Incrementality Testing: How to Prove Your Ads Actually Work

Attribution reports tell you which ads were present before a sale. They do not tell you whether the sale would have happened anyway. That gap is why incrementality testing has moved from specialist practice to the mainstream. According to a July 2025 EMARKETER and TransUnion survey, 52% of US brand and agency marketers already use incrementality testing and experiments, and 36.2% plan to increase spending on it over the following 12 months ¹.

The cost of entry has also fallen. In November 2025 Google said that an experiment which "once might have cost upwards of $100,000" can now be run for $5,000 ². This guide covers the main types of test, what the platforms offer, how to design a test properly and where tests most often go wrong.

What incrementality means

Incrementality is the difference between what happened with your marketing and what would have happened without it. EMARKETER describes incrementality testing as using "causal-based logic to weigh whether results were caused by marketing or not" ³. Google's definition is similar: knowing what happens because of your marketing, "and what would not have happened otherwise" ².

The principle is the same as a medical trial. One group is exposed to the advertising, a comparable group is not, and the difference in outcomes is the incremental effect. The design challenge is making sure the two groups really are comparable.

The three main test types

Test type	How it works	Best for
User-level holdout (conversion lift)	Users are randomly split; the control group is prevented from seeing the ads	Single digital platforms with logged-in users
Geo experiment	Regions are assigned to test or control, and media is changed only in test regions	TV, out-of-home, retail media, cross-channel effects
Campaign or channel holdout	A campaign or channel is paused for a period or a share of the audience	Checking whether always-on spend, such as brand search, is adding value

User-level tests are statistically efficient but only work inside a platform that can control who sees ads. Geo experiments work for almost any channel. EMARKETER recommends them for TV, out-of-home and retail media, where user-level holdouts are impractical ¹. They need enough regions with comparable behaviour, which can be a constraint in smaller markets.

What Google and Meta offer

Google has made incrementality testing far more accessible within Google Ads. Its November 2025 update lowered the practical minimum spend to around $5,000, allowed advertisers to choose a custom test size and view results at their preferred level of statistical confidence, and said results would be delivered "up to 50% more frequently" ². Google also recommends using experiment results to calibrate marketing mix models ².

For tests that span publishers, Google has built Meridian GeoX, an open-source library for geo experiments. In June 2026, Think with Google described it as a way to run "transparent, cost-effective, publisher-agnostic experiments" to calibrate its Meridian marketing mix model ⁴. Google's documentation says GeoX supports holdback, go-dark and heavy-up designs, and can compare several treatments against a shared control in one study to reduce cost and time ⁵.

Meta's open-source mix model, Robyn, is designed to take incrementality experiment results as a calibration input. Calibration adds an error measure against the experiment results that the model's optimiser tries to minimise alongside its usual fit metrics ⁶. This makes lift tests the anchor for Meta's own long-term measurement approach, not just a one-off check.

One caution applies to any platform-run test: it measures that platform's ads only, and the platform both runs the test and sells the media. Geo experiments run independently, or with open-source tools, give a cross-platform check.

How to design a good test

A well-designed incrementality test follows a clear sequence.

1. Write a specific question. For example: "Does our paid social prospecting drive incremental online sales in the UAE?" A vague question produces an unusable answer.
2. Choose one primary KPI. Use a metric close to business value, such as revenue or qualified leads, and agree it before the test starts.
3. Pick the right design. Use a user-level lift test inside one platform, or a geo test for cross-channel or offline media.
4. Size the test. Run a power analysis to estimate how large the test and control groups must be, and how long the test must run, to detect a realistic effect.
5. Set the duration. EMARKETER's guidance is to run tests for at least three to four weeks with properly sized holdouts ¹.
6. Include a post-test window. Effects can continue after media stops. In geo experiments for TikTok campaigns run by measurement firm Haus, the post-treatment period showed an additional 68% lift in the primary KPI ¹.
7. Agree in advance how results will be used. Decide what budget change a positive, negative or inconclusive result will trigger.

Common pitfalls

• Too little power. Small tests produce wide confidence ranges that cannot distinguish a real effect from noise. This is the most common reason for inconclusive results.
• Contaminated control groups. If control regions still receive the media, through national TV or users travelling between regions, the measured lift will be understated.
• Changing other things mid-test. Price changes, promotions or new creative in only one group distort the comparison.
• Ignoring the cost of holdouts. Holding back media means giving up some short-term revenue, which needs to be budgeted for ¹.
• Comparing incremental ROAS with platform ROAS. Incremental return is usually lower than reported return because it sets a higher bar ¹. That is expected, not a failure.
• Treating one test as permanent truth. Results change with seasons, creative and competition. The IAB's 2026 research found incrementality testing shifting from a few tests a year towards always-on experimentation ⁷.

Reliability is also a concern for the industry as a whole. In a retail media study by Skai and the Path to Purchase Institute, cited by EMARKETER, 44% of respondents named doubts about the accuracy or reliability of results as the top barrier to incrementality testing, followed by difficulty applying it across ad types and retailers (43%) and limited tools (41%) ¹. Careful design addresses most of these doubts.

Key takeaways

• Incrementality testing measures what your marketing caused, not just what it touched ²³.
• Over half of US marketers already run experiments, and adoption is growing ¹.
• Google has cut the cost of in-platform tests to around $5,000 and offers open-source GeoX for geo experiments ²⁵.
• Good tests start with a specific question, a single KPI, a power analysis and a pre-agreed decision rule.
• Use results to calibrate marketing mix models and plan a regular testing calendar ⁶⁷.

What this means for your brand

Begin with the channel where you have the most doubt and the most spend, often brand search, retargeting or paid social prospecting. Run one well-powered test before attempting several small ones. For channels without click-based tracking, use geo experiments, and for brands across EMEA, choose test regions carefully so they are genuinely comparable. Feed every result into your planning and mix model, and repeat important tests at least once a year.

Sources

1. FAQ on incrementality: How to prove your ads actually work in 2026 - EMARKETER, April 2026
2. Strengthen media measurement and ROI clarity with incrementality testing improvements - Google Ads, November 2025
3. MMM, incrementality, and other measurement trends that will define 2026 - EMARKETER, November 2025
4. Marketing mix modelling with Meridian - Think with Google, June 2026
5. Meridian GeoX - Google for Developers, 2026
6. How Facebook Robyn Works - Recast, October 2024
7. 75% of marketers say their measurement systems are falling short - MarTech, February 2026

Figures and product details are as reported by the sources listed, at the time of writing. Platforms change quickly, so check the latest documentation before acting on any single number.