The marketing mix modeling papers worth reading fall into three groups: Google’s 2017 Bayesian MMM series (Jin, Chan and Perry, Sun, Wang and colleagues), the field-experiment papers that show how hard advertising is to measure (Lewis and Rao, Blake, Nosko and Tadelis, Gordon and colleagues at Facebook), and recent work on calibration and identifiability. This page is an annotated bibliography of 30 of them, each with a summary in plain English, a note on why it matters in practice, and a link to the library page that uses it.
I put this list together because most “MMM white paper” searches end at a vendor’s gated PDF, while the research the vendors build on is free and surprisingly readable. I’ve built marketing mix models for brands like Monzo and Booking.com and advised Google and Recast on MMM, and these are the papers I find myself pointing people to when they ask why a model said what it said. They’re also the papers behind the defaults in Robyn, Meridian and PyMC-Marketing, so reading them is the fastest way to understand what your library is assuming on your behalf. The filterable list is further down; the sections before it explain how the papers fit together.
Where to start if you only read five marketing mix modeling papers
If you have an afternoon, read Chan and Perry (2017) first. It’s 16 pages, it was written by Google statisticians rather than salespeople, and it lists everything that makes MMM results unreliable: too few data points, correlated channels, a narrow range of historical spend, selection bias from targeting, and the uncomfortable fact that several equally well-fitting models can recommend completely different budgets. It’s the paper I’d hand to a CFO who has been told an R² of 0.98 means the ROAS is right.
Then read Jin et al. (2017), which is the model itself: delayed adstock, Hill saturation, Bayesian estimation and the definitions of ROAS and marginal ROAS that every Google tool since has used. Its simulations are the useful part, because they show that two years of weekly national data leave the response curves biased downwards and the priors doing much of the work. After that, Sun et al. (2017) shows the main way out of that hole (geo-level data with partial pooling), which is the architecture Google Meridian is built on today.
The other two on my shortlist aren’t MMM papers at all. Lewis and Rao (2015) explains why even large randomized experiments struggle to measure advertising returns precisely, which tells you how much (or how little) to expect from a regression on 104 weekly data points. Gordon et al. (2019) takes 15 Facebook experiments, throws away the randomization, and shows that sophisticated observational methods often fail to recover the experimental answer. Together they’re the case for calibrating your MMM with experiments rather than trusting it on its own.
The Google media mix modeling papers, and how they build on each other
Google published four MMM papers in 2017 (the Jin and Wang PDFs are dated April 2017), written by an overlapping team (Yuxue Jin, Yueqing Wang, Yunting Sun, David Chan, Jim Koehler and Mike Perry), and they’re best read as one project. Chan and Perry set out the problem. Jin and colleagues proposed a Bayesian model with flexible carryover and shape effects for a single brand’s national data, then showed it needs more information than a typical dataset contains. Sun and colleagues supplied that information from regional data, and Wang and colleagues supplied it from other brands in the same category, both with hierarchical Bayesian models that share strength across groups. Zhang and Vaver’s simulator paper, published the same year, gave the team a way to test all of this against a known truth.
The series continued in 2018 with Chen et al. on paid search bias, which uses a causal diagram to justify adding search query volume as a control, so that the model stops giving search the credit for demand it merely captured. Then there was a gap until Google’s Meridian team published two papers by Yingxiang Zhang and colleagues: one in 2023 adding reach and frequency to the geo-level model, and one in 2024 that rewrites the model so ROAS is a parameter you can put a prior on directly. If you’ve wondered why Meridian asks for ROI priors rather than coefficient priors, that second paper is the answer.
It’s worth keeping in mind that none of the 2017 papers were published in a journal; they live on Google Research’s site, which is why you’ll see them cited as “Google Inc.” The LightweightMMM README cites Jin, Chan and Perry, and Sun as the basis for that library, and it now points everyone to Meridian (the repository was archived in January 2026), whose own reference list cites most of the series (Jin, Sun, Wang, Chen and both Zhang papers). So when someone says “the Google MMM paper”, they usually mean Jin et al., but the honest answer is that it’s a set of six or seven papers that only make sense together.
