attribution model, you can show a CFO exactly which markets make up the comparison and how closely they matched before the money was spent.
How is synthetic control different from a standard geo test?
A classic geo experiment randomly assigns markets to test and control before launch. Synthetic control is usually applied when assignment was not random, or when there are too few treated markets for a matched-pair design to work.
| Approach | When it fits | Key strength | Key weakness |
|---|---|---|---|
| Randomized geo test | You control launch markets and have 20+ comparable regions | Strongest causal claim | Requires planning before the buy; needs many markets |
| Matched-market test | A few test markets paired with look-alike controls | Simple to explain | One bad match distorts the result |
| Synthetic control | One to a handful of treated markets, chosen for business reasons | Builds a better comparison than any single market | Needs long, stable pre-period data |
| Bayesian time series (CausalImpact style) | Single treated series with strong covariates | Gives probability ranges and handles seasonality | Sensitive to which covariates you include |
| Audience holdout | Addressable media such as CTV with household IDs | Measures at household level | Not possible for most out-of-home or broadcast buys |
For most luxury programs the right answer is a combination. A household holdout test measures addressable CTV directly, while synthetic control reads the market-level effect of channels that cannot be held out, such as billboards near a flagship, regional podcast sponsorships or a sports partnership.
When should a luxury brand use synthetic control?
The method earns its place in five situations we see often:
- A regional launch chosen for business reasons. A private club opening a second location, a developer selling a new tower or an aviation operator adding a base will advertise where the product is. Randomization was never on the table.
- High-cost channels with few markets. Premium out-of-home, airport media and regional sports sponsorships are bought market by market. Synthetic control is often the only credible read.
- Small conversion counts. A wealth manager booking 30 qualified consultations a month cannot split the country into 50 test cells. Pooling donor markets gives the comparison enough weight to be useful.
- Retroactive questions. A campaign already ran and the board wants to know if it worked. As long as clean pre-period data exists, synthetic control can still produce an honest estimate.
- Validating an attribution or mix model. A synthetic control read on one market is a useful calibration point for a marketing mix model, which otherwise relies on historical correlation.
It is a poor fit when the treated market is unlike anything in the donor pool. If your only flagship is the largest affluent market in the country, no weighted mix of smaller markets will reproduce it reliably.
What data does the synthetic control method for advertising measurement need?
Data quality decides the result more than the algorithm does. Before committing to a synthetic control design, confirm you have the following.
- A long pre-period. Aim for at least 26 weekly observations before launch, and 52 if your business is seasonal. Abadie's own guidance, summarized in a 2021 Journal of Economic Literature review, stresses that a good pre-period fit over a long window is what makes the estimate believable.
- A consistent outcome metric. Use the metric closest to revenue that you can measure by market every week: qualified inquiries, booked appointments, test drives, membership applications or closed revenue. Avoid metrics whose definition changed during the window.
- A clean donor pool. Donor markets must not receive the campaign, must not be exposed through spillover national media and must not have their own unusual events, such as a new showroom or a competitor's exit.
- Geography that matches the media. Define markets the way the media was bought, usually DMAs for CTV and audio, and match your CRM records to the same boundaries using billing or property ZIP codes.
- Covariates, if available. Search volume for the category, macro data such as local equity market exposure, or historic seasonality can strengthen the model, but each must be unaffected by the campaign itself.
If CRM records are the outcome, connect them first. Our guide to CTV CRM matchback covers how to join household exposure and revenue records, and the same discipline applies to market-level reads.
How do you run a synthetic control test, step by step?
Here is the workflow we use when designing a read for a regional campaign.
- Write the question and the decision. For example: "Did eight weeks of premium CTV and audio in three markets increase qualified consultations enough to justify national expansion?" Agree on the decision threshold before you see results.
- Fix the pre-period and the test period. Lock the dates in writing. Moving them after the fact is the most common way analysts talk themselves into a result.
- Build the donor pool. Remove markets with contaminating events. Keep 15 to 40 donors if you can.
- Fit the weights on the pre-period only. Tools such as Meta's open-source GeoLift package, which uses an augmented synthetic control approach, and Google's CausalImpact library automate this step. Inspect the fit: the synthetic line should track the treated line week by week, not just on average.
- Run a power analysis before launch. Simulate fake campaigns in the pre-period to see how large an effect you could actually detect. If the minimum detectable lift is 40 percent and you expect 10, change the design or the budget.
- Launch and leave it alone. Do not change creative, budgets or targeting mid-test unless you document it.
- Measure the gap and test its significance. Placebo tests, where you pretend each donor market was treated and see how often a gap as large as yours appears by chance, are the standard check.
- Translate lift into money. Convert incremental consultations or inquiries into expected revenue using your close rate and average order value, then compare that with spend.
What results are realistic for luxury and high-consideration brands?
