Incremental return on ad spend (iROAS) is the revenue that advertising caused, divided by what the advertising cost. It differs from ordinary return on ad spend (ROAS) in its numerator. ROAS counts every sale an attribution rule credits to an ad, including purchases that would have happened anyway. iROAS counts only the difference between what happened and what would have happened without the ads, a quantity that cannot be observed directly and has to be estimated, usually by withholding ads from a comparison group. The metric exists because platform-reported returns tend to flatter the media that generated them.

The arithmetic, worked through

Google's Conversion Lift documentation gives the standard formula: incremental conversion value, meaning the treatment group's conversion value minus the control group's, divided by total ad spend. Incremental conversions, which Google also calls absolute lift, are treatment minus control conversions, and incremental cost per action (iCPA) is spend divided by them.

A retailer spends 20,000 euros on a campaign and its platform attributes 120,000 euros of sales to it, a reported ROAS of 6.0. An experiment holds back a randomly chosen 10% of the target audience. After scaling the control group to the same size as the exposed group, exposed users generated 260,000 euros of sales and the comparable unexposed users 230,000 euros. The incremental value is 30,000 euros, and the iROAS is 1.5. Most of the attributed revenue was sales the campaign intercepted rather than created.

Revenue is not profit. At a 40% gross margin, 30,000 euros of incremental sales yields 12,000 euros of incremental gross profit against 20,000 euros of spend. The break-even iROAS is the reciprocal of the margin, 2.5 in this case, so a campaign reporting a ROAS of 6.0 lost money on the first purchase.

The denominator is not settled either. Google's Conversion Lift product divides by total spend. Google researchers Aiyou Chen and Timothy Au, in a 2019 paper on paired geographic experiments, define iROAS as the difference in revenue between treatment and control regions divided by the difference in spend. The distinction matters in a heavy-up test, where spend rises from 50,000 to 80,000 euros in some regions: only the extra 30,000 euros belongs in the denominator.

How the counterfactual is built

Three families of method supply the "what would have happened" figure.

User-level holdouts randomly split an audience before a campaign starts and suppress ads for the control group, the design described in PPC Land's explainer on the holdout study. Google's Conversion Lift and Meta's Conversion Lift both work this way. Google cut the minimum budget for its incrementality experiments to 5,000 dollars on November 11, 2025, from figures approaching 100,000 dollars, and recommends studies of at least 14 days, with a seven-day floor. Shorter studies showed drops of up to 17% in absolute lift for businesses with long conversion lags.

Geographic experiments split markets rather than people. Spend changes in treatment regions while control regions continue as before, and a regression on pre-test data predicts what the treatment regions would have sold. The method needs no user identifiers. It struggles with few regions, the problem a 2017 time-based regression method by Jouni Kerman, Peng Wang and Jon Vaver addressed.

Marketing mix models (MMM) estimate incremental return from years of weekly sales and spend data, without withholding anything. They cover every channel at once but rely on statistical assumptions rather than randomisation. Google's Meridian reports, for each channel, incremental outcome, return on investment, marginal return and a response curve, each with a credible interval. Meta's open-source package Robyn uses ridge regression and a budget allocator that aims to equalise marginal ROAS across channels, according to PPC Land's examination of the tool. Both can be calibrated with experiment results.

Where platforms report it

The advertiser or agency commissions a test; the selling platform usually runs it. Google's interface shows incremental conversion value and incremental ROAS in Conversion Lift reports, and Google Ads API v25.1, announced August 19, 2026, added 24 Conversion Lift metrics as read-only resources for allowlisted accounts.

Meta went further than reporting. In June 2025 it rolled out incremental attribution globally, a setting that optimises delivery towards conversions predicted to be incremental, trained on Conversion Lift data. Meta's white paper on a "suite of truth" cited a three-cell test in which the makeup brand Laura Geller recorded 3.3 times higher incremental ROAS in the incremental attribution cell than in its standard campaign.

Retail media networks report attributed ROAS by default. Amazon moved its Marketing Mix Modeling API to general availability in 14 countries on May 1, 2026, and on May 15, 2026 opened a workflow letting Amazon DSP advertisers self-serve more than 50 third-party measurement products, including offline sales lift studies. In November 2025, IAB and IAB Europe published guidelines for incremental measurement in commerce media, grading methods by causal rigour for optimisation, return validation and platform calibration, according to IAB New Zealand.

Origin and evolution

The ratio emerged from Google's geo-experiment research. Jon Vaver and Jim Koehler published a geo-based regression framework in 2011. A May 31, 2016 post on the Unofficial Google Data Science blog by Kerman, Vaver and Koehler used "iROAS" and defined it as "the slope of a curve of the response metric plotted against the underlying advertising spend" - a marginal definition rather than an average one. Chen and Au's Trimmed Match paper, posted to arXiv in August 2019, stated that advertisers "frequently find the iROAS" more informative and actionable.

Academic work supplied the motive. Randall Lewis and Justin Rao analysed 25 large field experiments with US retailers and brokerages, covering about 2.8 million dollars of digital spend, in the Quarterly Journal of Economics in 2015. The median confidence interval on return on investment was more than 100 percentage points wide. In 2019, Brett Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky showed in Marketing Science that observational methods often failed to recover the results of 15 randomised Facebook experiments.

Signal loss then turned the concept into a product category: Google unveiled Meridian in March 2024 and opened it to all users on January 29, 2025.

Why it matters

Attributed ROAS is routinely the figure used to justify budgets, and platforms grade their own homework. A ROAS of 10 on Meta may indicate worse performance than a ROAS of 2, according to marketers who argued that view-through credit and retargeting of existing customers inflate the higher figure. iROAS gives finance teams a number closer to the economic question, and commerce media's growth has sharpened it. The IAB's one-page incrementality framework of September 2025 explicitly separated incrementality from media-attributed ROAS.

