Incrementality is the share of business outcomes that advertising actually caused. It rests on one question - what would have happened if the campaign had not run - and every complication in the discipline follows from the fact that the answer can never be observed. A conversion recorded by an ad platform is a fact. Whether it required the ad is a counterfactual, and counterfactuals have to be constructed, normally by withholding advertising from a comparable group and measuring the difference.

The concept exists because attribution counts rather than tests. Last-click models, data-driven models and view-through windows distribute credit among touchpoints that preceded a purchase. None establishes that removing a touchpoint would have removed the purchase. The IAB and IAB Europe formalized the distinction in a joint paper dated November 3, 2025, defining incrementality as the additional business outcomes directly driven by a campaign or tactic compared with what would have occurred without marketing activity. According to that document, attribution and return on ad spend show what happened, not whether marketing caused it.

The arithmetic

A study splits a population into a treatment group, eligible to see the advertising, and a control group, which is not. Absolute lift is the difference in outcomes between the two, scaled to equal group sizes; relative lift expresses that difference against the control baseline; incremental return on ad spend, written iROAS, divides incremental revenue by the spend that produced it.

The result rarely matches platform dashboards, and the direction of the mismatch is contested. Reported return normally exceeds incremental return, because ads placed in front of people already heading toward a purchase collect credit for a decision already made. Retargeting and branded search are the standard illustrations. The gap can also run the other way: measurement firm Measured reported in May 2026 that platform reporting can understate the incremental value of YouTube and Demand Gen by as much as four times.

Test designs

The IAB paper sorts methods into four families graded by causal strength. Experiment-based approaches, covering randomized control tests, holdouts, ghost ads and matched markets, are rated strong but narrow, usually confined to one platform. Model-based counterfactuals, including synthetic control and propensity models, are rated strong to moderate. Econometric approaches such as marketing mix modeling are rated moderate to weak with high scope. Hybrid proxies are rated weak, and that category is worth reading closely: it holds new-to-brand percentages, baseline versus exposed analysis, simple multi-touch attribution, and platform-reported incrementality.

Inside the experimental family the designs differ mainly in cost. A public service announcement test serves a charity or house ad to the control group, keeping auction dynamics comparable but paying for impressions that generate nothing. An intent-to-treat design avoids that cost by randomizing before delivery, but dilutes the effect across everyone assigned to treatment, including people who never saw an ad. Ghost ads resolved the trade-off: the platform runs its normal auction for control users, logs the impression that would have been served, and serves something else, turning the log entry into a counterfactual at no media cost. Garrett Johnson, Randall Lewis and Elmar Nubbemeyer described the method in a 2015 working paper published in the Journal of Marketing Research in 2017, reporting that a variant called predicted ghost ads was already recording more than 100 million entries a day at roughly a tenth of the cost of the alternatives.

Geo experiments avoid user-level identity altogether. Non-overlapping regions are assigned at random, spend is changed through geo-targeting, and only region-level outcomes are needed. Jon Vaver and Jim Koehler set out the framework in a Google paper in 2011. It survived the collapse of third-party identifiers because it never depended on them, which is why go-dark and matched-market tests became the default in connected television.

Inside a platform study

Google's Conversion Lift illustrates the plumbing. The geography-based version uses Google Marketing Areas, sub-country units generated by a spectral clustering algorithm that holds down contamination - exposure in a treatment region followed by conversion in a control region - while keeping enough units for statistical robustness.

Before launch the interface returns a feasibility rating of high, medium or low alongside a minimum detectable iROAS, and Google advises against proceeding on a low rating. The user-based version carries explicit floors: results appear from the fourth reporting day and require 150 conversions in treatment and 65 in control, alongside spend of at least $5,000. Duration guidance sets a floor of seven days and recommends fourteen, since shorter studies showed drops of up to 17% in absolute lift where conversion lags are long.

