Measurement and Attribution Under Opt-In Consent

In the best-documented comparison available, an attribution-style estimate put return on ad spend above 4,100% while a randomised experiment on the same spend returned minus 63%. That is not a measurement discrepancy. It is two methods disagreeing about whether the money made anything at all, and only one of them is making a causal claim. Everything else on this page follows from that single comparison, because in Europe — where opt-in consent compounds with browser restrictions to produce the weakest attribution coverage of any major market — the method you trust determines the budget you set.

A capability page from Digital, Ecommerce & Performance Marketing in Europe. Read after performance marketing under opt-in consent, which covers why the signal is degraded in the first place. Last reviewed August 2026.

1. The three methods and what each can prove

They are routinely presented as competing answers to one question. They are separate answers to three different questions, and only one of them is causal.

Based on 2026 analysis distinguishing the three approaches, noting that multi-touch attribution is bottom-up and tactical but only sees the trackable slice, MMM is top-down and covers all media including offline and brand while accounting for seasonality and pricing, and that only incrementality testing makes a causal claim.

Source: 2026 measurement analysis citing this as the best-documented comparison available. A single documented case is not a general law — the useful takeaway is the direction and the magnitude of the possible gap, not that every attributed figure is inflated by that amount.

2. Why attribution broke, and how far

Multi-touch attribution depends on user-level cross-device tracking requiring persistent identifiers. Reported figures put the damage clearly: attribution coverage fell from over 90% to roughly 60–80% following iOS 14.5 and Safari tracking prevention, and in some channels to between 30% and 60%.

Layer European opt-in consent on top of that. As covered in the consent analysis, published opt-in rates for marketing cookies run around 46% across the EU with Germany reported near 36%. A European advertiser is therefore losing signal twice — once to the browser and once to the banner.

Attribution has not died. It has been demoted from a source of truth to a tactical signal for optimising digital channels, and treating it as anything more in European digital marketing is where the budget errors start.

3. Why MMM suits Europe specifically

Marketing mix modelling is a technique from the 1960s enjoying a revival, and the reason is structural rather than fashionable: it requires no cookies, no device identifiers and no user-level tracking at all.

In a market where roughly half your visitors decline consent, a method that never needed their individual data is not a workaround. It is the only major measurement approach whose accuracy is unaffected by the consent rate.

Based on reported analysis that MMM uses aggregate historical data without tracking individuals, and that server-side conversion APIs recover a reported 20–30% of conversions. The bottom two rows are the reason this cluster keeps recommending them for European ecommerce marketing.

For European digital marketing teams that is a structural advantage rather than a consolation prize. What changed in 2026 is access rather than mathematics. Google open-sourced Meridian, Meta maintains Robyn, and PyMC Labs ships PyMC-Marketing — three free, production-grade libraries that together removed the six-figure consulting engagement which once restricted MMM to large enterprises. The IAB published a vendor-neutral Modernizing MMM guide in December 2025, which is a reasonable signal the discipline has re-entered the mainstream.

4. Europe is unusually good for geo experiments

This is the point I would emphasise most to any European performance marketing team, and it is rarely made.

Geographic incrementality testing requires clean test and control regions where media can be bought and withheld independently. Most large markets have to construct those splits artificially from metro areas. Europe has them already: twenty-seven member states plus the UK, with separate languages, separate media buying, separate currencies in several cases, and comparable economies in many pairings.

Operational argument rather than a published finding. Note the caveats: language borders do not always align with national borders, some markets share media and language (Germany and Austria, Belgium and the Netherlands), and pairing countries with different competitive intensity or seasonality will produce a misleading read. Design carefully.

Google announced Meridian GeoX in May 2026, described as an open-source geo-incrementality solution integrating directly into the MMM to run publisher-agnostic geo-experiments. It was framed as newly introduced rather than broadly generally available, so check its current release status before building a plan around it.

5. The independence problem in free tooling

The free libraries are genuinely useful and the caveat is genuinely important: Google and Meta also sell the media their libraries measure.

A model built by a company that sells one of the channels it grades deserves the same scepticism as a platform reporting its own conversions. That is not an accusation. It is a reason to validate against something the vendor did not build.

Sources: 2026 MMM software guidance, including the explicit recommendation to weigh the independence question when a model grades its owner’s own channels. Practical guidance also suggests choosing Meridian where Google Ads is the largest spend channel and geo experiments are planned, and Robyn where Meta is the largest channel — note that this advice points each brand toward the library built by its biggest supplier.

