Your ROAS Dashboard Is Lying: Why Incrementality Testing Is Replacing Platform Attribution

Why does a paid ad campaign show a strong return on ad spend in the platform dashboard while contributing little to actual revenue growth? The answer is that platform-reported ROAS measures attributed conversions, not causal lift, and the two numbers diverge far more than most advertisers assume. Incrementality testing, the practice of comparing customers exposed to an ad against a matched control group that was not, is what actually isolates the revenue a campaign caused rather than the revenue it merely claimed credit for. Bayesian marketing mix models such as Google's Meridian and Meta's Robyn, paired with geo-based holdout experiments, are becoming the standard correction for a measurement layer that Google Ads and Meta Ads Manager alone cannot provide.

The clearest evidence that platform attribution diverges from causal truth comes from a peer-reviewed study published in Marketing Science by Brett Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky. Analyzing 12 large-scale US advertising lift studies at Facebook, encompassing 435 million user-study observations and 1.4 billion ad impressions, the researchers found that standard observational measurement techniques, including last-touch attribution and matching on thousands of behavioral variables, still produced conversion estimates that diverged sharply from the results of true randomized experiments. In several cases, observational methods indicated a campaign was profitable when the randomized experiment showed it was not, and vice versa. The paper's conclusion is blunt: without a randomized control group, no amount of data sophistication reliably recovers the true causal effect of an ad.

Geo-based incrementality testing supplies that missing control group without tracking individual users. The method splits comparable geographic regions into a test group that receives a campaign and a control group that does not, then measures the aggregate difference in conversions between the two. A 2020 AdKDD paper by Joel Barajas, Tom Zidar and Mert Bay, examining Google's Universal App Campaigns, found that a properly designed geo-experiment reduced the advertiser budget required to reach statistical significance by up to 85 percent compared to prior ad hoc testing practices, while also correcting for a documented bias in which last-touch attribution undervalued display and native inventory against experimentally measured incremental conversions. Geo-experiments do not replace platform reporting, they calibrate it, supplying the one number, actual incremental lift, that pixel-based attribution structurally cannot produce.

Google formalized this correction at scale with Meridian, an open-source marketing mix modeling framework built on Bayesian causal inference, maintained publicly on GitHub and documented through Google's own developer portal. Meridian ingests spend and conversion data across online and offline channels and estimates each channel's incremental contribution to revenue, explicitly modeling reach and frequency rather than crediting the last click before a purchase. Google opened Meridian to all advertisers in 2025 and added a no-code Scenario Planner interface in February 2026, extending Bayesian MMM analysis to marketing teams without a data science function, with the current release at version 1.7.0.

Meta's parallel answer is Robyn, an open-source MMM package from Meta Marketing Science distributed under an MIT license and documented in a 2024 arXiv paper by Meta's research team. Robyn combines ridge regression with a multi-objective evolutionary algorithm to search hyperparameters automatically, and critically, it is designed to be calibrated against the results of real-world lift tests and geo-experiments rather than trained on historical spend data alone. That calibration step ties the two halves of the current measurement stack together: geo-experiments supply ground-truth incremental lift for a handful of campaigns, and the MMM extrapolates that causal relationship across the full media plan so a business does not need to run a controlled experiment on every channel every quarter. Google previewed a similar integration in May 2026 with Meridian GeoX, an experimentation layer that feeds geo-experiment results directly into the MMM as a causal prior.

The practical shift for any business running paid ads is that the dashboard number is a starting hypothesis, not a verdict. A campaign reporting a 4x platform ROAS may be capturing demand that would have converted anyway, brand searches, retargeting a warm list, regions with strong organic presence, while a campaign showing a modest 1.5x ROAS may be generating the only genuinely new revenue in the account. Distinguishing the two requires periodically pulling a subset of campaigns or regions into a holdout, measuring the real lift, and feeding that number back into budget allocation. This is the same signal-quality discipline that governs server-side conversion tracking: platform-reported numbers are useful for real-time bid optimization, but they are not a substitute for a periodic causal check on whether the spend is actually growing revenue.

Italian DesAIgns treats incrementality measurement as a required layer of paid advertising management rather than an optional audit. Campaigns are structured to support geo-holdout testing where budget allows, and the resulting lift data is used to sanity-check platform-reported ROAS before it drives further budget decisions. A quick AI visibility check is a useful companion diagnostic, since a business whose organic and paid channels are not clearly separated in the data is often the business most likely to be crediting ad spend for revenue it would have earned regardless.