Meridian is a free software library that estimates how much each advertising channel contributed to a business outcome, using aggregated spend and sales data rather than records of individual people. It exists because the older way of answering that question, which followed a user from an ad impression to a purchase through cookies and device identifiers, stopped working reliably once browsers restricted cross-site tracking and privacy regulation narrowed what could be collected. Marketing mix modelling (MMM) is the statistical alternative, and Meridian is Google's implementation of it: a Python package published under the Apache 2.0 licence that any advertiser, agency or consultancy can download, inspect and modify.

The output is a set of channel-level figures. For each channel in the model, Meridian produces an estimate of incremental outcome, a return on investment figure, a marginal return figure describing what the next unit of spend would earn, and a response curve showing how the outcome changes as spend rises. Every one of those comes with a credible interval rather than a single number, which is a direct consequence of the Bayesian machinery underneath.

How the model works

The published model specification is a regression in which weekly outcome in each geographic unit is explained by a time-varying intercept, a geo-specific intercept, control variables, non-media treatments such as price, and transformed media variables. Two transformations do the heavy lifting.

Adstock handles carryover, the fact that an impression served in week one can still influence a purchase in week four. It computes a weighted average of current and lagged spend, with weights decaying according to a single parameter bounded between zero and one. Meridian offers two decay shapes, geometric and binomial, and the maximum lag defaults to eight periods.

Hill curves handle saturation, the fact that the tenth million spent on a channel does less than the first. The function takes two shape parameters and produces the familiar flattening response. Meridian applies adstock first and the Hill curve second by default, though the order can be reversed with a single argument.

The model is hierarchical across geographies. Each channel gets a national-level coefficient plus a variance parameter governing how far individual regions are allowed to deviate from it, so a small market with thin data borrows strength from the national estimate rather than producing noise. Trend and seasonality enter through knots, a set of anchor points between which the time-varying intercept is linearly interpolated. Estimation runs through Markov chain Monte Carlo sampling, with convergence checked through R-hat diagnostics.

Priors are where domain knowledge enters, and they are not neutral. Meridian's default prior on channel return is a log-normal distribution with parameters 0.2 and 0.9, which according to Google's documentation implies an expected return of 1.83 before the model sees any data, with 80 percent of the prior mass between 0.5 and 6.0. Analysts who have run incrementality experiments can replace that default with a tighter, experiment-derived prior, which is the mechanism Google calls calibration.

What it takes to run

Meridian is a library, not a service, and the data burden is substantial. Google's guidance recommends geo-level input and treats the national single-geo variant as a fallback. Its worked example puts 12 channels, six controls and eight knots at 26 parameters, and calls two years of weekly data, or four observations per parameter, too low to estimate reliably. A geo model with 105 regions and three years of weekly data yields 16,380 observations, landing between eight and 74 data points per parameter depending on how much the hierarchy shares across regions.

Google states that the model is built for channel-level analysis and advises against campaign-level input, because campaigns with hard start and stop dates break the adstock memory it depends on.

Origin and evolution

The statistical foundations predate the product. Google Research published the papers Meridian builds on in 2017, covering Bayesian methods with carryover and shape effects and geo-level hierarchical media mix modelling, with reach and frequency extensions in 2023. An earlier package, LightweightMMM, implemented part of that research as an experimental project.

Google unveiled Meridian in March 2024 on a limited-availability basis, and deprecated LightweightMMM shortly afterwards. The repository changelog records a first tagged release in July 2024 and version 1.0.0 on January 24, 2025, five days before Google opened the framework to all marketers and data scientists on January 29 following testing with hundreds of brands. That release was announced by Harikesh Nair, Senior Director of Data Science at Google, alongside a certified partner network of more than 20 providers.

The September 30, 2025 update added non-media variables such as pricing and promotions, channel-level contribution priors, and the binomial adstock decay option, and expanded the partner network to 30 providers including Publicis Media, dentsu, Monks, Adswerve, KINESSO and Accenture. A Scenario Planner arrived on February 19, 2026, converting a trained model into a dashboard for people who do not write code, with the open beta guide documenting per-channel spend constraint arithmetic in August 2026.

