Media incrementality: how to measure what ROAS alone cannot answer
What media incrementality is, how companies measure the additional effect of campaigns and how to use incremental ROAS in investment decisions.

Media incrementality is the additional result a campaign causes compared with what would have happened without it. The central question is how much the media added: sales, sign-ups or opportunities that would not exist if the investment had not been made. Attributed ROAS relates credited revenue to ad spend, but on its own it does not show that the revenue depended on the ad.
A campaign can reach people who would have bought anyway. It can also influence purchases that never show up in the platform's attribution window. For large advertisers, confusing these two readings can steer budget toward channels that capture existing demand and undervalue efforts that create future demand. Measuring incrementality is how you separate what the media caused from what it merely accompanied.
What media incrementality means in practice
Every incremental measurement compares two scenarios: the world in which the campaign ran and a counterfactual, that is, an estimate of what would have happened without it. The counterfactual is never observed directly. That is why incrementality methods are different ways of building a reliable comparison: a group that did not receive the media, a set of similar markets or a model that projects the expected trend.
The difference between the two scenarios is the incremental effect. It can be expressed in units sold, revenue, margin or another business metric. How much that difference is worth depends on the quality of the comparison: if the groups were already different before the campaign, part of what looks like a media effect is the effect of something else.
Incrementality, ROAS and incremental ROAS
Attributed ROAS divides the revenue an attribution system credits to media by the investment. Incremental ROAS, or iROAS, divides the incremental revenue estimated by a test by the investment that generated it. The difference sounds technical, but it changes the decision. The first answers how much of the recorded revenue passed through the ad according to a crediting rule. The second answers how much of the revenue exists because of the ad.
The two numbers rarely match. Channels that operate close to the purchase, such as branded search or remarketing, can receive a lot of credit in attribution because they sit at the last step of the path, and some of those purchases would have happened anyway. Channels that build demand can show up little in attribution and still produce a meaningful additional effect over longer horizons. iROAS does not replace attributed ROAS in day-to-day monitoring, but it calibrates how you read it.
For investment decisions, it is worth going one step further: compare the incremental effect with contribution margin, not with gross revenue. An iROAS that looks positive may not pay back the investment once product cost, logistics and discounts are taken out.
Start with the decision, not the test format
Define the choice the study should support: keeping an investment, expanding into a region or comparing levels of advertising pressure. Specify the outcome, the population and the time horizon. A vague question produces a test that ends with a number, but not with a decision.
Whenever feasible, random assignment between groups helps build a causal comparison. In other contexts, geographic tests or quasi-experimental methods may be considered, along with their assumptions and limitations. The choice requires statistical, operational and commercial evaluation.
How companies measure media incrementality
There is no single method. The choice depends on the channel, the scale, how much control the company has over delivery and the time available. The most common designs fall into four families.
- User-level control groups: part of the eligible audience is randomly held out from the ads, and the comparison between exposed and unexposed users estimates the effect. Media platforms offer studies of this kind, and it is worth understanding how each one defines the control group and the metric.
- Geographic tests: comparable regions, cities or markets are split into test and control. Media changes in one group and stays stable in the other. They work well when sales can be measured by region, including offline.
- Quasi-experimental methods: when there is no randomization, techniques such as difference in differences or synthetic control build a counterfactual from historical series. They rely on stronger assumptions and require careful analysis.
- Pause tests: stopping or restarting a channel for a defined period. They are simple to run, but vulnerable to seasonality and to events that coincide with the pause.
Beyond these families, marketing mix models estimate contributions from aggregate data and can be calibrated with experiment results. The topic is covered in detail in the article on marketing mix modeling.
Step by step of an incrementality test
- Frame the decision: which investment will be kept, increased or reduced based on the result.
- Choose the primary metric and the source that measures it outside the media platform, such as the sales system, CRM or ERP.
- Define the design: user, region or time period, depending on how much control the company has over delivery.
- Calculate the minimum detectable effect based on historical variability and adjust duration and group size until the test can answer the question.
- Record the plan before starting: groups, dates, metric, analysis rule and what will be done with each result.
- Monitor execution: actual delivery, budget spent, inventory, pricing and parallel initiatives in each group.
- Analyze with uncertainty intervals and document the context so the learning can be reused.
What needs to be agreed before the campaign
- Primary metric: sales, contribution margin, qualified opportunity or another clearly defined outcome.
- Minimum relevant effect: the difference that would justify changing the investment decision.
- Test duration and size: consistent with variability, the purchase cycle and the ability to detect the effect.
- Contamination: cross exposure, promotions, price changes and parallel initiatives that could affect the comparison.
