Google's Ads DevCast episode on Meridian explains a measurement option for marketing teams: marketing mix modeling, or MMM. Meridian is an open-source framework that advertisers can use to build in-house models of channel contribution and budget allocation.
It can be valuable when user-level attribution is incomplete or inappropriate. It is also a statistical modelling project, not a switch in Google Ads.
What marketing mix modeling does
MMM analyses changes in a business outcome, such as revenue, sales or sign-ups, against media activity over time while accounting for relevant non-media factors. Meridian uses aggregated data rather than cookies or person-level journeys.
The framework is designed to help answer questions such as:
- How much did each channel contribute to the selected outcome?
- What return on investment does the model estimate?
- How might a different budget allocation perform?
Those are modelled estimates with uncertainty, not transaction-level facts.
What Meridian adds
Google describes Meridian as a Bayesian causal-inference framework that can work with geo-level or national data. Its published capabilities include:
- hierarchical geographic modelling;
- media saturation and lag effects;
- reach and frequency inputs for suitable channels;
- search-query volume as a control for underlying demand;
- prior information from experiments; and
- budget optimisation and scenario planning.
The open-source code and documentation make the methodology inspectable. They do not remove the need for modelling judgement.
The 2026 product announcements
Google's May 2026 Ads DevCast highlighted three related surfaces:
- Meridian in GA360, described as an integrated MMM experience using available Analytics 360 data;
- Meridian Studio, a Google Cloud environment intended to support model building, quality checks and scenarios; and
- Meridian GeoX, an open-source approach to geographic incrementality testing and calibration.
Availability, eligibility and product terms can change. Check the official documentation before planning around any announced surface.
Data required for a credible model
At minimum, a model needs a consistent outcome series, media execution data and relevant control variables over a sufficient period. Geo-level models require comparable geographic observations.
Before modelling, test:
| Data area | Questions |
|---|---|
| Business outcome | Is the KPI measured consistently across the whole period? |
| Media | Are spend or exposure definitions comparable by channel and week? |
| Geography | Are regions stable and large enough for meaningful variation? |
| Controls | Are pricing, promotions, seasonality and demand factors represented? |
| Experiments | Do prior tests measure the same outcome and timeframe as the model? |
| Governance | Can assumptions, transformations and exclusions be reproduced? |
Meridian's documentation recommends exploratory data analysis before modelling to find missing values, anomalies and problematic correlations.
Calibration and causal limits
Meridian can use experimental results or domain knowledge as priors. Calibration can keep estimates within a plausible range, but inappropriate prior information can also bias the result.
Google's documentation is explicit that causal inference is difficult to validate directly without well-designed experiments. A model with good predictive fit is not automatically a good causal model. Important confounders, incorrect controls and weak variation can produce misleading ROI estimates.
This is why a qualified analyst should document uncertainty rather than presenting one output as exact channel truth.
When Meridian is a good fit
It may suit an advertiser that has:
- meaningful spend across several channels;
- a consistent weekly or geographic outcome series;
- enough history and variation to estimate effects;
- access to statistical and data-engineering expertise; and
- decisions large enough to justify the modelling cost.
It may be premature when conversion tracking is unreliable, budgets are small, campaigns change constantly, or the business lacks enough observations. In those cases, first repair analytics, CRM outcomes, platform experiments and basic reporting.
A practical adoption path
- Define the budget decision the model needs to support.
- Audit the outcome, media and control data before choosing software.
- Run exploratory analysis and document transformations.
- Build a baseline model and review diagnostics and uncertainty.
- Calibrate with comparable experiments where available.
- Compare recommendations with commercial constraints.
- Test material budget changes incrementally instead of treating the model as an instruction engine.
Meridian expands the measurement toolkit. It does not eliminate uncertainty or replace business judgement.
Sources
- Google Ads Developer Blog: The “What” and “Why” on Meridian
- Google for Developers: Meridian
- Google Meridian open-source repository
- Meridian documentation: Assess model fit and results
- Meridian documentation: Exploratory data analysis