Key takeaways
- Advanced measurement is now standard on the buy side, but most buyers say it falls short. Between 60% and 75% of senior planning and analytics leaders say attribution, incrementality testing and MMM underperform on rigor, timeliness, trust or efficiency (IAB and BWG Global, State of Data 2026).
- Controlled experiments keep showing that non-experimental ad measurement can be badly wrong. In one Facebook checkout study, a simple exposed versus unexposed comparison showed a 416% lift when the randomized test showed 77% (Gordon, Zettelmeyer, Bhargava and Chapsky, Kellogg and Facebook, 2019).
- Even with more than 5,000 user-level features across 663 Facebook experiments, the best observational methods could not reliably recover the true effect (Gordon, Moakler and Zettelmeyer, Marketing Science 2023).
- Marketers are shifting trust toward MMM and experiments. 27.6% of US marketers rate MMM as the most reliable method, ahead of multi-touch attribution at 19.4% (EMARKETER and TransUnion, 2025).
- Signal loss is real but uneven. Chrome kept third-party cookies and Google retired most Privacy Sandbox APIs in October 2025, while iOS App Tracking Transparency opt-in sits around 35% (Adjust, Q2 2025).
Platform-reported results vs incremental results
Every ad platform reports conversions it can link to its own ads. That is attribution. Incrementality asks a different question: how many of those conversions would not have happened without the ad? The gap between the two answers is the core measurement problem in 2026.
The clearest public evidence comes from studies where researchers had both a randomized test and the observational data side by side.
The first big one used 12 US advertising lift studies at Facebook, covering 435 million user-study observations and 1.4 billion impressions (Gordon, Zettelmeyer, Bhargava and Chapsky, Kellogg and Facebook white paper, published in Marketing Science 2019). The randomized test was the benchmark. Observational methods, the kind of logic many attribution reports rely on, were then asked to estimate the same effect.
The results varied widely. In one checkout study, comparing exposed with unexposed users gave a 416% lift against a true lift of 77%. In another, observational estimates ran from 1,413% to 3,288% when the randomized lift was 2.5%. Across the checkout studies, exact matching missed the true lift by an average of 661 percentage points. The authors also noted that it was hard to predict in advance when the methods would be close and when they would be far off.
A follow-up study scaled this up to 663 experiments, each described by more than 5,000 user and experiment features (Gordon, Moakler and Zettelmeyer, Marketing Science 2023). Even with deep learning models, the methods did not perform well.
| Funnel stage | Median true lift (RCT) | Median absolute error of observational methods |
|---|---|---|
| Upper funnel | 28% | 115 percentage points |
| Mid funnel | 19% | 107 percentage points |
| Lower funnel | 6% | 62 percentage points |
Source: Gordon, Moakler and Zettelmeyer, Marketing Science 2023, as summarised by Kellogg School of Management. The earlier arXiv version of the same paper reports slightly different figures, so treat these as orders of magnitude rather than exact constants.
Search tells a similar story. In large field experiments at eBay, turning off brand keyword ads showed that 99.5% of the lost paid clicks came back through organic search (Blake, Nosko and Tadelis, Econometrica 2015). For non-brand search, the experimental return was minus 63%, while simple regression on the same business suggested returns above 1,400%.
These studies are about specific advertisers and time periods. They do not prove that every platform number is inflated by a fixed multiple. What they do show is that correlation-based credit can be wrong by a large and unpredictable amount, in either direction.
How marketers measure in 2026: adoption of MMM and incrementality testing
Advanced measurement is no longer niche among larger advertisers. The IAB and BWG Global surveyed 430 senior planning and analytics decision-makers at US brands and agencies for State of Data 2026 (published February 2026). Within that sample, 315 used attribution, 292 ran incrementality tests regularly and 287 used MMM.
Other research puts adoption lower. A July 2025 EMARKETER and TransUnion survey found 52% of US brand and agency marketers use incrementality testing and experiments, and 36.2% plan to increase incrementality spending over the next 12 months. The difference likely reflects who was surveyed: the IAB sample is senior planning and analytics staff, while the EMARKETER sample is broader.