Google moved the statistics from frequentist to Bayesian during 2025, using historical campaign data as priors. That changes what the interval means: an 80% credible interval carries an 80% probability of containing the true lift, which is not what a confidence interval says.

Origin and how it hardened

Academic field experiments established that the gap was large. Thomas Blake, Chris Nosko and Steven Tadelis, running experiments at eBay and publishing in Econometrica in 2015, found that brand keyword advertising produced no measurable short-term benefit, and that average returns on non-brand search were negative because most spend reached frequent buyers whose behavior the ads did not change.

Brett Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky then tested whether observational methods could reproduce experimental answers. Using 15 US experiments at Facebook covering 500 million user-experiment observations and 1.6 billion impressions, published in Marketing Science in 2019, they reported that observational models often failed to recover the experimental effect even after conditioning on extensive demographic and behavioral variables. Privacy regulation and identifier deprecation did the rest: as user-level tracking degraded, methods built as research tools turned into commercial products.

Why it matters

Incrementality now sits at the center of budget arguments rather than the edge. Research cited around Smartly's March 2026 agreement to acquire INCRMNTAL put proving incremental return at the top of measurement priorities for 67.4% of marketers, ahead of cross-channel attribution accuracy at 55.1%. Confidence in the methods is lower than adoption: IAB's State of Data 2026 report, published February 7, 2026, recorded up to 75% of buy-side decision-makers rating attribution, incrementality tests and mix models as underperforming.

Platform language has adjusted accordingly. Meta published "Building a Suite of Truth" on May 28, 2025 and now calls Conversion Lift the gold standard in measurement while conceding full adoption takes time. Its Incremental Attribution product rolled out globally in June 2025 on self-reported testing showing a 46% average rise in incremental conversions.

Limitations and disputes

Every experiment has an opportunity cost, which Google's documentation states plainly: holding back part of an audience forgoes conversions the ads would have produced. Google reported that most advertisers run one to two studies a year, a cadence below what structural calibration methods require.

Contamination is the second constraint. Control-group members exposed on an unmeasured platform, or converting across a geographic boundary, compress the measured gap, which the IAB paper names as a weakness of experiment-based methods in multi-platform settings.

The third dispute concerns who grades the work. The IAB and IAB Europe classified platform-reported incrementality as a hybrid proxy, the weakest of the four tiers, because it relies on platform-specific attribution rather than an independent control group. The task force behind that classification included staff from Criteo, Meta, Instacart, Walmart, Uber and Google. Neither Google's studies nor the vendor analyses disputing them have been published in a form permitting independent replication, and vendor lift figures are typically self-reported.

Statistical power is the quiet failure mode: an underpowered study returns a wide interval containing zero, which is not evidence of no effect, and the budget is spent anyway.

Disambiguation

Attribution assigns credit among observed touchpoints. Incrementality tests whether the touchpoint mattered. The two can disagree completely without either being computed incorrectly.

Marketing mix modeling estimates channel contribution from aggregate time series. It is econometric rather than experimental, and the IAB rates its causal strength moderate to weak while rating its scope high. Experiments increasingly calibrate it.

Brand lift measures survey-reported changes in recall, awareness or purchase intent among an exposed group against a control. The design is experimental, the outcome attitudinal rather than transactional.

New-to-brand classifies buyers by prior purchase history inside a retailer's data. It is a proxy indicator rather than a causal estimate.

Recent developments

Tooling has moved toward continuous measurement and lower thresholds. Google cut its minimum experiment budget from figures approaching $100,000 to $5,000 on November 11, 2025, citing statistical changes delivering up to 50% more conclusive results. It announced Meridian GeoX, an open-source geographic testing tool feeding results into its mix model, on May 5, 2026, added Uplift Experiments for Demand Gen on May 20, 2026, and exposed 24 lift metrics through the Google Ads API in August 2026.

