The ROAS Illusion: Why Your Ad Platform's Return Number Isn't the Real One

The Percentage That Doesn't Exist

Search for the gap between platform-reported ROAS and true incremental return and you will find a version of the same claim everywhere: platforms overstate results by 20 to 60 percent. It is a clean, quotable number, and it is not supported by any independent, cross-platform, peer-reviewed source. Every rigorous piece of evidence we could locate and verify shows something more complicated and more useful: the bias is channel- and funnel-stage-dependent, running heavily in one direction for some spend and, credibly, in the opposite direction for other spend. There is no single haircut you can apply to a reported ROAS number and trust the result.

That is not a comfortable finding for anyone selling a simple fix, and it is worth sitting with. Two of the sources that came closest to a flat 40 to 60 percent figure turned out, on inspection, to be a vendor's explicitly hypothetical sales example and an unattributable aggregation of ad-tech blog commentary with no named study behind it. Both were set aside. What survives is smaller, older in places, and more honest.

There is no single haircut you can apply to a reported ROAS number and trust the result.

What the One Peer-Reviewed Study Actually Found

The strongest evidence available is also the oldest and the least commercially motivated. In 2019, researchers working with Facebook published a study in Marketing Science, a peer-reviewed journal of the Institute for Operations Research and the Management Sciences, examining 15 large-scale field experiments covering more than 500 million user-experiment observations and 1.6 billion ad impressions. They compared the observational and attribution methods marketers normally use against results from true randomized controlled experiments, the gold standard for isolating what an ad actually caused. The finding was not a clean overstatement number. Commonly used attribution methods generally overestimated advertising effectiveness relative to the randomized-experiment truth, but in some cases they significantly underestimated it.

That single sentence is the most academically credible fact in this entire subject, and it already contradicts the flat-overstatement story before a single vendor benchmark enters the picture. The direction of the bias is not fixed. It depends on what is being measured and how.

The $200 Million Real-World Test

Before any of the modern incrementality-testing vendors existed, Procter & Gamble ran the largest natural experiment on record for this exact question. Over nine months in 2017 and 2018, Chief Brand Officer Marc Pritchard cut roughly $200 million of digital ad spend, about 20 to 50 percent of spend on some digital channels, largely in response to concerns about ad fraud, viewability, and walled-garden measurement. The result, corroborated across multiple trade publications at the time, was no discernible negative impact on P&G's business performance. Reach actually increased by about 10 percent, because the spend that was cut had been buying largely redundant or unmeasurable impressions rather than incremental audience.

This is not a controlled experiment with a clean causal design the way the 2019 Marketing Science study is. It is a real company making a real, large cut and living with the consequences in public. But it points the same direction as the peer-reviewed research: a meaningful share of reported digital ad performance was never doing the work its reported ROAS implied.

Where the Money Actually Goes Before It Reaches Anyone

Part of why platform-reported numbers are so unreliable is structural, not statistical. In December 2023, the Association of National Advertisers published a log-level analysis of 21 major advertisers, $123 million in ad spend and 35.5 billion impressions, tracing money through 12 participating supply-chain companies from the moment it entered a demand-side platform to the moment it reached a consumer. The finding: only 36 cents of every dollar effectively reached the consumer as working media. Roughly 29 percent of spend was consumed by transaction costs, the fees that sit between the advertiser and the audience, and about 35 percent went to inventory that was non-viewable, invalid, or otherwise unmeasurable.

That is not an incrementality problem, it is a supply-chain problem, but it compounds the measurement problem. A platform reporting ROAS on the dollar it received is reporting on a dollar that, on average across this study's sample, had already lost nearly two-thirds of its working power before it ever reached a person capable of converting.

Only 36 cents of every dollar effectively reached the consumer as working media.

The Reliable Distortion: Branded Search and Retargeting

Where the evidence does point toward a consistent, sizable overstatement is at the bottom of the funnel. Across 225 geo-based incrementality tests run between August 2024 and December 2025, the incrementality-testing platform Stella reported an overall median incremental ROAS of 2.31x, with branded search and retargeting showing inflation of 5 to 10 times against platform-reported numbers. In the channel-level breakdown from the same panel, Google Search branded spend came in at a median incremental ROAS of just 0.70x, meaning that on an incremental basis it was likely losing money even though the platform's own attributed ROAS almost certainly looked strong.

This is the one piece of quantified evidence in this entire subject that should be read carefully rather than dismissed: it comes from a single vendor that sells incrementality testing and has a direct commercial incentive to show a large, fixable gap, so it is labeled contested here rather than established. But the mechanism it describes, that branded search and retargeting largely capture customers who were already going to convert, is well understood and consistent with the peer-reviewed evidence above. Treat the specific multiplier as directional, not as gospel.