A global view is more sobering. Only 32% of marketers measure traditional and digital media spend holistically, falling to 23% in Europe and 29% in Latin America (Nielsen Annual Marketing Report 2025, 1,400 marketers surveyed February to March 2025).
Frequency is also low. The IAB report says regular testers run about 3 to 5 incrementality tests a year today, and attribution and MMM are often refreshed only annually or twice a year. Respondents expect AI to push testing to 11 or more a year and models to monthly runs within one to two years (IAB State of Data 2026).
Free, open tools have lowered the cost of entry. Google opened Meridian, its open-source MMM, to everyone on 29 January 2025, with support for using incrementality experiments as calibration priors (Google, 2025). Meta's open-source Robyn MMM documentation says that applying results from randomized experiments "may improve the accuracy of your marketing mix models dramatically" and recommends recalibrating on a recurring basis (Meta Robyn documentation).
What marketers trust, and where trust breaks down
No single method has won marketers' confidence. When asked which approach is most reliable, 27.6% of US marketers chose MMM, 19.4% chose multi-touch attribution and 18.9% chose unified measurement (EMARKETER and TransUnion, 2025). That means fewer than one in three picked any single method.
The IAB data shows why. Users of each method were asked whether it falls short on a list of core promises.
| Falls short on | Incrementality | Attribution | MMM |
|---|---|---|---|
| Captures performance across all channels | N/A | 72% | 61% |
| Produces reliable or consistent results | 68% | 72% | 67% |
| Delivers quickly enough to inform planning | 73% | 66% | 68% |
| Is transparent and explainable to stakeholders | 69% | 67% | 63% |
| Is trusted by planning and analytics teams | N/A | 61% | 61% |
Source: IAB and BWG Global, State of Data 2026.
MMM has its own blind spots. The same report found that respondents see every channel as underrepresented in their models to some degree, with gaming, commerce media, creator and influencer, and traditional media cited most often (IAB State of Data 2026).
Trust is also a people problem. Marketing analytics influenced only 53% of marketing decisions in a Gartner survey of 377 analytics users (May to June 2022). A third of respondents said decision makers cherry-pick data to support opinions they already hold. In a later Gartner survey of 378 senior marketing leaders (April to May 2024), only 52% said they could prove marketing's value and receive credit for it. Leaders who met regularly with their analytics teams reported a 62% success rate versus 30% for those who met infrequently.
Signal loss: cookies, iOS and consent
The signal loss story has changed since the "cookieless future" headlines.
Chrome did not remove third-party cookies. Google confirmed it would keep its current approach to user cookie choice, and on 17 October 2025 announced it was retiring most Privacy Sandbox technologies, including the Attribution Reporting API, Topics and Protected Audience, citing low adoption (Google Privacy Sandbox blog, 2025). Google said it will continue work on an interoperable attribution standard through the W3C.
So for Chrome users, the cookie is still there. That does not mean signals are complete. Safari and Firefox restrict third-party cookies, consent banners reduce what can be collected, and users switch devices.
On iOS, App Tracking Transparency remains the larger gap. The industry-wide opt-in rate, among users shown the prompt, was 35% in Q2 2025, up from 34.5% in Q2 2024 and 34% in Q2 2023 (Adjust, 2025). Rates vary widely by category and country. Sports games reached 50% and education apps 14%. Brazil led at 50%, while Canada and Australia were at 29% and 27%.
The practical effect is that platform attribution increasingly relies on modelled conversions to fill gaps. That is not wrong in itself, but it makes the platform's number an estimate, which is exactly why independent checks matter.
Where the research disagrees
It is worth being clear about what is not settled.
- How widespread incrementality testing is. The IAB sample suggests most senior buy-side teams test regularly. EMARKETER and TransUnion put the figure at about half of US marketers. Both can be true for different populations.
- Whether MMM is the answer. MMM ranks highest on reliability in the EMARKETER and TransUnion data, yet 61% to 69% of MMM users in the IAB study say it falls short on coverage, trust or timeliness.
