Developing
All results are synthetic. There is no validation on real advertiser data, and the 2.5x bias is specific to this simulation.
In 60 seconds
- Niklas Heusch's preprint (arXiv 2608.21128) was submitted on August 21, 2026.
- On synthetic data with a true paid search ROAS of 4.20x, a realistic MMM reported 10.61x; an MMM with oracle controls reported 8.41x.
- A structural method using geo-experiment results recovered 4.14x with four tests, with the true value inside its interval.
- The author's affiliation is not listed, though the contact address uses a zalando.de domain.
Impact Layer
- What changed
- A new method proposes using geo-test results to pin down marketing mix model parameters instead of just checking them.
- Why it matters
- If MMMs inflate channels that are correlated with demand, budget shifts based on them can be wrong.
- Who is affected
- Analytics and finance teams that rely on MMMs for budget allocation.
- Martech stack area
- Analytics, Advertising
- Stack impact
- Treat MMM output as a model, and calibrate it with geo or holdout experiments where you can run them.
- Business impact
- Moderate A single-author preprint using synthetic data, as summarised by PPC Land. Not peer reviewed.
- Related technologies
- What to watch
- Replications on real advertiser data.
A preprint by Niklas Heusch, titled “Structural Estimation of Marketing Mix Model Parameters from Geo-Experiments” (arXiv 2608.21128, submitted August 21, 2026), tests how well marketing mix models recover the true effect of paid search. PPC Land covered it on August 31.
The setup
The data are synthetic: an online retailer over 156 weeks with three channels (paid search, social, TV) and four geo-tests. True paid search ROAS is 4.20x. The method differences treatment and control outcomes, which cancels the common component, including unobserved confounders, if assignment is random.
Results
- MMM with realistic controls: 10.61x (90% interval 6.56 to 14.36), which excludes the truth.
- MMM with oracle controls: 8.41x (6.91 to 9.81), also excludes it.
- Structural method, two tests: 4.31x (3.87 to 4.75); four tests: 4.14x (3.79 to 4.48). Both contain the truth.
- The realistic MMM estimated carryover (adstock) at 0.50 against a true 0.20; the four-test method estimated 0.19.
Limits
PPC Land stresses that this is simulated data with no real-advertiser validation. The method assumes the assumed adstock and saturation forms are correct, no spillover between regions and constant parameters, and it needs tests with enough variation in spend, which costs money. The 2.5x gap is a property of this simulation, not a general constant.
Our read
The lesson is not “MMM is wrong” but “calibrate it”. Where a geo-test is feasible, feed its result into the model rather than only comparing afterwards.
Sources
- PPC Land: MMM overstates paid search ROAS by 2.5 times, Zalando researcher finds (opens in a new tab)
Source of the method, results table, simulation design and caveats. The PPC Land article's first paper link appears to point to a different paper.