Why MMM came back, and why now
Marketing mix modeling fell out of favour for roughly a decade. Pixel-based attribution was cheaper to run, faster to produce, and gave marketers the granular channel-by-channel credit splits their reporting decks demanded. That was convenient rather than accurate, and the accuracy problem was building quietly underneath it.
Three things broke the convenience story at the same time. Platform walled gardens meant each channel reported its own conversions, and the sum of those reported conversions reliably exceeded the total conversions the business actually recorded. Multi-touch attribution models tried to distribute the credit across touches, but they were still working from the same biased dataset. Then Google retired the Privacy Sandbox APIs, including Topics, Protected Audience, and Attribution Reporting, in October 2025, while keeping third-party cookies alive but in continued decline. The signal environment got noisier, not cleaner.
The result: a gap between what platforms claim and what a CFO can verify. MMM fills that gap because it works from aggregated business outcomes, not from pixels dropped into browsers. It does not care whether users accepted a consent banner. It does not care whether a platform attributes the same conversion three different ways. It asks the older and harder question: holding everything else equal, how much revenue moved when this channel's spend changed?
For marketing teams in Singapore, Australia, the US, Canada, and Malaysia, the pressure has an additional layer. Regulators and boards are asking for measurement that can be explained to a non-technical audience and audited. An output from a Bayesian model with quantified uncertainty ranges is easier to defend than a platform-native attribution report whose methodology is proprietary.
What Meridian changes
Google's Meridian, released as open-source software on GitHub, is a Bayesian MMM framework built on causal inference. The methodological choice matters: Bayesian models produce a probability distribution over outcomes rather than a point estimate. You get a range of plausible effects with quantified uncertainty, not a single number that implies more precision than the data can support.
Meridian uses the No-U-Turn Sampler for Markov Chain Monte Carlo estimation, which is the same family of methods used in academic causal inference work. It is designed to run on geo-level data, meaning it can model how a media change in one region compares to a region where spend was held flat. That geographic variation is what allows the model to isolate causal effects rather than just correlations. It supports reach-and-frequency modeling for video, which is relevant for brands running connected TV or YouTube at scale, and it integrates search query volume as a lower-funnel signal to improve accuracy on channels with longer attribution windows.
The open-source distribution changes who can afford to run serious MMM. Previously, a credible implementation required either a specialized measurement team or an expensive vendor engagement. Meridian moves the barrier from "do we have a team of econometricians" to "do we have the data and the analytical capacity to configure and calibrate a model." That is still a real bar, but it is a different one.
At Google Marketing Live in May 2026, Google announced that Meridian will be integrated into Google Analytics 360, repositioned as what Google called "a new command center for modern measurement." The announcement described the integration as forthcoming, using "soon, you'll be able to" framing, so this is a roadmap commitment rather than a currently available feature. Google also announced Qualified Future Conversions, a Gemini-powered signal linking upper-funnel spend to future sales via brand search, described as eventually integrating with Meridian to improve model accuracy. As of late July 2026, both remain announced but not generally available.
For teams in regulated markets, the open-source architecture carries a governance advantage. You can inspect the methodology. An auditor can see the model specification. A board-level question about "how did you get that number" has a documentable answer that is not "the platform said so."
The honest limits of MMM
The case for Bayesian MMM is real. The hype version of that case papers over constraints that will determine whether a particular team gets a model that helps them or one that produces plausible-looking noise.
Data volume is the first constraint. MMM needs enough variation in spend across channels and time periods to distinguish signal from noise. A brand that has spent consistently on two channels for three months does not have enough variation to estimate reliable effects. The working assumption in measurement practice is that you need two to three years of weekly data, covering multiple spend-level changes per channel. Meridian's geo-level design helps here because geographic variation creates additional observations, but the requirement for adequate variation does not go away.
Calibration is the second constraint. A Bayesian model needs priors, which means it needs assumptions about how channels are expected to perform before the data speaks. Well-specified priors improve estimates. Poorly specified priors can produce confident-looking outputs that reflect the modeler's assumptions more than the data's signal. Calibration with holdout experiments, which is what Meridian's GeoX capability targets, is the standard correction, but it requires the organizational discipline to actually run controlled experiments rather than spending everywhere all the time.
