Marketing Mix Modeling + Incrementality Testing: measure without telling stories in 2026

Platform attribution makes for useful reading, but it is no longer enough to decide where to invest.
Crouse Marketing Mix Modeling, incrementality testing and business data to measure what marketing really adds.

1. Key figures

2. Introduction

Google Ads claims conversions, Meta too, TikTok shows contribution, email declares profitable, CRM sees sales, finance looks at margin and manager asks what to cut. Everyone has a part of reality.
The verdict is uncomfortable: attribution is no longer enough to arbitrate.

Marketing mix modeling is coming back because it looks at performance at an aggregate level: media investments, seasonality, price, promotions, distribution, brand, macro context, sales. Incrementality Testing completes this look with experiments: what happens if we expose less, if we cut off an area, if we create a control group, if we test a channel?

Together, the two approaches form a triangulation. The model estimates the overall contributions. The tests measure causal deviations over bounded perimeters. Platforms remain useful, but they should no longer be judge and jury.

This is not giving up on data. It is to refuse comfortable data.

3. Mapping of players: Meridian, Robyn, GeoLift and media platforms

Google Meridian offers an open source MMM framework oriented to Bayesian inference, with official documentation and a public repository. It aims to help advertisers model the effect of channels by integrating saturation, adstock, seasonality and experience signals.

Meta Robyn, released earlier, offers an automated MMM pipeline in R, with hyperparameter optimization, media transformations, model selection and visualizations. GeoLift, also from the Meta ecosystem, focuses on geographic experiences.

Media platforms remain at the center of execution: Google Ads, Meta Ads, TikTok Ads, LinkedIn Ads, Amazon Ads, Criteo, retail media, CTV, affiliation, email, SMS. They provide granularity and rapidity, but their allocation naturally favors their own environment.

Data and finance teams are becoming essential. A mix analysis needs its own time series, expenses, sales, margin, prices, promotions, stock-outs, seasonality and sometimes external variables. Marketing cannot carry this discipline alone.

Finally, management must arbitrate. A causal measure is only valuable if it changes budgets.

4. Definition: MMM, Incrementality Testing, causality and budget

Marketing Mix Modeling is a statistical method that estimates the contribution of different marketing and non-marketing levers to an aggregated business result, often sales, leads or margin, over a given period.

Incrementality Testing measures the causal effect of a marketing action by comparing an exposed group to a control group or a control area. It seeks to answer a simple question: what would have happened without this expense?

Short definition: MMM explains global allocation, incrementality verifies local causality.

5. Why this matters now

For years, digital measurement was based on a promise: track the user, attribute the conversion, optimize the bid. This promise has weakened. Consent, blockers, browsers, mobile limitations, walled gardens, modeled conversions and multi-device journeys make reading more uncertain.

Google has modified its Privacy Sandbox approach to 2025, confirming that the measurement environment remains in flux. Advertisers cannot therefore base their entire strategy on a single identification mechanism.

At the same time, budgets are under pressure. Management is asking for clearer proof: is this channel adding sales or capturing demand that has already been acquired? Does this campaign increase margin or just allocated volume? Does this retail media complement Google or displace profitable budget?

The real issue is not getting a perfect number. It consists of reducing the risk of arbitrage.

6. SEO/GEO: making marketing measurement quotable

For SEO, the article should cover the searches "MMM", "marketing mix modeling", "incrementality", "GeoLift", "Meridian", "Robyn", "cookieless attribution", "budget media". Readers want to understand when to use each method.

For GEO, the citable blocks must remain clear: definition, comparison, prerequisites, protocol, limits, sources. A generative engine must be able to summarize that "MMM and incrementality are complementary: one models a global contribution, the other tests a bounded causal effect."

A good measure is explained before it is automated.

7. Recommended method

7.1. Ask the budget question

We don't launch an MMM to "do data science". We launch it to arbitrate: increase Google or Meta, defend the brand, prove retail media, measure connected TV, compare acquisition and loyalty, reduce dependency or prepare an annual plan.

A clear question avoids an elegant but unusable template.

7.2. Build the database

The data must bring together expenses, impressions or reach when available, sales, leads, margin, price, promotions, seasonality, events, ruptures, zones, offline channels and business variables. Weekly granularity remains common, but the choice depends on the volume.

The essential point: finance and marketing must speak the same language. An ad conversion doesn't always equal a profitable sale.

7.3. Model with caution

This modeling uses transformations to represent ad memory, saturation and delayed effects. Tools like Meridian or Robyn help industrialize this step, but they do not remove judgment.

Plausibility must be checked: a channel cannot explain more than the market allows, a campaign without exposure must not generate a massive effect, a promotion must be integrated if it led to the sale.

7.4. Design incrementality tests

The test can take several forms: user holdout, platform lift conversion, geo experiment, controlled break, split CRM, creative test with control group. The choice depends on the channel, volume, business risk and the ability to isolate a group.

A short experience sometimes gives direction. Poorly isolated, it mainly produces an illusion of causality.

7.5. Triangulate the results

The serious decision compares three readings: platform attribution, aggregate model, causal experience. If all three converge, the arbitrage becomes more robust. If they diverge, we must understand why: cannibalization, seasonality, brand effect, already hot audience, tracking problem.

