Attributed, Incremental, and Why They Diverge
An attributed conversion is one your measurement system credited to an ad, according to whatever window and model you configured. An incremental conversion is one the ad caused.
Every attributed conversion falls into one of two buckets. Either the person bought because of the ad, or they were going to buy anyway and the ad happened to be in the path. The second bucket is real, it is often large, and nothing in your dashboard separates it from the first.
Incremental ROAS, usually written iROAS, is incremental revenue divided by media spend. It is the number that actually decides whether a campaign should exist. Reported ROAS is incremental ROAS plus whatever credit the model took for conversions that were going to arrive regardless.
The 2026 ADKDD paper out of TikTok gave the second bucket a name worth borrowing: the cannibalization rate, defined as the fraction of nominal paid-attributed conversions that is not truly incremental. Their framing is exactly right. Paid-attributed conversions overstate true growth when the paid channel overlaps with organic demand, brand-driven traffic, or another channel that was already going to close the sale.
The Evidence That the Gap Is Large
I am going to be careful here, because incrementality is a topic where vendor blogs quote each other into a circle and a made-up percentage acquires a citation trail. What follows is peer-reviewed or platform-published, and I have said which is which.
eBay, brand keywords, 2015. In a field experiment published in Econometrica, eBay switched off brand-keyword advertising on Yahoo and MSN while holding Google as a control. 99.5% of the forgone paid click traffic came back through natural search instead, and the authors describe that as a lower bound on retention. The same paper reported non-brand search at an observational return of 1,632% with time and geography controls, against an experimental return of negative 63%.
Facebook, fifteen experiments, 2019. Gordon and co-authors compared observational estimates against randomised ones across 15 US Facebook experiments, 500 million user-experiment observations and 1.6 billion impressions. In half the studies, the estimated percentage increase in purchases was off by a factor of three. In one study, the true randomised lift was 73%, while a naive comparison of exposed against unexposed users reported 316%, and exact matching on age and gender still reported 222%.
The 222% row is the important one: cleaner data cut the error roughly in half and still overstated the truth by a factor of three.
That is the finding I would put in front of anyone who thinks better tracking is the answer to this. It is not. Better tracking makes the attributed number more accurate as an attributed number. The distance between 222% and 73% is not measurement error, it is selection: the people who saw the ad were already more likely to buy.
Retargeting, measured properly. The honest published numbers are more modest than the dashboards. A ghost-ad controlled retargeting experiment found 17.2% more site visits and 10.5% more purchases. A separate randomised field experiment found retargeting caused 14.6% more users to return within four weeks, with 33% of the first week's effect landing on day one. A January 2026 geo experiment across both Google and Facebook found large and consistent returns on prospecting while retargeting returns were significantly lower; that paper's full text is paywalled, so I am quoting its direction and not inventing a figure for it.
You will see "20 to 30% of retargeting conversions would have happened anyway" quoted with great confidence. I could not find a study behind it. Treat it as folklore.
How Incrementality Is Actually Measured
Three methods, in descending order of how much you can trust the answer.
Randomised holdout. Split your audience, show ads to one group, withhold from the other, compare outcomes. This is the gold standard because assignment is random, which removes selection entirely. Platform lift products do this internally.
Geo experiment. Split geographies rather than people. Pick matched markets with similar historical behaviour, run the ad in one set and go dark in the other, then model the counterfactual for the dark set from its own history and the control's movement. This is what you use when you do not control the ad delivery well enough to randomise users, which describes most affiliates.
Observational modelling. Propensity matching, media mix modelling, anything that reconstructs a control group from data you already have. Cheapest, weakest, and the thing the Facebook paper found off by a factor of three.
What the Platforms Offer in 2026
Meta. Incremental Attribution launched around April 2025 and now sits in Ads Manager both as an optimisation setting and as a reporting column. Meta's own newsroom in January 2026 reported that its Q4 model rollout drove a 24% increase in incremental conversions against the standard attribution model. Conversion Lift, the actual experiment product, is not self-serve: Meta's help documentation directs you to your account representative to find out what tests are available and what minimums apply, and publishes no dollar figure. The self-serve route is A/B testing inside Experiments, which reports once there are at least 100 conversion events per strategy tested. Note also that the Conversions API is mandatory for Conversion Lift tests measuring standard web events started on or after 1 October 2021.
Google. Conversion Lift also requires a rep. Google Ads Help still states plainly that it is not available for all accounts and that you should contact your Google account representative. What did change is the cost: Google's documentation says an experiment that once cost upwards of $100,000 can now be run for $5,000, and attributes that to moving the methodology from Frequentist to Bayesian, using historical campaign data as priors and reporting 80% credible intervals. Geo-based tests run on Google Marketing Areas, built by spectral clustering with an explicit model for people travelling between areas, and the setup screen tells you a feasibility rating and a minimum detectable iROAS before you commit.
Cheaper is not the same as open. As of August 2026 the $5,000 figure and the contact-your-rep gate sit in the same set of Google help pages.
If Nobody Will Give You a Rep
Which, if you are an affiliate, is the likely outcome. The do-it-yourself path is a geo holdout analysed with open-source tooling.
- CausalImpact, Google's Bayesian structural time-series package, is actively maintained. Version 1.4.1 was published to CRAN on 26 September 2025.
- Meridian GeoX, announced May 2026, is Google's open-source, publisher-agnostic geo experiment framework, with holdback, go-dark and heavy-up designs and results that feed back into a media mix model as priors.
- GeoLift, Meta's synthetic-control package, is the one everybody cites. Its repository is public and carries an active badge, but its most recent release, v2.6.06, is dated 19 May 2023. Roughly three years without a release is worth knowing before you build a measurement practice on it.
The design constraint every one of these is fighting is contamination: someone exposed in a treatment market who converts in a control market drags the measured lift down. Google's own documentation explains this clearly and it is the reason matched markets are constructed algorithmically rather than by picking states that feel similar.
Why This Matters More to Affiliates Than to Anyone Else
Affiliates run exactly the campaign types with the worst incrementality profile and the best-looking dashboards.
Branded and brand-adjacent search converts beautifully and is the single clearest case of credit without cause. Retargeting shows the highest ROAS in every account I have ever looked at and measures, in controlled experiments, in the low tens of percent. Both are the campaigns a media buyer defends hardest in a budget meeting, because the reported numbers are extraordinary.
There is a second-order problem specific to affiliates. If your refunds, rebills and chargebacks never reach the ad platform, the attributed number is not only overcounting causally, it is overcounting arithmetically. Two independent errors, stacked, both pointing the same direction.
So What Do You Do About It
Incrementality is not a metric you can add to a report. It is a thing you have to go and measure, and measuring it costs you revenue while the holdout runs. That is the entire reason it is rare.
I want to be plain about my own position here. ClickerVolt does not measure lift and cannot. What it does is make the data going into a lift test complete and correctly timed: the full fifteen-field Meta payload so the platform-side numbers are not degraded by poor matching, Refund Sync so a reversal reaches Google, Meta and TikTok rather than sitting in a report, and unlimited retention so the pre-period you compare against still exists. You can see how that part works here.
The free thing to do this week is smaller and more useful than buying anything. Take the campaign with your best reported ROAS, and write down what you would expect to happen to total revenue if you turned it off for two weeks. Then ask what evidence you have for that number. If the honest answer is that the dashboard said so, you have found the campaign worth testing first.
