Geo Experimentation: How to Measure the Incremental Impact of Marketing

Marketing is a high-risk, high-reward investment for companies. On the one hand, effective marketing can have a significant impact on a product’s growth. On the other hand, marketing often involves massive budgets, making it critical to invest those resources where they can generate the greatest impact.

As analysts, our goal is to determine whether, and to what extent, marketing investments actually drive incremental impact.

How can we do that? In this blog, we introduce geo experiments, a powerful statistical approach for measuring the incremental impact of marketing campaigns. We’ll walk through what geo testing is, when it should be used, and why it can provide a more reliable measure of causal impact than common attribution models. If you’re looking for a robust way to measure the true impact of your marketing campaigns, this is where you should start.

What is Geo Experimentation?

Geo experimentation is a powerful tool for measuring the incremental impact of a marketing campaign: how much did the campaign actually increase the outcomes we care about, beyond what would have happened without the campaign?

The idea behind a geo experiment is intuitive: we divide geographic regions into treatment regions, where marketing activity is modified (e.g., turning a channel on or off, increasing or decreasing spend, or introducing a new campaign strategy), and control regions, where marketing activity remains unchanged. By comparing how outcomes evolve in the treatment and control regions, we can estimate the incremental impact of the campaign.

For this estimate to be valid, the treatment and control regions must be meaningfully comparable. This can be challenging because geo experiments typically involve a relatively small number of units, while regions can differ substantially in their underlying characteristics and behavior. As a result, simple random assignment may not produce sufficiently comparable groups.

Instead, advanced methods can be used to construct a synthetic control, a weighted combination of control regions that closely matches the pre-treatment behavior of the treated regions. This provides a data-driven approximation of the counterfactual: what would have happened in the treated regions had the campaign not been implemented. The difference between the observed outcome in the treatment regions and this counterfactual represents the estimated incremental impact of the campaign.

Figure 1. Construction of a Synthetic Control. Control regions (Texas, Florida) are weighted to mimic the pre-treatment baseline of the target region (New York). Optimal weighting yields a balanced comparison (top), while improper weighting introduces bias and fails to track the target (bottom).

This incremental impact is the key to evaluating marketing efficiency. For action-based KPIs such as purchases, subscriptions, or installs, we can translate the incremental impact into incremental Cost per Acquisition (iCPA): the additional marketing spend required for each incremental acquisition generated by the campaign. For value-based KPIs such as revenue, we can calculate incremental Return on Ad Spend (iROAS): the incremental revenue generated for each dollar spent on the campaign. Together, these measures connect the causal impact measured by the experiment to the business value of the campaign, helping us determine whether the investment was truly worthwhile.

Why Do We Need Geo Experimentation?

The holy grail of statistical inference is causal inference: determining whether a manipulation actually caused an observed outcome, rather than simply being associated with it. Geo experimentation is particularly valuable in marketing because it provides a rigorous approach to measuring causal impact, while common alternatives such as attribution models rely on observed associations and cannot establish whether marketing activity actually caused an outcome.

Analysts using attribution models may think they are measuring whether a campaign drove an outcome, but in reality, they are measuring whether the campaign was associated with that outcome. To illustrate this point, consider a user who searches for a product, sees a paid ad, and then makes a purchase. An attribution model may assign some or all of the credit for the purchase to the ad because it appeared in the user’s journey.

However, the fact that ad exposure and purchase occurred together does not mean that the ad caused the purchase. This is particularly problematic because marketing exposure is rarely random: users with stronger purchase intent, greater product familiarity, or higher brand awareness may be both more likely to encounter or engage with an ad and more likely to convert. The observed association between exposure and conversion may therefore reflect these underlying differences rather than the causal impact of the advertising itself.

Figure 2. Comparison of attribution models (top) and geo-experiments (bottom). (A) Attribution models assign conversion credit to campaigns based on users’ conversion paths. The right panel illustrates four possible paths and the number of conversions associated with each, while the left panel shows how conversion credit is distributed under first- and last-click attribution. (B) Geo-experiments divide geographic regions into treatment (campaign on) and control (campaign off) groups. The upper panel shows the KPI over time for treatment (red) and control (blue) regions, with the dashed line marking the start of the campaign. The lower panel shows the residuals, calculated as treatment minus control, illustrating the change in the difference between the two groups following campaign launch.