Three further products were disclosed on May 5, 2026 ahead of Google Marketing Live: a Data Manager Map View, Meridian GeoX for geographic incrementality experiments, and Meridian Studio, a Google Cloud platform for running large numbers of models under governance. At the event on May 20, Google placed Meridian inside Google Analytics 360 and introduced Qualified Future Conversions. A developer episode on May 28 walked through the Hill curve and adstock mathematics and the reasoning behind publishing the code.

Why it matters to the marketing community

Meridian sits where budget decisions get made. Mix modelling returned to prominence from 2024 because it works on aggregate data, and adoption followed: research in November 2025 recorded 46.9 percent of marketers planning to increase mix modelling investment, the highest priority among measurement methods surveyed. IAB Australia's vendor landscape of September 2025 listed Meridian among twelve profiled providers, one of two entries offered as open source in a field otherwise made up of commercial engagements.

The framework also shapes the commercial market around it. Prescient AI positioned its July 2025 launch explicitly against open-source mathematical foundations, and data supply has moved to match, with Amazon's mix modelling API reaching general availability across 14 markets in May 2026.

Limitations and disputes

The most serious criticism is not about Meridian specifically but about what any observational mix model can identify. Advertising budgets are not randomly assigned; they rise when demand is expected to rise, and algorithmic bidding raises spend when conversion rates rise. That creates endogeneity, correlation between spend and the unmeasured drivers of sales, which a regression reads as effectiveness.

An August 2026 preprint by Niklas Heusch put a number on it. A conventionally specified model, using the seasonal controls Meridian, Robyn and pymc-marketing ship as standard, reported 10.61 times return on a synthetic paid search channel whose true return was 4.20. Handed the true confounders, it still read 8.41. Adjusting the priors moved the figure to 11.38 rather than fixing it. Every result is synthetic, a constraint the paper states: ground truth for advertising effectiveness does not exist in observational data.

A 2024 paper by Ryan Dew, Nicolas Padilla and Anya Shchetkina found nonlinear response and time-varying effectiveness frequently indistinguishable in mix model data, with conflation in up to 99 percent of simulation settings under high carryover, and priced one resulting misallocation at 227,000 dollars over 14 weeks.

Governance is the second dispute. CIMM warned on July 31, 2026 that default priors and embedded assumptions in widely adopted open-source frameworks can reconfigure data asymmetry rather than remove it. A PPC Land analysis on April 2, 2026 examined the structural position of a platform that sells media and supplies the allocation model, using Robyn as the case; the same argument applies to Meridian, which ships with Google search query volume as a modelling input.

Cost is the third. Meridian is free, but running it requires data engineers, data scientists and calibration experiments. Google cut the minimum incrementality test budget to 5,000 dollars in November 2025 and reported most advertisers run one or two studies a year, below what rigorous calibration would need. IAB's State of Data 2026, published February 7, 2026, found up to 75 percent of buy-side decision-makers rating attribution, incrementality and mix modelling as underperforming. The IAB argued separately on April 7, 2026 that standard mix modelling is structurally misaligned with retail media.

Distinguishing Meridian from adjacent terms

Marketing mix modelling is the method; Meridian is one implementation of it. Rival implementations include Meta's Robyn, which uses evolutionary hyperparameter search and ridge regression rather than Bayesian sampling, and the community project pymc-marketing.

LightweightMMM was Google's earlier Python package, now unsupported. Google's migration documentation describes Meridian as the official evolution of its approach and notes that Meridian fixes the Hill-adstock order for reach and frequency channels.

Meridian GeoX, Studio and Scenario Planner are separate products in the same family, not the model. GeoX designs and analyses geo experiments, Studio operates models at scale on Cloud, and Scenario Planner is a reporting layer over an already trained model.