Why the experiment papers belong on an MMM reading list
An MMM is an observational model, so its biggest weakness is that it can’t tell spend that caused sales apart from spend that followed demand. The experiment papers on this list are the evidence for how big that problem is. Blake, Nosko and Tadelis (2015) switched off eBay’s paid search and found the returns were a fraction of what non-experimental estimates had suggested, mostly because the people clicking brand ads would have bought anyway. Shapiro, Hitsch and Tuchman (2021) estimated TV effects for 288 brands and found them much smaller than the published literature, with negative marginal ROI for more than 80 percent of brands.
The design papers matter just as much, because they’re how you generate the ground truth an MMM needs. Vaver and Koehler (2011) is the template for geo lift testing, Kerman, Wang and Vaver (2017) adapts it for when you only have a few regions, and Chen and Au (2019) adds an estimator that trims outlying pairs for when one odd market would otherwise swing the answer. The CausalImpact paper (Brodersen et al., 2015) and the synthetic control papers by Abadie and colleagues cover the cases where you can’t randomize at all and have to build a counterfactual from markets that weren’t treated. Johnson, Lewis and Nubbemeyer (2017) explains the “ghost ads” logic behind platform conversion lift studies.
Recent MMM research (2023-2026): calibration, identifiability and better benchmarks
The most useful recent work is about what MMMs can’t learn from data alone. Dew, Padilla and Shchetkina’s 2024 paper, provocatively titled “Your MMM is Broken”, shows that diminishing returns and effectiveness that drifts over time often can’t be told apart in standard MMM data, and that the two explanations imply very different budgets. Their remedy is experiments that move spend deliberately, which lines up with Zhang et al.’s calibration paper and with Robyn’s own approach: Runge, Skokan, Zhou and Pauwels (2024) describe how Robyn trades off model fit against decomposition distance and calibration error, and suggest switching those objectives on one at a time and comparing.
The other strand is better test data. Heusch (2026) publishes a synthetic dataset where spend isn’t random but reacts to promotions, quarterly results and recent performance the way real budgets do, which is a much harder and more honest exam for an MMM library than the usual random spend. On LLMs, I couldn’t find peer-reviewed or even well-documented work that uses language models to estimate media effects. The closest I found is an LLM agent simulation of shoppers (Chu et al., 2025), which is about testing a promotion before launch rather than measuring one afterwards. I’ve included it so you can see where the line sits, and the AI and MMM page covers what agents are actually useful for today.
The full annotated bibliography
-
1979
Aggregate Advertising Models: The State of the Art
Little reviewed the aggregate models of the day and lined them up against what empirical studies had actually observed: sales rising and falling at different speeds when advertising goes up or down, response curves that can be concave or S-shaped, positive sales at zero advertising, competitive effects and effectiveness that drifts over time. His verdict was that the popular structural forms missed some of these, most notably different rise and decay rates.
Why it matters in practice: It’s a checklist that still works. If your model can’t produce an S-curve, a non-zero baseline or changing effectiveness, it has assumed those away rather than tested them.
FoundationsAdstock & saturation Used in: Saturation curves -
1979
One Way TV Advertisements Work
Broadbent, then at Leo Burnett in London, gave the industry the word “adstock”: the idea that an ad builds a stock of awareness that decays week by week rather than vanishing the moment the flight ends. The original article isn’t online, so the link goes to a reference page that cites it; the decay-rate idea it introduced is what every modern MMM library calls geometric adstock.
Why it matters in practice: Every carryover parameter you set in Robyn, Meridian or PyMC-Marketing is a descendant of this idea, and it’s why a model that ignores last month’s TV gives this month’s search the credit.
FoundationsAdstock & saturation Used in: Adstock -
2003
The Economic Costs of Conflict: A Case Study of the Basque Country
The paper that introduced the synthetic control: to estimate what terrorism cost the Basque Country, the authors built a weighted blend of other Spanish regions that tracked Basque GDP before the conflict, then read the gap afterwards as the effect (about 10 percentage points of per capita GDP).
Why it matters in practice: Swap “region” for “market” and “terrorism” for “a TV flight” and you have the logic behind most matched-market and synthetic-control geo tests.
Counterfactuals & synthetic controlFoundations Used in: Synthetic control method -
2010 Start here
Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program
The formal treatment of synthetic control, illustrated on California’s 1988 tobacco program: a weighted combination of other states stands in for the California that didn’t pass Proposition 99, and inference comes from placebo tests that pretend each donor state was treated instead.
