Marketing Mix Modeling (MMM) and Incrementality Testing in 2026: Measuring True Marketing Performance Without Cookies

By Million Marketing · Reading time: 12 min

How much does one euro invested in Paid Social, display or awareness campaigns really return? In 2026, the question is harder than ever for the marketing departments of enterprise accounts and mid-market companies. Since March 2024, Google has required explicit consent to use Google Ads and GA4 in the European Economic Area [1]. In the « basic » Consent Mode configuration, the loss of measurement data can reach 30 to 70% [2]. And the CNIL leaves no room for error: 83 sanctions and €486.8 million in cumulative fines in 2025, including 21 sanctions related to cookies and trackers [3].

In this context, click-based attribution is no longer enough to steer a media budget. Two methods are emerging as the new standard for performance measurement: Marketing Mix Modeling (MMM), which models the impact of each lever from aggregated data, and incrementality tests, which establish causality through experimentation. This guide explains how they work, which open-source tools to favour, and how to deploy them in 90 days.

Why click-based attribution is no longer enough

Deterministic measurement, based on tracking individual users and last-click attribution, relies on a signal that degrades year after year. To keep using remarketing and ad personalisation in the EEA, advertisers must send Google two explicit consent parameters, ad_user_data and ad_personalization [1]. Without the user’s consent, tags do not fire in basic mode; advanced mode sends degraded signals and rebuilds part of the conversions through modelling, but the CNIL has not formally approved it and consent remains the default principle [2]. We detail these technical trade-offs in our article on privacy-first SEA.

The problem is not only quantitative. Conversion modelling allows each platform to optimise its own campaigns, but it does not give a marketing director or a CFO an objective, cross-channel reading of performance. Last click mechanically overestimates bottom-of-funnel levers, such as SEA on brand queries, and underestimates awareness investments, display, video or Paid Social, which generate few immediate clicks. To allocate an overall budget, you need a method that depends neither on cookies nor on personal data.

Marketing Mix Modeling explained simply

MMM is a statistical approach that estimates the contribution of each marketing lever (media spend, promotions, pricing) and of external factors (seasonality, trend, public holidays) to a business indicator: revenue, sales, leads. It works exclusively on data aggregated over time, for example weekly spend by channel. As Qonto’s data team puts it, MMMs « rely on aggregated data rather than individual data » and require no personally identifiable data [12]. This is what makes them GDPR-compliant by design and insensitive to falling consent rates.

Adstock and saturation: the two key transformations

To reflect the reality of advertising, an MMM applies two transformations to media investment [7][9]:

• Adstock (carry-over effect): a campaign running this week continues to influence sales in the following weeks. The model applies a gradual decay of the effect rather than giving all the credit to the week it ran.
• Saturation (diminishing returns): doubling a budget does not double sales. A non-linear response curve identifies the point beyond which each additional euro returns less and less.

Why the Bayesian approach dominates

Modern frameworks favour Bayesian inference. It makes it possible to incorporate prior knowledge (« priors »), for example the result of an incrementality test, and produces estimates with credible intervals rather than a single misleading figure. Google presents Meridian as a model based on Bayesian causal inference [5], and PyMC-Marketing allows custom priors and likelihoods to be defined [9]. Qonto stresses this point: demand confidence intervals and out-of-sample validation rather than relying on R² alone [12].

What MMM measures… and what it does not

MMM measures the incremental ROI of each channel, the effect of offline and online levers on the same indicator, the share of « baseline » sales independent of marketing, and the saturation curves that guide budget reallocation. It does not measure an individual customer’s journey, the performance of a specific creative or day-to-day bid optimisation: these tactical decisions remain the remit of platforms and analytics.

Google Meridian, Meta Robyn, PyMC-Marketing: which tool should you choose?

Long reserved for econometrics consultancies, MMM has been democratised thanks to three major open-source frameworks.

Criterion	Google Meridian	Meta Robyn	PyMC-Marketing
Publisher	Google	Meta Marketing Science	PyMC Labs (Apache 2.0 licence)
Method	Bayesian causal inference, hierarchical geographic or national model	Ridge regression, multi-objective optimisation via Nevergrad	Fully configurable Bayesian modelling (PyMC)
Strengths	GPU acceleration, calibration with experiments and priors, frequency optimisation using reach and frequency data	Automated hyperparameter search, built-in budget allocator, trend/seasonality decomposition	Custom priors and likelihoods, time-varying effects, calibration with lift tests, CLV modules
Limitations	Python 3.11 to 3.13, GPU recommended; ecosystem designed around Google data	Production version in R; Python version still in beta	Requires genuine Bayesian statistical expertise
Sources	[5][6]	[7][8]	[9][10]

Meridian, whose version 2.0.0 is available on GitHub, can model data at geographic as well as national level, accelerate computation on GPU and calibrate the model with experiment results and prior information [5]. It also includes ad frequency optimisation based on reach and frequency data, an asset for advertisers investing heavily in video and YouTube.

