Google Meridian 2.2.0: Better MMM Calibration, Model Validation, and Budget Optimization
On November 5, 2026, Google released version 2.2.0 of Meridian, its open-source Marketing Mix Modeling (MMM) framework. The update expands model validation and calibration, makes historical media data easier to handle, improves sampling progress monitoring and budget optimization, and fixes issues that could affect the estimated contribution of certain business drivers.
Meridian builds statistical models that estimate how advertising channels contribute to sales and other business KPIs from aggregated data. Unlike user-level attribution, MMM does not require linking every purchase to a particular ad click or user identifier.
More Balanced Holdout Samples for Evaluating MMM Models
The RandomHoldoutSpec class now supports a balanced option, allowing analysts to construct validation samples that balance observations across both time periods and geographic regions.
A holdout sample tests how well a trained model predicts outcomes for observations that were excluded from model fitting. In geographic MMM, a random split can leave validation observations unevenly distributed across periods or markets. The new option addresses that distribution across both dimensions.
This is particularly relevant for advertisers with strong seasonality, regional demand differences, or uneven media activity. The change improves how the evaluation sample is constructed; it does not, by itself, guarantee more accurate forecasts.
New Charts for Reviewing Media Calibration
Meridian adds ReviewSummary.plot_calibration_overview() and ReviewSummary.plot_calibration_details(). The first visualizes the overall calibration picture, bringing together experimental evidence, the model’s prior assumptions, and the posterior estimates obtained after fitting. The second provides detailed views of adjustments for individual channels and experiments used in calibration.
Calibration lets MMM incorporate evidence from independent incrementality experiments when estimating channel effects. For example, results from a geo experiment can help inform the expected return from advertising in a particular channel.
These charts help analysts assess how fitted results align with experimental findings and the original priors. They extend diagnostic reporting rather than introducing a new tool for running experiments automatically.
Separate Time Coordinates for Historical Media Activity
The DataFrameInputDataBuilder.with_media() and with_reach() methods now accept an optional media_time_col parameter. It supports a separate timeline for media activity, reach, and frequency, including observations from before the main modeling window.
Historical media observations matter because advertising may have delayed effects, commonly modeled through Adstock. A campaign can end before the analysis window begins and still influence sales in subsequent weeks.
Previously, passing the additional media history required extending the data window and using missing values for spend. Now spending can remain aligned with the primary time_col, while media activity uses the extended timeline. This simplifies input preparation, particularly for channels with long carryover effects.
Progress Monitoring During Bayesian Sampling
The Meridian.sample_prior(), sample_posterior(), and sample_posterior_and_review() methods add a progress_callback argument, allowing sampling progress to be reported to external monitoring interfaces.
Google also introduces TqdmProgressBar, a ready-to-use progress display. It can report the computation stage, chain progress, the number of generated samples, and divergent transitions encountered during sampling.
For JAX, posterior sampling progress is updated after each MCMC draw, including adaptation. For TensorFlow, updates are emitted after each batch of chains. This visibility is useful when fitting large models takes substantial computation time and analysts need to monitor progress.
Faster JAX Computation and Budget Optimization
Version 2.2.0 includes changes designed to reduce computation time and memory usage. In JAX, posterior-result reconstruction is compiled once per sample_posterior() call instead of repeatedly compiling separate portions of the work. Related improvements apply to coefficient calculations for media, organic, and other factors during prior sampling.
Google also speeds up the construction of computational grids used by BudgetOptimizer. Meridian precomputes reusable Adstock and Hill components and combines portions of reach-and-frequency channel calculations. Repeated calls to analytical methods, including calculations underlying media response curves across spend levels, have also been optimized.
The result is lower overhead when fitting models and comparing numerous budget-allocation scenarios. The actual speedup depends on model configuration, data size, and hardware.
New Warnings and a Fix for Unallocated Budgets
BudgetOptimizer now checks additional properties of channel response curves. It warns when a response curve is not monotonic within the specified constraints or when the distribution of the slope_m parameter includes values above one.
Under those conditions, the optimization algorithm may not find the globally optimal allocation. The warnings allow analysts to examine the model and constraints before acting on the recommended spending mix.
Google also fixed a case in which the optimizer could stop prematurely, leaving some of a fixed budget unallocated. This happened when the next most attractive spending change exceeded the remaining budget. The algorithm now continues to evaluate other feasible allocation options. The fix makes optimization more reliable without eliminating the need to review assumptions and constraints.
An Important Fix for Non-Media Prior Contributions
One of the most consequential corrections in v2.2.0 concerns the contribution_n parameter. The bug affected models that scaled non-media factors by population and used nonzero baseline values for those factors.
When converting a prior contribution in contribution_n to the coefficient gamma_n, Meridian incorrectly divided the baseline value by population twice. As a result, the intended prior contribution distribution could differ from the contribution calculated by Analyzer.
Google explicitly recommends refitting affected models so their prior assumptions are applied correctly. This matters for MMM implementations that include non-media business variables such as distribution, pricing, or other sales drivers. The bug does not apply to every Meridian model, only to the specified configuration.
Additional Model Review and Reporting Updates
The release also fixes ModelReviewer checks in cases where ROI priors apply to only some types of media channels. Input validation now warns when frequency is below one despite positive reach, or when a geography has zero or negative population.
The data structure returned by Analyzer.expected_vs_actual_data now includes an evaluation_set dimension so analysts can access training, holdout, and full-data results separately.
New options in Analyzer.summary_metrics support separate calculations for prior and posterior metrics. Under certain configurations, this makes it possible to calculate optimal frequency and perform budget optimization without first calling sample_prior(). The release also fixes issues involving ROI calculations, frequency optimization, JAX processing, and selected input data types.
What Meridian 2.2.0 Means for Marketing Measurement
Google Meridian is an open-source framework for Bayesian Marketing Mix Modeling, evaluating the incremental contribution of marketing channels, and optimizing marketing investment. It is designed for marketing analysts, data science teams, large advertisers, and agencies that assess effectiveness using aggregated historical data.
Version 2.2.0 strengthens validation and calibration, makes it easier to model delayed advertising effects, and improves computational efficiency. For organizations already using Meridian, the contribution_n correction deserves particular attention: models using the affected configuration should be refitted before relying on their contribution estimates or budget-allocation recommendations.
Trust Us Transform Your Performance Marketing.
Understand how each marketing investment contributes to business outcomes. We help advertisers audit measurement frameworks, validate incrementality assumptions, and connect marketing mix insights with practical paid media decisions.
Audit Your Marketing Analytics



