Marketing Mix Modeling and Incrementality: Meridian, Robyn, Geo Experiments, and Lift Tests

Who this is for

• Advertisers spending across several channels (search, social, video, CTV, retail media, offline) who need to decide where the next dollar goes.
• Brands whose platform-reported conversions add up to more than total sales.
• Teams losing user-level tracking to consent choices and browser limits who need a privacy-safe, aggregate method.
• Finance leaders who want a defensible estimate of marketing ROI with uncertainty ranges, not a single number.

What's included

• Readiness assessment. A written answer on whether MMM, experiments, or both fit your data, spend variation, and decision cadence.
• Data assembly. Weekly spend, impressions, KPI, and control variables (price, promotions, seasonality, distribution) by geography where available.
• Experiment design. Conversion lift, geo experiments, or holdouts with pre-registered success metrics, power estimates, and run dates.
• Model build. Meridian, Robyn, or PyMC-Marketing, chosen for your data and team, with priors calibrated to experiment results.
• Validation. Out-of-sample fit, sanity checks against known events, and stated uncertainty for every channel estimate.
• Budget scenarios. Response curves and a reallocation plan with ranges, plus the experiment we'd run next to test it.

How John approaches incrementality and MMM

Common measurement mistakes John fixes

• Treating platform ROAS as incremental. Retargeting and branded search often look best in-platform and turn out smallest in tests.
• Fitting MMM on flat spend. If a channel's budget barely moved for two years, no model can tell what it did. The fix is to create variation on purpose.
• Trusting one model run. Google's own MMM paper shows that with small samples, priors drive the result and the "optimal" mix carries large variance (Jin et al., 2017).
• Ignoring paid search's special bias. Search ads respond to demand that would have converted anyway. Google's researchers documented the bias this creates in MMM (Chen et al.).
• Underpowered tests. Ad effects are small relative to the noise in sales, so tests need more volume than people expect (Lewis and Rao, 2015). We run a power estimate before committing budget.
• Using observational attribution as ground truth. Large Facebook field experiments found observational methods often failed to recover the lift that randomized experiments measured (Gordon et al., 2019).

When MMM is (and isn't) the right tool

We don't use a spend threshold, because what matters is data, not dollars. These are the criteria we apply, drawn from the Meridian and Robyn documentation and the research they're built on.

Choosing the open-source framework

• Google Meridian (Python): Bayesian and hierarchical across geos, with priors for experiment calibration, reach and frequency inputs, and budget optimization.
• Meta Robyn (R, with a Python port): ridge regression with automated hyperparameter search, returning a set of candidate models the analyst chooses from.
• PyMC-Marketing (Python): a general Bayesian toolkit for teams that want full control of the model specification.

Pricing and engagement

Measurement work is quoted as a flat fee per phase: readiness assessment, experiment design and readout, and model build. Many clients stop after the first two, and that's fine. The scope depends on how many channels and regions are involved and how clean the data is, so we don't publish a single price. Platform lift studies run on your media budget, and we don't charge on top of it. The pricing page has the general terms: no percentage-of-spend billing, and you can stop after any phase.

Platforms we work in

Open-source modeling frameworks, platform lift tools, and the data stack we run them on.

Google Meridian Meta Robyn GeoLift PyMC-Marketing Google Ads Conversion Lift Meta Conversion Lift BigQuery Python R

Logos via the Simple Icons project. Trademarks belong to their owners; no endorsement implied.

Proof you can check before you call

John's published thinking on measurement, attribution, and testing.

Frequently asked questions

Talk to John about incrementality

Tell John which channels you run, roughly how long you've been running them, and what decision you're trying to make. You'll get a straight answer on whether MMM fits.

Top 26 references

The primary sources, standards, research, and tools we rely on for this work. Every link was checked on 2026-10-11. We aren't affiliated with these publishers unless noted.

Official documentation

1. An introduction to Meridian — Google for Developers
   Meridian's design: priors, geo models, reach and frequency, and controls.
2. Meridian: collect and organize your data — Google for Developers
   Data granularity and minimum history recommendations.
3. Meridian: geo-level modeling — Google for Developers
   Choosing geos and handling national-only media.
4. About Bayesian methodology in Conversion Lift — Google Ads Help
   The $5,000 minimum and conversion thresholds for lift results.
5. Incrementality testing improvements — Google Ads Help
   Google's announcement of lower lift-test spend thresholds.
6. Set up Conversion Lift measurement — Display & Video 360 Help
   Eligible campaign types and the 7-to-56-day study length.
7. Robyn key features — Meta (Robyn docs)
   Ridge regression, Nevergrad search, and experiment calibration explained.

Research & studies

8. Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects — Google Research (Jin et al., 2017)
   The foundational Bayesian MMM paper on adstock, saturation, and prior sensitivity.
9. Geo-level Bayesian Hierarchical Media Mix Modeling — Google Research
   Why pooling across regions tightens estimates; the basis of Meridian's geo model.
10. Media Mix Model Calibration With Bayesian Priors — Google Research
    How experiment results become priors that calibrate an MMM.
11. Challenges and Opportunities in Media Mix Modeling — Google Research (Chan and Perry, 2017)
    An honest list of what MMM can and can't identify.
12. Bias Correction for Paid Search in Media Mix Modeling — Google Research
    Why search ads look better than they are in naive models.
13. Measuring Ad Effectiveness Using Geo Experiments — Google Research (Vaver and Koehler, 2011)
    The original Google paper on randomized geo experiments.
14. Estimating Ad Effectiveness using Geo Experiments in a Time-Based Regression Framework — Google Research (Kerman, Wang, and Vaver, 2017)
    The time-based regression method for analyzing geo tests.
15. Robust Causal Inference for Incremental Return on Ad Spend with Randomized Paired Geo Experiments — Google Research (Chen and Au)
    Trimmed match, a robust estimator for paired geo tests.
16. The Unfavorable Economics of Measuring the Returns to Advertising — Quarterly Journal of Economics (Lewis and Rao, 2015)
    Why ad experiments need far more volume than most teams expect.
17. A Comparison of Approaches to Advertising Measurement — Marketing Science (Gordon et al., 2019)
    Large Facebook experiments compared with observational methods.

Leading tools

18. google/meridian — Google (GitHub)
    The official open-source Meridian repository.
19. facebookexperimental/Robyn — Meta (GitHub)
    The official Robyn repository for R and Python.
20. GeoLift — Meta
    Meta's open-source synthetic-control geo testing package.
21. google/trimmed_match — Google (GitHub)
    Google's library for designing and analyzing paired geo experiments.
22. google/matched_markets — Google (GitHub)
    Google's library for picking matched test and control regions.
23. Introduction to Media Mix Modeling — PyMC-Marketing
    A fully Bayesian alternative for teams that want custom models.

Expert guides

24. An analyst's guide to MMM — Meta (Robyn docs)
    Meta's practical walkthrough from data collection to budget allocation.
25. Causal Inference: The Mixtape — Scott Cunningham
    Free textbook on synthetic control, difference-in-differences, and experiments.
26. Forecasting: Principles and Practice (3rd ed.) — Hyndman and Athanasopoulos
    Free textbook on the time-series methods MMM builds on.

Related services