- Analysis rule: what will be concluded in the face of a positive, negative or inconclusive result.
Common mistakes in incremental measurement
- Using the platform's own metric as the only source on the test result.
- Ending the test early because the partial number looks good or bad.
- Choosing control groups for convenience, such as regions that were already underperforming.
- Ignoring the effect that appears after the campaign ends, especially in planned purchase categories.
- Generalizing the result from one investment level to any level. The effect of doubling the budget is not necessarily double the measured effect.
- Testing without deciding in advance what will be done with the result, which turns the study into a report with no consequence.
An inconclusive result also needs interpretation
Failing to detect a difference does not prove the effect is zero. The study may lack precision, run for the wrong length of time or operate under conditions different from those planned. Present uncertainty intervals and explain what the data allow you to state. Avoid turning a point estimate into financial certainty.
Hypothetical example: a geographic test from start to finish
Consider a hypothetical retail chain, selling in stores and through e-commerce, that wants to decide whether to increase its investment in online video. The figures below are illustrative and serve to show the reasoning. They do not represent market or client data.
The chain selects 20 markets with similar sales histories and randomly assigns 10 to receive an additional R$500,000 in investment over six weeks. The other 10 keep the normal plan. The primary metric is total revenue per market, combining stores and e-commerce, taken from the sales system. Before launch, the team records that expansion depends on a return that covers the investment in contribution margin.
At the end, the test markets sell more than expected based on the control markets. The central estimate of incremental revenue is R$1.2 million, with an interval from R$0.4 million to R$2 million. The central iROAS is 2.4. Over the same period, the platform reported a higher attributed ROAS for the campaign, because it credited purchases from customers who already shopped at the stores.
The reading does not end there. With a 30% contribution margin in the category, the central incremental revenue generates R$360,000 in margin, below the R$500,000 invested. The chain concludes that the increase, in this format and at this level, does not pay for itself within the measured horizon. It also records two conditions: there were stockouts in two test markets, and the post-campaign effect was not tracked. With inventory failing in part of the markets, the reading is no longer a simple comparison of advertising pressure. The next round tests a lower investment level and measures four additional weeks.
How to combine the evidence
Attribution helps observe recorded paths. Experiments investigate causal effects under defined conditions. Marketing mix models help analyze aggregate relationships and scenarios, depending on their assumptions. The Meridian documentation presents a modeling framework that can incorporate prior knowledge and experimental results. No method removes the need to assess data quality.
How to decide where to test first
Organize a testing agenda by the value of the decision. There is no need to test every channel at once. Prioritize where there is significant investment, material doubt and the ability to act on what is learned. Keep designs, results and conditions on record to avoid repeating studies without building knowledge.
A practical way to prioritize is to ask three questions for each channel: how much of the budget is in it, how much leadership trusts the current number and whether there is a real reallocation alternative if the test shows a smaller effect than expected. Channels with high investment, low confidence and room to maneuver are natural candidates.
Frequently asked questions about media incrementality
Does a high ROAS mean high incrementality?
Not necessarily. Attributed ROAS measures the revenue a crediting system associates with the ad and can include purchases that would have happened without the campaign. Incrementality measures the difference caused by the media compared with a scenario without it. A channel can have a high ROAS and a small additional effect, or the opposite. They are measures with different interpretations that answer different questions.
What is incremental ROAS (iROAS)?
It is the incremental revenue estimated by a test divided by the media investment that generated it. Unlike attributed ROAS, it relies on a comparison with a control group or with a constructed counterfactual. For budget decisions, it is worth converting it into contribution margin, because a positive iROAS on revenue may not cover the investment once product and operating costs are deducted.
How do companies measure media incrementality?
With experiments and methods that build a reliable counterfactual. The most common designs are user-level control groups, geographic tests across comparable markets, quasi-experimental methods on historical series and channel pause tests. Marketing mix models complement these readings with aggregate data. The choice depends on the decision at stake, the scale and the control over delivery.
Is it possible to measure everything through experiments?
Not always. Constraints of scale, time and execution prevent testing every channel and investment level. Some media do not allow groups to be isolated, and some effects appear over horizons too long for a test. That is why large advertisers combine experiments, attribution and modeling, and organize a testing agenda prioritized by the value of each decision.
How long should an incrementality test run?
Long enough to detect the minimum relevant effect, given the variability of the metric, and to cover the purchase cycle of the category. There is no universal duration. The calculation should be made before launch, using historical data, and the test should not be stopped early because of a partial result. For planned purchases, it is worth tracking for weeks after the campaign ends.
To bring this testing agenda into planning, explore our Media area and the discussion on media budget allocation.