Vendors moved in parallel. Nexxen built ghost bidding into its demand-side platform on March 24, 2026, submitting control-group bids without serving ads. Smartly signed a letter of intent to acquire INCRMNTAL on March 16, 2026, adding always-on measurement that reads natural fluctuations in spend rather than pausing media. Kochava added self-serve on/off pulse testing in the second quarter of 2026, and Jamloop opened household-level holdouts on July 21, 2026, reporting 3,224 incremental subscriptions at 99.98% confidence.

The most pointed recent contribution runs against how results are consumed. A paper posted to arXiv on August 21, 2026 by Niklas Heusch argues that collapsing a geo-experiment into one return figure discards the temporal structure identifying carryover and saturation, reporting on synthetic data that a standard mix model returned 10.61x against a true 4.20x while structural estimation from four experiments recovered 4.14x.

Timeline

  • 2011 - Jon Vaver and Jim Koehler publish Google's geo experiment framework for measuring ad effectiveness through regional assignment
  • 2015 - Blake, Nosko and Tadelis publish eBay field experiments in Econometrica, finding no measurable short-term benefit from brand keyword advertising
  • June 2015 - Johnson, Lewis and Nubbemeyer circulate the ghost ads methodology, published in the Journal of Marketing Research in 2017
  • August to November 2015 - The Edmunds.com branded search experiment runs across 210 designated market areas
  • 2019 - Gordon, Zettelmeyer, Bhargava and Chapsky publish the Facebook comparison finding observational methods often fail to recover experimental effects
  • January 2025 - Google opens Meridian globally, incorporating experiment results as model calibration
  • May 22, 2025 - Google announces Bayesian incrementality testing at Google Marketing Live
  • May 28, 2025 - Meta publishes "Building a Suite of Truth"
  • June 2025 - Meta rolls out Incremental Attribution globally
  • September 9, 2025 - IAB and IAB Europe publish "Demystifying Incrementality in Commerce Media"
  • November 3, 2025 - IAB and IAB Europe publish the Guidelines for Incremental Measurement in Commerce Media
  • November 11, 2025 - Google reduces the minimum incrementality experiment budget to $5,000
  • March 16, 2026 - Smartly signs a letter of intent to acquire INCRMNTAL
  • March 24, 2026 - Nexxen adds ghost bidding incrementality to its demand-side platform
  • May 5, 2026 - Google announces Meridian GeoX
  • May 20, 2026 - Uplift Experiments announced for Demand Gen at Google Marketing Live
  • July 21, 2026 - Jamloop opens household-level holdout methodology in its self-serve platform
  • August 21, 2026 - Structural estimation of mix model parameters from geo-experiments posted to arXiv

Summary

Who: Advertisers, agency measurement teams, retail media networks and platform measurement groups, with standards work led by the IAB Commerce Board, IAB's Task Force on Incrementality and IAB Europe's Retail Media Committee, and product implementations from Google, Meta, Nexxen, Kochava, Jamloop, Smartly and INCRMNTAL.

What: Incrementality is the causal impact of a campaign, defined as the additional business outcomes it produced compared with a counterfactual in which it did not run. It is estimated through randomized experiments, holdouts, ghost ads, geo tests, modeled counterfactuals or econometric models, and expressed as absolute lift, relative lift or incremental return on ad spend.

When: The experimental methods date to 2011 for geo experiments and 2015 for ghost ads and the eBay field experiments. Standardized definitions arrived on September 9 and November 3, 2025. Platform access widened on November 11, 2025 when Google cut the minimum experiment budget to $5,000.

Where: Inside buying platforms as native lift products, in third-party measurement vendors, in clean rooms, and increasingly as a calibration input to marketing mix models.

Why: Attributed conversions and reported return on ad spend describe outcomes without establishing causation. As identifier-based tracking degraded and budget scrutiny increased, causal estimates became the reference point for allocation decisions, while remaining costly, slow and contested over who is qualified to produce them.