Attribution models can take many forms, from simple first- and last-touch models to more sophisticated multi-touch and probabilistic approaches. While these methods differ in how they distribute credit, they share the same fundamental limitation: they infer marketing impact from observed associations. Establishing causality requires a fundamentally different approach, one that can answer the counterfactual question: What would have happened if users had not been exposed to the campaign?

As discussed in the previous section, geo experimentation is designed to answer this question exactly. By assigning comparable regions to treatment and control and switching on the marketing campaign only in the treatment regions, we can compare their outcomes and estimate the incremental impact of the campaign.

So, why do we need geo experimentation? Because it allows us to move beyond asking “Was the campaign associated with the outcome?” to the question that ultimately matters for decision-making: “How much additional outcome did the marketing activity actually cause?” This is the evidence businesses need to determine whether a campaign truly works and whether additional investment is justified.

When Should You Use Geo Experimentation?

Geo experimentation can be used to establish causality across virtually any marketing channel. However, certain campaign settings create measurement challenges that other methods cannot adequately address, making geo experimentation particularly valuable. The main cases are:

• Brand search campaigns: Whenever a company pays for ads that appear in response to users’ branded search queries (e.g., “IKEA sofa”), it faces an important challenge: which traffic and conversions were generated by the paid ads, and which would have occurred organically anyway? This challenge is particularly relevant for branded search, where users already have some familiarity with, or intent toward, the brand.

While attribution models cannot disentangle this incremental effect from organic demand, a geo-experimentation framework can. By comparing regions with and without paid brand campaigns, geo-experiments can reveal whether paid brand search creates additional demand or merely captures demand that would have materialized through organic search anyway.

Want to see a geo experiment in action? Check out this case study.

• Retargeting campaigns: Companies may use retargeting campaigns to encourage past users to return to their product. As with brand search, these users are already familiar with the product, making it difficult to determine from observed behavior whether they returned because of the campaign or would have returned anyway.

Geo-experiments can address this challenge by manipulating campaign exposure across regions and comparing outcomes between regions with and without retargeting. This makes it possible to estimate the incremental impact rather than simply crediting the campaign with users who were already likely to return.

• Brand awareness campaigns: Top-of-funnel campaigns, such as those on TikTok, YouTube, or Display, are designed to build awareness and influence future behavior rather than to drive immediate conversions.

This makes their impact particularly difficult to measure: there may be no direct or immediate link between ad exposure and a subsequent conversion. This poses a major challenge for attribution models, which rely on observable user journeys to connect marketing touchpoints to outcomes. For brand awareness campaigns, the effect may occur much later, through a different channel, or without any observable connection to the original exposure. Geo-experiments, in contrast, can overcome this limitation by comparing regions exposed to different levels of marketing activity and measuring incremental changes in downstream outcomes.

• Offline campaigns: Marketing activities such as TV, radio, billboards, and other out-of-home advertising are inherently geographic, while individual-level exposure is difficult to observe, control, or attribute. Geo experiments provide a natural framework for measuring their impact by varying campaign exposure across regions and comparing the resulting outcomes. This allows companies to estimate the incremental effect of the campaign at the regional level, even when it is impossible to determine which individual users were exposed to the advertising.

Takeaway: Measure Causal Impact, Not Association

Geo experimentation makes causal inference possible in marketing settings where randomizing individual users is not an option. By randomizing at the regional level and employing sophisticated counterfactual methods like synthetic control, geo testing provides a credible measure of impact for decision-making. This addresses the core limitation of attribution models, which remain correlational and cannot separate marketing influence from selection bias or pre-existing intent. Ultimately, geo experiments let organizations move beyond association, isolate the true effect of their marketing, and decide where to spend budget on evidence rather than correlation.