Why it matters in practice: Placebo tests are the part marketers skip. If your “lift” is no bigger than the fake lifts you get from untreated markets, you haven’t measured anything.
Counterfactuals & synthetic control Used in: Synthetic control method -
2011 Start here
Measuring Ad Effectiveness Using Geo Experiments
Google’s blueprint for geo experiments: split a country into non-overlapping regions, randomly assign them to test and control, change ad spend only in the test regions, and estimate the return with a regression of test-period results on pre-period results. The pitch is that this is simple to run and simple to read.
Why it matters in practice: This is the design most geo-lift tools still follow, and the paper is short enough to read before your first test.
Geo data & geo testsField experimentsGoogle MMM series Used in: Geo lift testing -
2012
Periodic Measurement of Advertising Effectiveness Using Multiple-Test-Period Geo Experiments
The follow-up: rotate regions between test and control across several periods, so you get a series of measurements instead of one. Pooling them tightens the confidence interval, or lets you reach the same precision with a smaller, cheaper change in spend.
Why it matters in practice: If leadership won’t accept a big holdout, smaller rotating ones are the honest compromise, and they feed a calibrated MMM with a steady supply of experiment results.
Geo data & geo testsField experimentsGoogle MMM series Used in: Incrementality testing -
2015
Inferring Causal Impact Using Bayesian Structural Time-Series Models
The CausalImpact paper. A Bayesian state-space model forecasts what the outcome would have done without the intervention, using control series that weren’t affected, and the gap between forecast and reality is the effect, with an interval that grows over time. The authors demonstrate it on the effect of an online ad campaign on site visits, and ship it as the CausalImpact R package.
Why it matters in practice: It’s the default tool for “we launched in one market, did it work?” questions, and understanding its assumptions (stable relationships with the control series) stops you over-reading it.
Counterfactuals & synthetic controlGeo data & geo tests Used in: Synthetic control method -
2015 Start here
The Unfavorable Economics of Measuring the Returns to Advertising
Across 25 large field experiments with US retailers and brokerages, the authors show that individual sales are so noisy relative to ad spend per person that even huge tests produce wide intervals on ROI (the median interval was over 100 percentage points wide). Informative tests can need more than 10 million person-weeks.
Why it matters in practice: If experiments struggle to see the signal, a regression on 104 weekly points will struggle more. Read it before you promise anyone a precise ROAS.
Field experimentsFoundations Used in: Lewis & Rao (2015) -
2015 Start here
Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment
eBay switched off paid search in experiments and found that returns were a fraction of what non-experimental estimates claimed. Brand keyword ads showed no measurable short-term benefit, and for non-brand keywords the people who responded were new and infrequent users, while frequent users (who would have bought anyway) soaked up most of the spend.
Why it matters in practice: It’s the cleanest demonstration that search clicks follow purchase intent, which is exactly the selection bias that makes MMMs overcredit search.
Field experimentsSelection bias Used in: Incrementality -
2016
Causal Inference in Economics and Marketing
Google’s chief economist writes a short primer for machine learning people: the critical step in any causal analysis is predicting the counterfactual, what would have happened without the treatment, and good prediction tools can help with that step.
Why it matters in practice: It’s the most readable bridge between “my model fits well” and “my model tells me what spend caused”, which is the gap most MMM arguments fall into.
Counterfactuals & synthetic controlFoundations Used in: Causal inference for marketers -
2017 Start here
Challenges and Opportunities in Media Mix Modeling
An unusually candid list of what goes wrong: too few data points for too many parameters, correlated channels, a narrow range of historical spend, selection bias from targeting and seasonality, funnel effects, and five equally good models that recommend different budgets. It ends with remedies: Bayesian priors, category and geo data, search-query controls, graphical models and simulation.
Why it matters in practice: It’s the best single document to hand a stakeholder who thinks an R² of 0.98 means the ROAS is right.
Google MMM seriesSelection biasFoundations Used in: Chan & Perry (2017) explained -
2017
Ghost Ads: Improving the Economics of Measuring Online Ad Effectiveness
“Ghost ads” log the control-group users who would have seen your ad, so you can compare like with like instead of comparing everyone. Against intent-to-treat and public service announcement tests, the method measures lift as precisely at a fraction of the cost; a retargeting campaign lifted site visits by 17.2 percent and purchases by 10.5 percent.