MMM is not real-time. The model describes what happened over a historical period. It cannot tell you whether yesterday's campaign is working. Teams that need in-flight optimization need a different tool for that job. MMM's role is in budget planning, channel mix strategy, and quarterly reallocation decisions, not in weekly bid management.
Finally, MMM cannot separate all effects cleanly. If a major earned media story drove a spike in branded search at the same time a paid campaign ran, the model will allocate some of that organic lift to the paid channel unless the organic event is explicitly modeled. Good implementations include these control variables. Implementations that skip them produce inflated paid media ROI estimates.
What the Analytics 360 integration means in practice
The announced integration between Meridian and Google Analytics 360 addresses one of the genuine friction points in MMM adoption: data assembly. Building the input dataset for an MMM is time-consuming, often accounting for a significant share of total project effort. You need channel-level spend and impression data, business outcome data at the appropriate geographic and temporal granularity, and control variables covering pricing, promotions, and macro indicators. Assembling this from disconnected sources is where many MMM projects stall before the model even runs.
What Google described at Marketing Live (May 2026), subject to the caveat that the feature is not yet live as of late July 2026, is a data layer that brings first-party, cross-channel data and metric signals into one place as the input to Meridian. For teams already running Google Analytics 360 alongside Google Ads and Display and Video 360, this would reduce the assembly step significantly.
The strategic implication is that Google is building a path from its ad measurement ecosystem into the methodology that was previously platform-agnostic. A team that runs its MMM inside the Google stack will naturally center its attribution within that ecosystem. This is not a reason to avoid the tool, but it is a reason to ensure that the model's data inputs include non-Google channels with equivalent rigor, and that any measurement governance review considers where the model lives and who controls the data pipeline feeding it.
For teams in Australia and Singapore with data residency requirements, the question of where the Analytics 360 integration processes data will matter. That detail was not specified in the May 2026 announcement and will need verification once the feature reaches general availability.
Is your data ready for MMM? A quick self-assessment
The three factors that most reliably predict whether an MMM implementation will produce useful outputs are data volume, channel diversity, and measurement discipline. The tool below gives a directional read on where your team stands. It is a planning signal, not a model output.
MMM readiness check
Answer three questions. The result is a directional planning signal only.
Measurement a board can audit
The phrase "board-auditable measurement" is doing specific work here. A board does not need to understand MCMC sampling. What it needs is a measurement methodology where the key assumptions are stated, the data inputs are documented, the outputs carry stated uncertainty, and an independent reviewer can check the work.
Platform-native attribution reports fail this test. They are not auditable by someone outside the platform ecosystem. The methodology is proprietary and can change without notice. The numbers can improve not because marketing performance improved but because the attribution model updated.
Bayesian MMM passes the audit test because the model is open. The prior assumptions are explicit. The outputs are probability distributions rather than point estimates, so the model itself tells you how confident to be. An independent analyst with the data and the model code can reproduce the result. For publicly listed companies or regulated businesses across Singapore, Australia, and the US, that reproducibility has governance value beyond marketing.
There is a coordination requirement that often gets missed. The team running the MMM needs to include people who understand the business context: promotions calendar, pricing changes, product launches. A model that does not control for a major promotional event during the modeling period will attribute the promotional uplift to whichever paid channel happened to be active at the same time. Getting the non-media inputs right is as important as the media data itself.
For teams ready to build this layer, the path connects directly to incrementality testing as a calibration input. An MMM calibrated against holdout experiments, where you can observe what happened to a test group versus a control group during a spend change, is more credible than one calibrated only on observational data. The incrementality testing guide covers how to design experiments that can feed this step. The cookieless measurement and Bayesian incrementality post covers the privacy context in more depth. For teams auditing what the existing analytics layer is producing before adding MMM on top, the GA4 actionable insights post is the right starting point.