Multi-touch attribution assigns credit along an individual user journey. Meridian does the opposite, working from aggregates and treating the causal question as one of incrementality against a modelled baseline.

Recent developments

Meridian GeoX has moved out of the closed beta described in May 2026 into its own repository, with guides, Colab notebooks and an API reference for holdback, go-dark and heavy-up designs. Version 2.0.0 of the core library, dated September 2, 2026 in the changelog, switched the default computation backend from TensorFlow to JAX, added prior calibration from incrementality experiments including GeoX, and introduced channel calibration recommendations flagging which channels would most benefit from a test. Both are breaking changes, so existing pipelines require code updates.

Timeline

  • 2017 - Google Research publishes the Bayesian methods paper on carryover and shape effects, and the geo-level Bayesian hierarchical media mix modelling paper
  • 2021 - Meta releases Robyn, establishing an open-source mix modelling category
  • 2023 - Google Research publishes the hierarchical model extension incorporating reach and frequency data
  • March 2024 - Google unveils Meridian on a limited-availability basis; LightweightMMM is deprecated shortly afterwards
  • July 16, 2024 - the first tagged release, version 0.2.0, appears in the public changelog
  • January 24, 2025 - version 1.0.0 is released
  • January 29, 2025 - Meridian opens globally after testing with hundreds of brands, with a certified partner network above 20 providers
  • September 4, 2025 - version 1.2.0 adds binomial adstock decay, per-channel decay functions and automatic knot selection
  • September 30, 2025 - non-media variables, channel-level contribution priors and binomial decay are announced; the partner network reaches 30 providers
  • November 11, 2025 - Google lowers the minimum incrementality experiment budget to 5,000 dollars
  • December 8, 2025 - version 1.4.0 introduces the modules underpinning the Scenario Planner
  • February 19, 2026 - the Meridian Scenario Planner launches as a code-free budget interface
  • April 29, 2026 - version 1.6.0 adds JAX support and exploratory data analysis checks including a data-to-parameter ratio
  • May 5, 2026 - Meridian GeoX and Meridian Studio are announced ahead of Google Marketing Live 2026
  • May 20, 2026 - Meridian is integrated into Google Analytics 360 at Google Marketing Live 2026
  • August 6, 2026 - the Scenario Planner open beta guide is updated with spend constraint formulas
  • August 21, 2026 - the structural estimation preprint questioning observational mix model identification is posted to arXiv
  • September 2, 2026 - version 2.0.0 makes JAX the default backend, adds prior calibration from incrementality experiments, and introduces channel calibration recommendations

Summary

Who: Google built and maintains Meridian, with Harikesh Nair, Senior Director of Data Science, fronting the public releases and Gaurav Bhaya, VP and GM of Buying, Analytics and Measurement, announcing the 2026 product layers. Users are advertisers, agencies, consultancies and a certified partner network that reached 30 providers in September 2025, including Publicis Media, dentsu, Monks, Adswerve, KINESSO and Accenture.

What: An open-source Python library, licensed under Apache 2.0, that fits a geo-level hierarchical Bayesian regression estimating each channel's incremental contribution, return on investment and response curve from aggregated spend and outcome data, using adstock transformations for carryover and Hill curves for saturation.

When: Unveiled in March 2024 on a limited basis, released globally on January 29, 2025 at version 1.0.0, extended with non-media variables in September 2025, a Scenario Planner in February 2026, Analytics 360 integration in May 2026, and version 2.0.0 with a JAX default backend on September 2, 2026.

Where: Distributed through the google/meridian repository on GitHub with documentation on Google's developer site, run locally or on Google Cloud, and available as a managed enterprise workflow through Meridian Studio and inside Google Analytics 360.

Why: Cross-site tracking restrictions and privacy regulation degraded the user-level joins that multi-touch attribution depended on, leaving aggregate modelling as the practical route to cross-channel measurement. Publishing the code was intended to make the method auditable, though researchers and industry bodies dispute whether any observational mix model can identify causal effects without experimental calibration.