The mitigation is straightforward and it is what Meridian’s own design implies: calibrate the model with experiments rather than trusting it unaided. Reported guidance describes Meridian as built to integrate incrementality results as priors, agnostic of channel — designed to be calibrated by experiments rather than to replace them.

6. What MMM actually requires

The libraries are free. The data is not, and this is where most European ecommerce marketing teams discover the real constraint.

Based on reported MMM data requirements and a build timeline of roughly 3–6 weeks for data collection and preparation, 2–3 weeks modelling, 2–3 weeks validation and 1–2 weeks stakeholder onboarding.

The two-year history requirement is the one that stops most projects. A brand that restructured its channel mix eighteen months ago does not have a usable series yet, and no amount of open-source tooling substitutes for the missing data.

7. Validating a model you built yourself

A model that fits your history is not the same as a model that predicts your future, and an in-house build removes the external party who would otherwise challenge it. Reported practice recommends three validation layers.

Based on reported three-layer validation practice: statistical validation with out-of-sample testing and MAPE under 15%, plausibility checking of channel contributions with experienced marketers, and causal validation via geographic holdout.

The middle layer is the one teams skip because it feels unscientific. It is the layer that catches a model confidently attributing 40% of revenue to a channel that was switched off for two months of the period.

8. Triangulation, and what none of it answers

The 2026 consensus is triangulation: run all three methods for what each is good at, and stop arguing about which is correct.

Based on 2026 guidance that MMM is excellent at budget allocation but blunt tactically, that only incrementality is causal, and that attribution remains useful as a tactical optimisation signal.

The final row is the honest limit, and reported commentary lands on it well: the teams getting this right are not the ones with the most sophisticated model, but the ones who stopped asking their measurement stack a question it was never built to answer and started asking their customers instead. The experiment tells you whether to keep spending. It does not tell you what to change.

9. What this page does not cover

Scope statement. Note also that the reported figure of 46.9% of marketers planning MMM investment comes from US survey data and should not be read as a European adoption rate.

10. The 90-day plan

Indicative sequencing. The holdout starts in week one deliberately: it takes weeks of live running to produce a result, and it is the thing you calibrate the model against rather than an optional extra afterwards.

11. Mistakes to avoid

Recurring errors in measurement rebuilds; illustrative.

12. What changes next

Geo-incrementality tooling is becoming standard. Meridian GeoX, announced in May 2026 as an open-source publisher-agnostic geo-experiment capability integrated into the MMM, points toward experiments becoming a routine input rather than a specialist project — though its release status should be checked before planning around it.

The cookie position has stabilised, not resolved. Google confirmed in April 2025 that it would maintain its existing approach to third-party cookie choice in Chrome without a new standalone prompt, which removes one source of uncertainty while leaving European consent requirements entirely intact.

Consent rules themselves may move. The Digital Omnibus proposes relocating cookie consent into the GDPR, and that negotiation was unresolved as at August 2026 — another reason to favour methods that do not depend on user-level permission.

Key Takeaways

• Attribution and experiment have been documented at 4,100% against -63% on the same spend. Only one is a causal claim.
• Attribution coverage has fallen to 30–60% in some channels, and European opt-in consent degrades it a second time.
• MMM requires no user-level data at all, which makes it structurally suited to a consent-gated market.
• Europe is unusually good for geo experiments — national borders provide ready-made test and control cells.
• The free libraries were built by companies that sell the media they measure. Calibrate with experiments.
• Two years of weekly data is the real barrier, not the software licence.
• No method explains why customers responded. For that, ask them.

Frequently Asked Questions

Conclusion

European measurement is harder than American measurement for a structural reason: opt-in consent removes roughly half the signal before browser restrictions have taken their share. The response most teams reach for — better attribution tooling — is an attempt to repair the method whose foundations were removed.

The better response is to stop depending on user-level data for the decisions that matter most. Marketing mix modelling needs none of it and is now free to run. Geographic experiments need none of it and Europe supplies unusually clean test cells for them. Between those two you can answer how to allocate budget and whether a channel is earning its place, which are the two questions worth being right about. Keep attribution for the daily work it still does well, hold every model to a causal check it did not generate itself, and remember that the most sophisticated stack in European performance marketing still cannot tell you why anyone bought anything.