Results vary by channel, creative and offer, so treat the figures below as Stillwater planning ranges for scoping, not promises.
| Metric | Stillwater planning range | Notes |
|---|---|---|
| Pre-period fit error (MAPE) | Under 10 to 15 percent | Above 20 percent, the read is usually not trustworthy |
| Minimum detectable lift, 3 to 5 treated markets | 8 to 20 percent | Lower with longer tests and stronger covariates |
| Test length | 6 to 12 weeks, plus a 4 to 8 week post window | Longer sales cycles need a longer post window |
| Observed lift in qualified inquiries, premium CTV plus audio | 5 to 25 percent | Wide range; creative and frequency drive most of the spread |
| Share of donor weight in the top three donors | 50 to 80 percent | Very concentrated weights deserve extra scrutiny |
Two numbers matter most to a finance team: the confidence interval and the cost per incremental outcome. A 15 percent lift with an interval of 2 to 28 percent is a real finding but an uncertain one. Report it that way.
What are the most common mistakes with the synthetic control method for advertising measurement?
The method is only as honest as the person running it. These errors appear again and again.
- Too short a pre-period. Eight weeks of history will fit almost anything. The fit looks great and means nothing.
- Contaminated donors. National CTV, paid social or a PR moment that reaches donor markets shrinks the measured gap and makes good media look weak.
- Choosing the outcome after the fact. If you measured inquiries, site visits and revenue and only report the one that moved, you have not run a test.
- Ignoring the sales cycle. In private aviation or luxury real estate, much of the revenue arrives months after exposure. Our guide to how long CTV advertising takes to work explains why the post window has to match the buying cycle.
- Treating a single read as permanent truth. Lift depends on creative, season and competitive pressure. Re-test at least once a year, and every time the channel mix changes materially.
- Over-precise reporting. Presenting "17.3 percent lift" without an interval invites the wrong decision. Round, show the range and show the placebo distribution.
How does synthetic control fit with incrementality testing and attribution?
Think of measurement as a stack. Platform attribution tells you what each channel claims. Household holdouts and geo tests, including synthetic control, tell you what actually changed. A mix model ties everything together over years of data. Synthetic control is most valuable as the bridge: it measures channels that cannot be randomized and provides calibration points for the model.
For a fuller view of how these pieces work together, see our guides to incrementality testing for luxury brands and marketing mix modeling versus incrementality testing.
What does a good synthetic control readout include?
Whether the analysis comes from an agency, a vendor or an in-house team, ask for these elements before you accept the conclusion:
- the list of donor markets and their weights;
- a chart of treated versus synthetic for the full pre-period and test period;
- the pre-period fit error;
- the placebo test results, ideally shown as a chart of all donor gaps;
- the minimum detectable effect from the pre-launch power analysis;
- the outcome metric definition and how it was matched to geography;
- the estimated incremental outcomes, revenue and cost per incremental outcome, each with an interval;
- a written note of anything that changed during the test.
A team that cannot produce these is asking you to trust a number, not a method.
A worked example
Consider a luxury home furnishings brand that ran premium CTV and streaming audio in four markets for ten weeks. Its donor pool held 28 markets with clean data. Using 52 weeks of booked design consultations as the outcome, the synthetic control matched the treated markets with a pre-period error of about 7 percent, built mostly from five donors.
During the test, the treated markets booked roughly 640 consultations against a synthetic expectation of about 545, an estimated 95 incremental consultations, or roughly 17 percent lift. Placebo tests on the donors produced a gap that large only once in 28 runs. With a close rate of 22 percent and an average order of $18,000, the incremental consultations imply about $375,000 in expected revenue against $210,000 of media. That is a strong, defensible case for expansion, presented with its uncertainty. (These figures are illustrative, built from Stillwater planning ranges, not a client result.)
Should you use synthetic control for your next campaign?
If your next campaign launches in a handful of markets for business reasons, uses channels that cannot be held out at household level, or needs to satisfy a skeptical finance team, the synthetic control method for advertising measurement deserves a place in the plan. Design it before launch, protect the donor pool, match the post window to your sales cycle and report ranges rather than point estimates.
Stillwater Media designs measurement into every engagement from the first week, combining household holdouts, geo designs and synthetic control reads so clients can see which media created revenue and which only claimed it.
Ready to measure what your media really earns? Apply to work with Stillwater Media. We take on a limited number of engagements each quarter.
About the author: Stillwater Media is a selective performance media agency for luxury and high-consideration brands, specializing in premium CTV, programmatic, affluent audience engineering and incrementality testing. Signal. Strategy. Scale.
Frequently Asked Questions
What is the synthetic control method in advertising?
It is a way to measure campaign lift when only a few markets receive the media. It builds a weighted blend of untreated donor markets that matches the treated markets' history before launch, then treats the gap after launch as the estimated effect.
How is synthetic control different from a geo test?
A randomized geo test assigns markets to test and control before launch. Synthetic control is used when treated markets were chosen for business reasons or are too few for matched pairs, and it builds a better comparison than any single control market.
How much pre-period data does synthetic control need?
Aim for at least 26 weeks of weekly data before launch, and 52 weeks for seasonal businesses. A long, close pre-period fit is what makes the estimate believable.
What tools run synthetic control for marketing?
Meta's open-source GeoLift package and Google's CausalImpact library are widely used. Many analytics teams also run synthetic control directly in R or Python.
What lift can luxury brands expect from a synthetic control read?
It depends on channel, creative and offer. As a Stillwater planning range, premium CTV plus audio often shows 5 to 25 percent lift in qualified inquiries, and every result should be reported with its confidence interval.