When the Platform Undercounts You

The more counterintuitive finding, and the one that most directly breaks the flat-overstatement narrative, comes from a different incrementality vendor looking at Meta specifically. Across 640 Meta geo-holdout experiments, Haus found that for every $100 in platform-attributed DTC revenue under a 7-day click attribution window, Meta actually generated $115 in incremental revenue in aggregate, a 15 percent platform understatement. In a separate panel of 46 Meta incrementality studies among ecommerce brands with $15 million to $100 million in annual revenue, Stella found Meta delivering 21 percent more incremental revenue than the platform itself reported.

Both findings come with important texture. Haus found that Advantage+ and other automated campaign types tended to over-report relative to manual campaigns, while mid- and upper-funnel campaigns tended to under-report, especially when measured on a DTC-only basis. Fully 32 percent of Meta's total measured incremental impact in that panel lifted retail and marketplace sales that a DTC-only ROAS calculation would miss entirely. Both of these findings are also vendor-sourced and commercially interested, so they carry the same contested label as the bottom-funnel numbers above. But they are directly relevant because they show the bias running in the opposite direction from the popular assumption, on the same platform, sometimes in the same campaign portfolio.

For every $100 in platform-attributed revenue, Meta actually generated $115 in incremental revenue, a 15 percent understatement.

Believing It's Broken, Still Running It the Same Way

Marketers are not naive about any of this in the abstract. A 2026 survey of 210 senior marketing leaders across brands and agencies found that 91 percent believe platform-reported results are overstated to some degree, and a third believe attribution and optimization lag wastes more than 26 percent of budget. This particular survey should be read with a caveat: live verification found it was sponsored content commissioned by Affinity Solutions through its own Outcomes Marketing Council, not independent research, and no sampling methodology or margin of error was disclosed. It is included here as a labeled, vendor-commissioned data point on marketer sentiment, not as validated measurement of the actual overstatement size.

What is more telling, and comes from an independent 2020 eMarketer analysis, is the gap between that stated distrust and actual behavior. Only 40 percent of US marketers and agencies required their media partners to use buyer-approved measurement tools, and major walled-garden platforms typically did not comply even when asked. Forty-four percent of marketing professionals surveyed separately rated social media among the hardest channels to attribute sales revenue to, and 63 percent said insufficient media-quality transparency would most likely affect their spend on Facebook specifically. The distrust is real and widespread. The day-to-day optimization decisions still mostly run on the number the platform hands you, because building or buying an independent measurement capability is expensive, and the dashboard is free.

Why This Hits Small and Midsize Businesses Hardest

Every rigorous source in this study, the 2019 Marketing Science RCT, the geo-holdout panels, the incrementality benchmarks, has one thing in common: it required spend and conversion volume far beyond what most small and midsize advertisers generate. Geo-holdout testing needs enough geographic markets and enough conversions per market to detect a statistically meaningful difference. Randomized controlled experiments at the scale of the Gordon et al. study require Facebook-level infrastructure. None of the sources reviewed for this study included an SMB-specific incrementality dataset for accounts spending in the sub-$2,500 to $10,000 monthly range that describes most small businesses.

That is a reasonable inference from what incrementality testing structurally requires, not a measured finding about SMB accounts specifically. A small business cannot run its own version of the P&G experiment or a Meta geo-holdout panel. It is left choosing between trusting a platform number that the best available evidence says is unreliable in an unpredictable direction, or spending money it does not have on testing infrastructure built for brands ten times its size. That asymmetry, not the size of the gap itself, is the real finding for a business this size.

What to Do With a Number You Can't Fully Trust

None of this means platform reporting is useless, and it does not mean the answer is to ignore ROAS entirely. It means treating a platform's self-reported number as one signal among several, weighted by what is now reasonably well established: be most skeptical of branded search and retargeting ROAS, since the mechanism for inflation there, capturing customers who were already converting, is well understood even where the exact multiplier is vendor-sourced. Be slower to cut upper-funnel, prospecting, or automated campaigns purely on a weak reported ROAS number, since at least one rigorous panel found that exact category understated on a large sample.

The deeper fix is not a better attribution model from inside any single platform, it is an independent view of where a market's attention actually sits and how it is moving, measured from outside the walled garden that has every incentive to report its own performance favorably. That is a different kind of map than a ROAS dashboard, and it is the one worth building next.