After iOS App Tracking Transparency, cookie restrictions, and consent requirements, platform-reported attribution became both noisier and more generous to the platforms reporting it. Many direct-to-consumer brands now see the sum of conversions claimed by Google, Meta, and TikTok exceed actual orders.
That gap is the reason marketing mix modeling and incrementality testing moved from enterprise luxuries to mainstream tools. Attribution models themselves are covered in What Is an Attribution Model?.
Three Methods, Three Questions
| Criterion | Attribution | Incrementality test | Marketing mix model |
|---|---|---|---|
| Data level | User and click | Test and control groups | Weekly or regional aggregates |
| Proves causation | No | Yes | Estimates it |
| Speed | Real time | Weeks | Months |
| Scope | Usually one platform | One campaign or channel | All channels, including offline |
| Hurt by signal loss | Heavily | Lightly | Barely |
Incrementality: Measuring the Counterfactual
An incrementality test randomly splits an audience or set of regions. One group sees ads and the other doesn’t. The difference in conversion rate between the groups is the incremental lift, meaning the conversions the ads actually caused.
Google Conversion Lift requirements
Google Ads offers Conversion Lift based on users or geography. According to Google’s documentation:
- The account needs at least one compatible conversion action.
- Studies should run for at least 14 days. Google observed up to a 17% drop in absolute lift for studies with long conversion lag that ran shorter.
- When study power is below 90%, Google provides budget guidance, and longer duration or higher budget increases power.
- Results come with a certainty level between 50% and 95%.
Meta offers its own lift studies, so the same logic can be applied to Meta campaigns.
Geo experiments
Geo experiments assign regions rather than users, which makes them independent of cookies, logins, and device IDs. In the US, the usual design matches designated market areas (DMAs) with similar sales history and seasonality, then turns spend off or up in one set. Very large markets such as New York or Los Angeles rarely have a comparable match and are often held out of the test. Google’s open-source Meridian GeoX library is built for running these experiments across advertising platforms.
Marketing Mix Modeling: The Top-Down View
An MMM uses weekly spend by channel, total sales, and non-media factors such as seasonality, pricing, and promotions to estimate each channel’s contribution. It never touches user-level data.
Two open-source options dominate:
- Meridian from Google uses Bayesian methods and produces ROI estimates, response curves, and budget optimization. Experiment results can be fed in as priors.
- Robyn from Meta is an open-source MMM package built around automated hyperparameter search and budget allocation.
Response curves are the part attribution can’t replicate. They show diminishing returns, meaning how much less each additional dollar in a channel produces, which is the real answer to “what happens if we scale this?”
MMM Data Requirements
Models need long, varied history. Two or more years of weekly data is a common starting point, and each channel’s spend must have moved over that period. A channel that spent the same amount every week gives the model nothing to learn from. For the model to distinguish seasonal cycles, longer history is essential, and the data requirements grow with the number of channels being analyzed.
A Practical Cadence for Mid-Size Brands
Running all three continuously is expensive. A realistic rhythm looks like this:
| Frequency | Activity |
|---|---|
| Daily and weekly | Platform attribution for bids, creative, and budget pacing |
| Quarterly | One lift or geo test on the largest or most questioned channel |
| Twice a year | MMM refresh, calibrated with the latest test results |
| Annually | Budget planning based on MMM response curves |
The tests keep the model honest, and the model keeps attribution in context.
Common Mistakes
Treating retargeting ROAS as real impact is the first. Retargeting reaches people already close to buying, and lift tests often show far less incremental value than attribution reports.
Ending tests early is the second. Short tests miss conversions that arrive after the exposure window.
Reading MMM output as fact is the third. An uncalibrated model carries its assumptions straight into its recommendations.
Summary
Attribution shows what customers touched, not what caused the sale. Incrementality tests isolate causal impact with control groups. Marketing mix models estimate every channel’s contribution from aggregate data and show diminishing returns. Brands that combine regular tests with a calibrated MMM get attribution’s speed without trusting its inflated claims.