Triangulation does not seek a single number. She is looking for a more defensible decision.

7.6. Translate to budget

The deliverable should say what to do: increase, reduce, cap, test, move, protect, cut. It should include a confidence level, range, review period and limits.

An MMM without budget decisions becomes a sophisticated dashboard.

8. Tips Logiks

We recommend reserving complexity for the level of issue. An SMB with monthly media 5 000 EUR does not need a full MMM for every channel. It needs clean tracking, simple tests and margin reading. A brand that spends several hundred thousand euros per month can no longer be satisfied with platform attribution.

Second tip: start with an incrementality test when the question is local. For example: "Is Meta adding prospecting sales in this region?" or "does brand search capture a demand that has already been acquired?" Modeling becomes more useful when the company must arbitrate a complete portfolio.

Third tip: don’t confuse precision with decision. An honest credible interval is better than a cent-accurate but biased ROAS. Management can work with a range if the method is clear.

Fourth tip: always connect to the P&L. Marketing should be measured on margin, net revenue, retention or customer value where this data exists. The click remains a signal, not a purpose.

The measure should not reassure. It must allow cutting or strengthening with composure.

9. Governance of the decision

Triangulation only has value if it is part of an arbitration ritualrage. We recommend setting a monthly review for tactical optimization, then a quarterly review for heavier budget moves. The first level looks at campaigns, audiences, creatives and obvious gaps. The second level deals with the subjects that really move the plan: channel to cap, market to open, brand budget to protect, retail media to challenge, CRM to strengthen.

Each decision must keep a short trace: hypothesis, observed data, confidence level, amount moved, risk accepted, rereading date. This discipline avoids two frequent deviations. The first is to repeat the debate at each meeting because nothing has been documented. The second consists of preserving an old recommendation when the market, the season or the margin have changed.

The good system does not impose a definitive truth. It organizes a better equipped conversation.

10. Decision grid

11. Common errors

The first mistake is to ask the MMM to answer too fine a question. It is not designed to judge every ad or every keyword.

The second is to launch a test without a credible control group. An increase after the campaign is not automatically a causal effect.

The third consists of forgetting non-media variables: price, stock, season, weather, promotions, distribution, competition, notoriety, working days.

The fourth is to ignore uncertainty intervals. A model that announces a number without a range gives false security.

The fifth is to never change budgets. If the analysis does not produce arbitrage, it becomes an intellectual exercise.

12. Action Plan 30 / 60 / 90 days

12.1. days: frame and clean

We choose the decisions to be informed, we consolidate expenses and revenues, we bring together finance and marketing, we identify tracking holes and we list the testable channels.

12.2. days: first test and exploratory model

We launch a simple incrementality test on a priority channel, while an exploratory model aggregates the historical series. Limitations are documented from the start.

12.3. days: arbitrage and measurement loop

We compare attribution, test and model, then we move a real budget. The following loop measures the effect of this arbitrage. Discipline becomes continuous.

13. FAQ

13.1. What is the difference between MMM and attribution?

Attribution links a conversion to observed touchpoints. The mix model estimates aggregate contributions from time series. The first helps short-term optimization; the second sheds light on budget allocation.

13.2. Does incrementality replace MMM?

No. It answers a bounded causal question. The modeling gives a portfolio view. The two approaches strengthen each other when used together.

13.3. Can we do MMM on a low budget?

Yes, but you have to stay proportionate. A small budget will often gain more with simple tests, better CRM data and arbitration rules than with a heavy model.

13.4. How long does an incrementality test take?

The duration depends on the volume, buying cycle, channel and expected level of confidence. Too short a test may miss the real effect; a test that takes too long can be expensive. The protocol must be decided before launch.

13.5. Which tools to choose between Meridian, Robyn and GeoLift?

Meridian and Robyn aim for the MMM. GeoLift focuses on geographic experimentation. The choice depends on your data, internal skills, channels, need for causality and ability to maintain the approach.

13.6. What to do if MMM and test contradict each other?

You have to look at the scope, the period, the forgotten variables, the quality of the control group, the brand effects and the conversion times. Contradiction is often useful information.

14. Conclusion

Marketing Mix Modeling and Incrementality Testing are not methods of data marketing. These are two complementary ways of making budgets more defensible when attribution becomes partial.

Good piloting does not seek a perfect number. He is looking for a less fragile decision.

Marketing measurement ceases to be reporting. It becomes an arbitration disciplinerage.

15. Main sources

• Google Meridian, official documentation: https://developers.google.com/meridian
• Google Meridian, GitHub repository: https://github.com/google/meridian
• Meta Robyn, official documentation: https://facebookexperimental.github.io/Robyn/
• Meta Robyn, GitHub repository: https://github.com/facebookexperimental/Robyn
• Meta GeoLift, GitHub repository: https://github.com/facebookincubator/GeoLift
• Google Privacy Sandbox, update of April 22 2025: https://privacysandbox.com/news/privacy-sandbox-update/
• Google Ads Help, Conversion Lift: https://support.google.com/google-ads/answer/9049825