Church one: the dashboard is truth. Platform ROAS is the scoreboard, scale what’s green, kill what’s red. Church two: the dashboard is fraud. Meta claims credit for people who were buying anyway, attribution is astrology for media buyers, burn it all down.
Incrementality testing is the audit that both churches said would vindicate them. Now the results are in at serious scale, and they’re inconvenient for everyone.
Incrementality testing measures the sales that would not have happened without the ads, usually by holding out a matched group of regions or users from exposure and comparing outcomes against business-as-usual. Not clicks. Not attributed revenue. Caused revenue.
Haus, the incrementality platform built by ex-Netflix and Meta economists, published an analysis of 640 Meta incrementality experiments run by its customers since the start of 2024, including head-to-head tests of Advantage+ against manual campaigns. It’s the largest public dataset of its kind. Almost every number in it breaks somebody’s favourite story.
The Meta-is-fraud camp has a problem
Start with the finding the sceptics won’t enjoy.
Meta drove an average 19% lift to brands’ primary KPI across account-wide studies (Haus, The Meta Report)
77 of the 100 highest-lift experiments ever run on Haus are Meta tests, including the single-channel record at 74% lift (Haus)
For advertisers reading 7-day click attribution only, Meta actually UNDER-reports its own incrementality by 15% on average (Haus)
Read that last one again. The platform whose entire villain arc is inflated self-reporting is, for click-only DTC measurement, selling itself short.
And for omnichannel brands, roughly 32% of Meta’s measured impact landed outside DTC, in retail and Amazon revenue the ad account never sees. Haus points out the maths: an omnichannel brand can break even at a 0.68x DTC iROAS. If you’re holding Meta to a 2x in-platform bar while half your volume moves through Coles, you’re not being rigorous. You’re being expensively wrong in the conservative direction.
The dashboard-junkies have a bigger problem
Now the other church.
58% of brands saw higher incremental ROAS on manual campaigns than on Advantage+ in head-to-head tests, despite Advantage+ reporting 2.4% higher platform ROAS (Haus)
Advantage+ over-reported relative to manual campaigns by 12 percentage points once post-treatment effects were counted (Haus)
Post-treatment lift, the sales that keep arriving after a test ends, averaged +32% for manual campaigns versus +17% for Advantage+ (Haus)
The pattern is almost elegant. Advantage+ is exceptional at finding people who are about to buy. That’s why it looks brilliant at the experiment midpoint (9% ahead of manual) and why it fades by the end (12% behind). An algorithm optimised to find high-intent users will, at the margin, find people who didn’t need an ad. Intent identification and incrementality are not the same job. Sometimes they’re opposite jobs.
Because the model gets rewarded for conversions, not for caused conversions. Because nobody’s dashboard has a column for “would have bought anyway”. Because the machine is doing exactly what it was told, which is not what you meant.
Two Meta researchers’ worth of foundational canon predicted this: Gordon, Zettelmeyer, Bhargava and Chapsky’s 2019 Marketing Science study compared attribution models against randomised experiments at Facebook and found attribution routinely failed to approximate true lift, in either direction. The 640 tests are that paper’s warning, now at industry scale.
But before anyone torches their ASC campaigns: for 42% of brands, Advantage+ won the head-to-head. It also drove a slightly higher share of impact from new customers than manual (70% vs 65%), and a bigger omnichannel halo (+51% vs +43%). It is not a retargeting smuggler. It’s a specific tool with a specific bias, and now the bias has a number.
Where incrementality testing gets weird
Three findings peer operators should sit with.
The 50/50 trap
Brands that moved from a skewed setup (75%+ of spend in their winning campaign type) to a “balanced” 50/50 split saw an 18% drop in iROAS on average (Haus). Diversification instinct, applied inside one platform, mostly just splits your signal. Pick a lane, feed it.
Mid-funnel is quietly under-priced
Mid-funnel optimisation showed 14% lower DTC iROAS than lower-funnel, but 9% higher post-treatment lift, a 70% versus 46% omnichannel halo, and a 78% versus 67% new-customer rate (Haus). Testing of mid-funnel tactics grew 121% across Haus customers. The cheap-looking thing is buying futures; the efficient-looking thing is harvesting.
The correction cuts both ways
Haus originally found Meta’s new Incremental Attribution product underperforming standard attribution. They re-ran the analysis in July 2026: it now outperforms. When the audit changes its mind in public, that’s not weakness. That’s the entire point of running audits.
What you can do right now
Be honest about the entry price first. A clean geo-holdout needs enough order density per region to detect lift, and the working consensus among practitioners puts that around $30-50k of monthly spend before results stop being noise. Nobody has published a precise floor; anyone quoting one to the dollar made it up. If you’re under it, you don’t run the 641st test. You steal the priors from the first 640:
Stop judging Meta on 7-day click alone; it’s biased low for DTC and blind to omnichannel spill. Judge account changes against MER and total revenue over 4-to-6-week windows instead.
Run your own crude holdout at the account level if you can stomach it: platforms off in one comparable state or fortnight, watch what total revenue actually does. It’s not a Haus experiment. It beats faith.
Don’t split ASC and manual 50/50 for tidiness. Skew hard toward whichever wins in your account, and re-check quarterly.
If you sell through retail or marketplaces, put the 0.68x arithmetic in front of whoever sets your ROAS targets. Today.
And if you’re above the entry price and still making seven-figure budget calls on attributed ROAS alone, you’re not measuring. You’re vibing with extra steps.
The dashboard was never truth. It was never fraud either.
It’s a witness. Cross-examine it.
If you want help turning this evidence into an actual measurement plan for your brand, book a growth call with Thrive. We’d rather argue with your data than agree with your dashboard.
FAQ
What is incrementality testing? Incrementality testing measures the sales caused by advertising, not just correlated with it, by comparing a group exposed to ads against a matched holdout group that isn’t. The difference is incremental revenue. It corrects for attribution’s core flaw: claiming credit for customers who would have purchased anyway.
Is Meta advertising actually incremental? At scale, yes. Across 640 incrementality experiments analysed by Haus, Meta drove an average 19% lift to brands’ primary KPI, and 77 of the 100 highest-lift tests ever run on the platform were Meta tests. The open question for most brands is efficiency (lift relative to spend), not whether lift exists.
Is Advantage+ less incremental than manual campaigns? Usually, but not always. In Haus head-to-head tests, 58% of brands saw higher incremental ROAS from manual campaigns, and Advantage+ over-reported by about 12 percentage points once post-treatment effects were included. But 42% of brands saw Advantage+ win, so the answer is brand-specific and worth testing.
How much do I need to spend to run an incrementality test? Geo-holdout tests need enough conversion volume per region for statistical power. Practitioner consensus puts the practical floor around $30-50k in monthly ad spend; below that, published results from large test libraries like Haus’s Meta Report are a better guide than an underpowered test of your own.