Robyn combines Ridge regression, evolutionary algorithms and multi-objective optimisation via Nevergrad, with a budget allocation module. The R version is considered stable, while the Python version is presented as a translated beta [7].

PyMC-Marketing is the most flexible: custom priors and likelihoods, adstock, saturation, time-varying intercepts and media contributions, budget optimisation, and customer lifetime value (CLV) models in the same toolkit [9]. Bolt has documented its use of PyMC-Marketing for its budget allocation decisions [14].

Data required. Whatever the tool, the basic rule remains that of the Robyn guide: « an MMM needs at least two years of weekly historical data », more (4 to 5 years) if only monthly data is available, with a number of observations roughly 7 to 10 times the number of variables [8]. You need to bring together the target variable (sales, revenue, leads), investment or impressions by channel, and control variables: promotions, pricing, seasonality, events. Media exposure must also vary over time: a perfectly constant budget makes it impossible to isolate the effect of a channel.

Incrementality tests: proof through experimentation

However sophisticated it may be, an MMM remains an observational model. To establish causality, you have to experiment. The Robyn guide explicitly recommends it: use experimental results, regarded as ground truth, to calibrate the MMM, and run these studies « continuously and regularly » [8].

Geo-lift and synthetic control

Geographic testing is the most robust method in the face of disappearing cookies, because it does not depend on any individual identifier. The reference open-source tool, GeoLift (Meta), is based on the augmented synthetic control method [11]:

1. the territory is divided into geographic units (cities, regions, catchment areas);
2. the campaign is only activated (or switched off) in the test areas;
3. the algorithm builds a « synthetic twin » of the test areas from a weighted combination of control areas with comparable sales history;
4. the gap between actual sales and the sales predicted by the twin measures the incremental lift.

Statistical power and budget

A poorly sized test proves nothing: the campaign’s effect gets lost in the natural noise of sales. The GeoLift documentation points out that a prospective power analysis is « fundamental » before launching a test [11]. It sets the minimum detectable effect (MDE), the number of test areas and the budget required, expressed as cost per incremental conversion. It also requires a clean pre-test history, with no missing values or incomplete areas [11].

Conversion lift and holdouts

Conversion lift tests, which isolate a control group of users directly within a platform, remain useful, particularly for Meta [8]. But they measure one channel at a time, within the environment and with the data of the platform concerned. Geographic tests have the advantage of measuring the effect on your own sales, across all channels.

Triangulation: combining MMM, tests and attribution

No single method is enough on its own. The most advanced marketing departments practise triangulation, combining three approaches so that each corrects the biases of the others [4]:

• MMM provides the strategic view: contribution of each lever, saturation, overall budget allocation;
• incrementality tests provide causal proof for the levers under debate;
• attribution (analytics, platforms) remains the tool for tactical management: audiences, creatives, bids.

Qonto suggests a simple safeguard: the effectiveness estimated by the MMM should fall between the bounds provided by attribution models and by incrementality experiments [12]. To structure your analytics data upstream, see our guide GA4 and SEA for enterprise accounts.

The calibration debate

How do you feed test results into the MMM? Two schools coexist. The first converts the test result into a prior on the channel’s ROI. The second, advocated by PyMC Labs, adds each test as a likelihood that directly constrains the saturation curve: the more tests you accumulate at different spend levels, the more precise the curve becomes [10]. The stakes are real: in the simulation published by PyMC Labs, an unobserved confounding variable leads an uncalibrated MMM to reverse the ranking of two channels, while the calibrated model recovers ROAS within about 2% of the true values [10].

Caution is still advisable. Qonto’s team recommends using tests to steer the priors moderately, without fixing the parameters: you have to « leave the model the possibility of disagreeing » [13]. A single test gives a direction, not an absolute truth; it is the repetition of tests and their comparison with the model that builds confidence.