The gap can run in either direction. Meta's white paper, based on 307 studies by 54 advertisers, argued that rules-based attribution undervalues Meta by 31% at the median. A vendor study by the measurement company Incremental, covering 150,000 campaigns and 350 million dollars of spend, estimated that siloed retailer attribution misses 36% to 53% of retail media impact because sales land at other retailers. Both findings come from parties that benefit from larger measured returns.

Limitations and disputes

Statistical power is the first constraint. Lewis and Rao's result implies that many campaigns cannot be measured precisely at any affordable sample size, and an interval that includes zero is not evidence that an ad failed. Holdouts also forgo sales from the control group.

Who runs the test is the second. When the platform selling the media designs the experiment, computes the control group and reports the result, independence is limited. Karan Dhir, a measurement product lead at Genentech, argued in September 2026 that standard methods miss experimental ground truth "frequently by a factor of 3 or more", a multiplier the 2019 paper itself does not publish.

Models versus experiments is the third. In a paper posted on August 21, 2026, Niklas Heusch showed on synthetic data that a standard mix model reported 10.61x ROAS for paid search against a true 4.20x, while structural estimation from four geo experiments returned 4.14x. The data were simulated, not drawn from a real advertiser.

Definitions remain inconsistent. Average and marginal iROAS answer different questions, denominators vary, and few published case studies state test length, confidence or margin. Short windows miss long-term brand effects.

Not the same as

ROAS divides attributed revenue by spend. iROAS estimates causation instead.

Incrementality is the concept, the causal effect of advertising. iROAS is one way of expressing it in money, alongside absolute lift, relative lift and iCPA.

Sales lift is the percentage or absolute increase in sales between exposed and unexposed groups, as PPC Land's sales lift explainer sets out. It measures effect size; iROAS divides that effect by cost.

Marginal ROAS (mROAS) is the return on the next unit of spend, read from a response curve. A channel can carry a healthy average iROAS while its marginal return has already fallen below break-even.

Recent developments

At Google Marketing Live on May 20, 2026, Google said advertisers building first-party data strength saw an 11% average increase in incremental ROAS on value-bidding Search campaigns, based on its own data, and announced Meridian's integration into Analytics 360. A Lifesight report on connected TV, circulated in August 2026, cited a 3.03x incremental ROAS for an anonymised brand without spend, duration or confidence interval.

Google's tooling continues to move from experiments into models. Meridian GeoX, an open-source geo-testing library, reached general availability on September 9, 2026, alongside Meridian 2.0.0, which can take GeoX results as priors. As of October 2026, PPC Land's incrementality explainer counts experiments, model-based counterfactuals and mix models as the main routes to the number, none of them yet independently audited across platforms.

Timeline

  • 2011: Jon Vaver and Jim Koehler publish Google's geo-based regression framework for measuring ad effectiveness.
  • 2015: Randall Lewis and Justin Rao publish "The Unfavorable Economics of Measuring the Returns to Advertising" in the Quarterly Journal of Economics, based on 25 field experiments.
  • May 31, 2016: Kerman, Vaver and Koehler use "iROAS" on the Unofficial Google Data Science blog, defining it as the slope of response against spend.
  • 2017: Kerman, Wang and Vaver publish the time-based regression method for geo experiments with few regions.
  • March 2019: Gordon, Zettelmeyer, Bhargava and Chapsky publish their comparison of 15 Facebook experiments with observational methods in Marketing Science.
  • August 2019: Chen and Au post the Trimmed Match paper on estimating iROAS from paired geo experiments.
  • December 2020: Resident begins running Meta's open-source Robyn mix model alongside its in-house model.
  • March 2024: Google unveils Meridian on a limited-availability basis.
  • January 29, 2025: Meridian opens to all marketers and data scientists.
  • June 2025: Meta rolls out incremental attribution globally.
  • September 9, 2025: IAB and IAB Europe publish "Demystifying Incrementality in Commerce Media".
  • November 11, 2025: Google lowers the Conversion Lift minimum budget to 5,000 dollars.
  • November 2025: IAB and IAB Europe release guidelines for incremental measurement in commerce media.
  • May 1, 2026: Amazon's Marketing Mix Modeling API reaches general availability in 14 countries.
  • May 20, 2026: Google reports an 11% average incremental ROAS increase tied to first-party data strength.
  • August 19, 2026: Google Ads API v25.1 adds 24 Conversion Lift metrics for allowlisted accounts.
  • August 21, 2026: Niklas Heusch posts a paper showing a mix model overstating paid search returns on synthetic data.
  • September 9, 2026: Meridian GeoX reaches general availability alongside Meridian 2.0.0.

Summary

  • Who: Advertisers and agencies commission iROAS measurement; platforms such as Google, Meta and Amazon, measurement vendors and in-house analysts run the tests and models; IAB and IAB Europe have issued guidance on method.
  • What: Incremental return on ad spend, the revenue advertising caused divided by its cost, estimated against a counterfactual of what would have happened without the ads.
  • When: The term appears in Google geo-experiment research by 2016, built on work from 2011; it became a reported platform metric and commercial claim through the 2020s, with Google, Meta and Amazon all expanding related tools between 2024 and 2026.
  • Where: In Conversion Lift reports in Google Ads and Meta, geo experiments, marketing mix models such as Meridian and Robyn, and retail media and connected TV measurement studies.
  • Why: Attributed ROAS credits ads with sales that would have occurred anyway, so it can overstate or understate value; iROAS attempts to measure the causal return that budget decisions depend on, though its estimates vary with method, denominator and who runs the test.