Why it matters in practice: It explains what platform lift studies are doing under the hood, and why an intent-to-treat result looks smaller than a treatment-on-treated one.
Field experiments Used in: Conversion lift studies -
2017
Estimating Ad Effectiveness using Geo Experiments in a Time-Based Regression Framework
Time-based regression (TBR) for geo tests with only a handful of regions, such as smaller countries or matched-market tests, where the earlier geo regression struggles. It regresses the test group’s time series on the control group’s during the pre-period, forecasts the counterfactual through the test, and reads the cumulative effect at the end.
Why it matters in practice: Most real geo tests outside the US have few usable regions. TBR is the method built for that situation.
Geo data & geo testsField experimentsGoogle MMM series Used in: Geo lift testing -
2017
Introduction to the Aggregate Marketing System Simulator
Describes AMSS, a simulator (with an open-source R package) that generates aggregate marketing data from a population split into segments by six “mindset” states, such as whether people are in the market, how far along they are towards purchase and how they feel about the brand. Ads, pricing, competitors and seasonality move people between segments, and because the truth is known (including ROAS and marginal ROAS), you can check whether an MMM recovers it.
Why it matters in practice: Simulated ground truth is the only way to see whether your modelling process, not just your model, gets the right answer. Chan and Perry suggest giving vendors simulated data as part of an RFP.
Google MMM seriesSimulation & validation Used in: MMM validation -
2017
A Hierarchical Bayesian Approach to Improve Media Mix Models Using Category Data
Pool several brands in the same category into one hierarchical model, because different brands advertise independently and so add variation a single brand lacks. The results can be used directly or passed to a single brand’s model as informative priors, which in the authors’ simulations and case studies improved estimates and reduced uncertainty.
Why it matters in practice: It’s the formal version of what agencies and vendors do informally with benchmarks: borrow strength from similar brands, but say so in the priors.
Google MMM seriesBayesian MMM Used in: Bayesian MMM -
2017 Start here
Geo-level Bayesian Hierarchical Media Mix Modeling
Fit the model on regions instead of the national total, with each region’s media effect drawn from a shared distribution (partial pooling). In simulations the geo model had tighter intervals and far less bias in ROAS than a national model, helped somewhat with targeting bias, and got worse when geo-level spend had to be imputed from national totals.
Why it matters in practice: This is the architecture Meridian runs on. If you have regional sales and spend, it’s the most practical way to get more information out of the same two years.
Google MMM seriesBayesian MMMGeo data & geo tests Used in: Geo-level hierarchical MMM explained -
2017 Start here
Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects
The paper that defined the modern Bayesian MMM: delayed adstock with a peak, a Hill curve for diminishing returns, ROAS and marginal ROAS computed from posterior draws, and simulations showing that two years of weekly national data leave the response curves underestimated and the priors doing much of the work.
Why it matters in practice: LightweightMMM cited it as its basis and Meridian’s reference list still does. If you want to know why your library has the parameters it has, start here.
Google MMM seriesBayesian MMMAdstock & saturation Used in: Jin et al. (2017) explained -
2018
Bias Correction For Paid Search In Media Mix Modeling
Paid search ads are shown to people who are already searching for you, so search spend tracks demand. Using a causal diagram of the search environment and the back-door criterion, the authors show that controlling for relevant search query volume can give an approximately unbiased search ROAS in common scenarios, and compare the result with randomized experiments.
Why it matters in practice: It’s why Meridian lets you add Google Query Volume as a control, and the first thing to try when your MMM says branded search is your best channel.
Google MMM seriesSelection bias Used in: Confounding variables in MMM -
2019
Robust Causal Inference for Incremental Return on Ad Spend with Randomized Paired Geo Experiments
Introduces Trimmed Match, an estimator for iROAS from paired geo experiments that trims the most unusual pairs so a few strange regions don’t dominate. It’s distribution-free, comes with a data-driven way to choose how much to trim, and is checked against assumption violations and real case studies.
Why it matters in practice: Geo tests usually have a small number of very different regions. Trimmed Match is the option to reach for when one big market would otherwise swing the answer.