B2B and luxury specifics

B2B: long cycles and CRM data

In B2B, the final sale is too rare and too late to serve as the target variable on its own. It is often more relevant to model intermediate pipeline stages from the CRM, such as qualified leads (MQL, SQL) or opportunities created, and then link these stages to revenue. Adstock parameters must reflect longer effects: expert content, events and LinkedIn campaigns produce their effects over several months. Our article on LinkedIn Ads and ABM for enterprise accounts and mid-market companies details how to measure these levers.

Luxury: omnichannel and desirability

In luxury, a large share of sales are completed in the boutique, and brand value is built over the long term. A model limited to e-commerce sales will systematically underestimate out-of-home, print, events or influence. You therefore need to include omnichannel revenue (boutiques, e-commerce, wholesale where the data allows) and track brand indicators (awareness, consideration) separately. Geographic tests are particularly well suited to boutique networks: a local activation can be measured directly on footfall and point-of-sale revenue. On measuring the new retail advertising formats, read our analysis of retail media for enterprise accounts and luxury.

A 90-day roadmap for a mid-market company or enterprise account

Days 1 to 30: audit and data collection

Gather at least two years of weekly history [8]: investment and impressions by channel (Google Ads, Meta, LinkedIn, programmatic, TV, out-of-home), sales or leads from the ERP and CRM, and control variables (promotions, pricing, seasonality, events). Document each source and correct breaks in the series. The quality of this dataset determines the whole project: Qonto points out that data must be « clean, ready to use and well understood across the company » [12].

Days 31 to 60: first model

Choose the framework according to in-house skills and the media mix: Robyn for its automation, Meridian for geographic granularity and frequency, PyMC-Marketing for flexibility. Set the adstock and saturation bounds with the business teams, validate out of sample and present the results with their uncertainty intervals.

Days 61 to 75: first incrementality test

Launch a geo-lift on the lever whose ROI is most debated by the executive committee, often Paid Social, video or influence, after a power analysis [11]. Then feed the result into the model as a calibration constraint [10].

Days 76 to 90: budget reallocation

Use the framework’s budget optimiser to simulate several constrained scenarios, then give the finance department recommendations based on marginal return: the return on the next euro invested, channel by channel. Schedule regular model refreshes and a test calendar for the following quarters [13].

Common mistakes to avoid

• Judging a model on its R² alone: a good historical fit does not guarantee accurate contributions [12].
• Locking a model on a single test: a test points the way, it does not settle the question [13].
• Forgetting control variables: without promotions, pricing or seasonality, the model attributes to media what does not belong to it.
• Delivering an MMM once a year: the value comes from a continuous model → test → reallocation cycle.

Our approach at Million Marketing

At Million Marketing, we approach measurement as a governance project as much as a data science one. We frame the business questions with the marketing and finance departments, audit the quality of media and CRM data, choose the right framework (Meridian, Robyn or PyMC-Marketing) and design a geographic test calendar that feeds the model over time. The recommendations are then activated through our SEA, Paid Social and SEO expertise. For organisations that want to steer these trade-offs at the highest level without recruiting, our part-time marketing leadership offer integrates this measurement framework into strategic management.

Sources and references

1. Digidop — Google Consent Mode V2 : guide complet pour mettre votre site en conformité (2024)
2. EdgeAngel — RGPD, CNIL et Analytics : Consent Mode, IA, DPF (2026)
3. CNIL — Sanctions et mesures correctrices : la CNIL présente le bilan 2025
4. Funnel — Marketing mix modeling, incrementality testing and attribution (Funnel Measure)
5. Google — Meridian, official GitHub repository
6. Google for Developers — About the Meridian project
7. Meta Marketing Science — Robyn, official GitHub repository
8. Meta — An Analyst’s Guide to MMM (Robyn documentation)
9. PyMC Labs — PyMC-Marketing, official GitHub repository
10. PyMC Labs — Media Mix Model Calibration with Lift Tests and Bayesian Priors
11. Meta — GeoLift Walkthrough (official documentation)
12. The Qonto Way (Medium) — Marketing Measurement series: Marketing Mix Modeling at Qonto, Part I
13. The Qonto Way (Medium) — Marketing Measurement series: Incrementality test scheduling
14. Bolt Blog — Bayesian models: Bolt’s story with PyMC-Marketing

Frequently asked questions about MMM and incrementality

Read next

On the same topic

Retail Media in 2026: The Third Pillar of Digital Acquisition for Enterprise Accounts and LuxuryRead the article → LinkedIn Ads for B2B in 2026: Account-Based Marketing, Predictive Targeting and Budget ArbitrageRead the article →ChatGPT vs Perplexity vs Gemini: Which AI Engine to Optimise First for Your 2026 AIO StrategyRead the article →