- How big the gap is for your account. The academic studies show large errors on average, but the size and direction vary by campaign. The Facebook researchers explicitly say you cannot predict the error in advance without an experiment.
- Whether cookies still matter. Chrome keeping third-party cookies reduces one source of loss. iOS opt-in at around a third and consent requirements keep others in place.
What to do with this
These are practical recommendations based on the research above, not results from any single study.
- Fix the data before the model. Every method depends on clean conversion data. Check that tags fire once, that server-side events deduplicate with browser events, and that consent settings are working as intended. An analytics and tracking audit is the usual starting point.
- Treat platform ROAS as a directional signal. Use it for day-to-day optimisation inside a platform, not as proof of total business impact. Our guide to GA4 attribution in 2026 explains how GA4's models assign credit and where they differ from ad platforms.
- Anchor on business-level numbers. Track total revenue against total marketing spend and new customer acquisition cost alongside channel reports. A ROAS calculator helps you compare platform ROAS with the return your margins actually need.
- Run at least one holdout test on your biggest line. Start with the channel that takes the most budget, or with brand search, where the eBay research found most paid clicks were replaced by organic. Meta Conversion Lift, Google's experiments and geo tests are all options.
- Calibrate MMM with experiments if you use it. Both Meridian and Robyn are built to take experiment results as inputs. An uncalibrated model is one more estimate, not a referee.
- Agree the rules in advance. Decide which number wins for which decision: platform data for creative tests, experiments for channel budgets, MMM for annual planning. Gartner's research suggests regular contact between marketing leaders and analytics teams matters as much as the tools.
- Connect measurement to conversion work. Better attribution is only useful if it changes what you do on site and in campaigns. Pairing measurement with CRO and analytics work closes that loop.
Sources
- IAB and BWG Global, State of Data 2026: The AI-Powered Measurement Transformation, February 2026
- Gordon, Zettelmeyer, Bhargava and Chapsky, A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook, Kellogg School of Management white paper, published in Marketing Science, 2019
- Gordon, Moakler and Zettelmeyer, Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement, Marketing Science, 2023 (arXiv preprint 2022)
- Blake, Nosko and Tadelis, Consumer Heterogeneity and Paid Search Effectiveness: A Large Scale Field Experiment, Econometrica, 2015 (NBER working paper 2014)
- EMARKETER, FAQ on incrementality: How to prove your ads actually work in 2026, April 2026, citing EMARKETER and TransUnion survey, July 2025
- Nielsen, 2025 Annual Marketing Report, 2025
- Gartner, Marketing Analytics Survey press release, September 2022
- Gartner, survey of senior marketing leaders on proving value, press release via Business Wire, September 2024
- Google, Meridian is now available to everyone, Google blog, January 2025; Meridian developer documentation
- Meta, Robyn open-source MMM documentation, calibration section
- Google, Update on Plans for Privacy Sandbox Technologies, October 2025
- Adjust, ATT opt-in rates: 2025 data and benchmarks, 2025
FAQ
Are platform-reported ROAS numbers wrong?
They are not necessarily wrong, but they measure attribution, not incrementality. Experiments at Facebook and eBay found that correlation-based estimates can overstate the true effect by large and unpredictable amounts, so platform ROAS works best as a directional signal checked against holdout tests.
Is marketing mix modeling better than multi-touch attribution?
More marketers trust it. 27.6% of US marketers rated MMM the most reliable method versus 19.4% for multi-touch attribution (EMARKETER and TransUnion, 2025). But most MMM users in the IAB State of Data 2026 study still say it falls short on coverage, trust or speed, so calibrate it with experiments.
Did Google remove third-party cookies from Chrome?
No. Google kept third-party cookies in Chrome under its existing user choice approach and in October 2025 announced the retirement of most Privacy Sandbox APIs, including Topics and the Attribution Reporting API, citing low adoption.
How often should a business run incrementality tests?
There is no single rule. Regular testers in the IAB State of Data 2026 study run about 3 to 5 tests a year and expect to reach 11 or more. A practical start is one holdout test on your largest channel or on brand search.