Geo data & geo testsField experimentsGoogle MMM series Used in: Geo lift testing -
2019 Start here
A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook
The authors took 15 Facebook ad experiments (around 1.6 billion impressions), threw away the randomization, and asked whether matching, regression and propensity-score methods on rich user data could recover the experimental answer. They often couldn’t, even with extensive demographic and behavioral controls, and the errors were frequently large.
Why it matters in practice: It’s the strongest evidence that “we controlled for everything” isn’t the same as an experiment, which is the case for calibrating an MMM with lift tests.
Field experimentsSelection bias Used in: Facebook experiments vs observational methods -
2021
TV Advertising Effectiveness and Profitability: Generalizable Results From 288 Brands
Estimating TV advertising elasticities for 288 brands with a consistent method, the authors find effects much smaller than the published literature, many statistically insignificant or negative. Marginal ROI was negative for more than 80 percent of brands, and the overall ROI of the observed schedule was positive for only a third.
Why it matters in practice: It’s a sobering prior. If your MMM says TV returns are spectacular, this is the benchmark that asks you to check your controls and your publication bias.
FoundationsSelection bias Used in: Marginal ROAS -
2021
Bayesian Time Varying Coefficient Model with Applications to Marketing Mix Modeling
Uber’s approach to letting channel effectiveness change over time: coefficients are weighted combinations of local latent variables in a hierarchical Bayesian model, fitted with stochastic variational inference. The authors implemented it in Uber’s open-source Orbit package.
Why it matters in practice: Most MMMs assume a channel is equally effective in 2024 and 2026. If your creative, targeting or market changed, this is one principled way to relax that.
Bayesian MMMOpen-source tools Used in: Open-source MMM libraries -
2023
Bayesian Hierarchical Media Mix Model Incorporating Reach and Frequency Data
Extends the geo-level hierarchical model so that channels with reach and frequency data enter as reach and frequency rather than raw impressions. The model can then say not only how much a channel returns but what frequency is optimal, for any channel where those metrics exist.
Why it matters in practice: This is the methodology paper Meridian’s own documentation points to for its reach and frequency feature, and it matters most for YouTube and TV buyers.
Google MMM seriesBayesian MMMGeo data & geo tests Used in: Reach and frequency -
2024
Packaging Up Media Mix Modeling: An Introduction to Robyn’s Open-Source Approach
The closest thing Robyn has to a methods paper, written by people from Meta’s marketing science team and academia. It walks through the design choices (ridge regression, Prophet for seasonality, Nevergrad multi-objective search that trades off fit against DECOMP.RSSD and calibration error against experiments) and frames them as ways to make MMM acceptable inside organizations and to reduce common biases. It suggests starting with fit alone, then switching the other objectives on and comparing.
Why it matters in practice: Robyn’s docs tell you what to click; this tells you why the tool is built the way it is, which helps when you have to defend its output.
Open-source toolsCalibration Used in: Robyn -
2024 Start here
Your MMM is Broken: Identification of Nonlinear and Time-varying Effects in Marketing Mix Models
Shows that saturation and time-varying effectiveness are often not separately identifiable from standard MMM data: a pattern that looks like diminishing returns can also come from a simpler model where effectiveness drifts. The problem is worse with autocorrelated spend, which adstock makes common, and the fix the authors propose is experiments that move spend in ways that pin down the model form.
Why it matters in practice: The two explanations imply very different budgets. If you’re going to act on a saturation curve, this is the paper that tells you to test it first.
Bayesian MMMAdstock & saturationSimulation & validation Used in: Saturation curves -
2024 Start here
Media Mix Model Calibration With Bayesian Priors
Rewrites the model so that ROAS itself is a parameter, which means you can put an experiment’s result directly into the prior on ROAS instead of guessing what prior on a coefficient would imply it. Simulations show the approach reduces bias and uncertainty in posterior ROAS.
Why it matters in practice: It’s the idea behind setting ROI priors in Meridian, and the clearest answer to “how do I actually put my lift test into the model?”
Google MMM seriesCalibrationBayesian MMM Used in: Calibrating MMM with experiments -
2025
Meridian documentation: model specification, priors and calibration
Meridian has no single paper; its technical documentation is the reference. The model specification page sets out the normalized adstock, the Hill function, an option to apply Hill before adstock, and the geo hierarchy with population-level media coefficients and geo-level draws. The reference list cites Jin, Sun, Wang, Chen and the two Zhang et al. papers above.
Why it matters in practice: If you want to know exactly what Meridian assumes, read this page rather than a vendor summary. LightweightMMM’s README points users to Meridian, which Google released as its official Bayesian MMM in January 2025.
Open-source toolsBayesian MMMGeo data & geo testsCalibration Used in: Google Meridian -
2025
LLM-Based Multi-Agent System for Simulating and Analyzing Marketing and Consumer Behavior
Not an MMM paper, but adjacent: generative agents in a sandbox make purchase decisions and talk to each other, and the authors use the setup to test a price-discount scenario before running it for real. It’s an example of LLM simulation as a pre-launch test rather than a measurement tool.
Why it matters in practice: LLM simulations can suggest hypotheses to test; they can’t replace an experiment or a calibrated model. Useful to know where the line sits.
LLMs & AISimulation & validation Used in: AI and marketing mix modeling -
2026
A Synthetic Benchmark Dataset with Endogenous Marketing Spend for Validating Marketing Mix Models
A seeded generator (and fixed reference dataset, 156 weeks and three channels) where spend isn’t random: budgets react to the quarter’s results, ramp up before promotions, arrive as TV bursts and chase recent performance, the way real budgets do. The true weekly decomposition is recorded, and a companion routine simulates go-dark geo tests.
Why it matters in practice: Most synthetic MMM data hands the model exogenous spend, which hides the main difficulty. This is a better exam for any MMM library you’re evaluating.
Simulation & validation Used in: MMM datasets -
2026
PyMC-Marketing: MMM documentation
The guide to PyMC-Marketing’s MMM: configurable adstock and saturation functions, custom priors, time-varying intercepts and media effects via Gaussian process approximations, calibration with lift tests and budget optimization. It’s documentation rather than research, but it’s where many teams first meet these ideas in Python.
Why it matters in practice: If you want the flexibility to change the model rather than accept a vendor’s defaults, this is the most actively developed open-source option in Python.
Open-source toolsBayesian MMMCalibration Used in: PyMC-Marketing
No papers match those filters. Try clearing the topic or widening the years.
How to read an MMM research paper as a practitioner
Most of these papers are written by statisticians for statisticians, but you don’t need to follow every equation to get the value out of them. The first thing I look for is the simulation setup, because it tells you how favorable the test was to the model. Jin et al. note that their simulated signal is as strong as or stronger than most real MMM datasets they had seen, and Sun et al. say in a footnote that their parameters are more favorable to the modeler than is typical. If a method only just works under those conditions, it will struggle on your data.
The second thing is the data size. Chan and Perry point out that three years of weekly national data is only 156 data points, often with 20 or more channels and three or four parameters each, against a rule of thumb of seven to 10 data points per parameter for a stable regression. When a paper’s real-data example has 106 weeks, like Jin et al.’s shampoo brand, and the posterior for a retention rate looks exactly like its prior, that’s the paper telling you the data didn’t contain the answer. Our MMM data requirements page turns that into a checklist.
The third thing is what the paper treats as ground truth. Simulation papers know the truth because they wrote it; experiment papers estimate it with randomization; and observational papers often have no truth at all, only fit. A paper that validates its model on R² or MAPE alone is answering a forecasting question, not the causal one you care about, which is the distinction behind our MMM validation guide. In practice, the papers that have aged best are the ones that were most candid about this.
What the research still doesn’t answer well
The research is thinnest on long-term and brand effects. Almost every paper here models carryover of a few weeks to a few months (Jin et al. cap it at 13 weeks; Sun et al.’s auto case uses three months), while the questions marketers argue about most, like whether brand building pays back over years, need evidence the MMM literature doesn’t really have. Binet and Field’s work is the usual reference there, and it sits on our page about The Long and the Short of It because it’s industry analysis rather than peer-reviewed research.
The other gap is honest benchmarking of the open-source tools against each other on the same hard dataset. Each library’s documentation shows it recovering parameters on data it generated itself, which is necessary but not sufficient. Until there’s a shared benchmark with endogenous spend that the community actually uses, the best you can do is test any library you’re considering on your own simulated data, read the open-source MMM library comparison, and treat every ROAS as an estimate with a range. If you’d rather start from the foundations, the pillar guide to marketing mix modeling and the MMM course are where I’d begin, and the books page covers the longer reads.
