Google Meridian

Google Meridian is Google's open-source Media Mix Modeling (MMM) framework, built on Bayesian causal inference. It enables advertisers to run their own in-house MMM models to measure the true incremental impact of media spend and optimise budget allocation. Meridian is free to use and open-sourced on GitHub (github.com/google/meridian). (developers.google.com/meridian, as-of 2026-08-14)

What Meridian is

Meridian is a Python library built on TensorFlow Probability (not PyMC or Stan — see Contradictions). It fits a Bayesian hierarchical regression that decomposes a KPI (typically revenue) into contributions from each paid media channel, organic/non-media variables, and a trend-plus-seasonality baseline. The model uses two canonical transformations shared with the broader MMM literature:

Adstock — models the carryover/lag effect of advertising; two weight functions supported: geometric and binomial (developers.google.com/meridian)
Saturation (Hill function) — models diminishing returns; applied before or after Adstock via a configurable hill_before_adstock boolean (pymc-labs.com, 2025-09-08)

Seasonality is modelled via splines (knots), providing a flexible time-varying baseline. Automatic Knot Selection (AKS) was introduced in v1.2.1 to automate knot placement. (pymc-labs.com, 2025-09-08)

As of July 2026, the current version is 1.7.1 and the product surface spans four pillars: (developers.google.com/meridian, as-of 2026-07-31; ppc.land, 2026-05-20)

Pillar	What it does
MMM	Core Bayesian model — fits channel contributions, ROI, mROI, response curves
GeoX	Open-source geo-experiment tool that calibrates MMM priors (not yet GA as-of Aug 2026)
Studio	Enterprise platform for high-volume MMM management on Google Cloud
Scenario Planning	No-code budget optimisation interface launched February 2026

Release timeline

Date	Milestone
March 2024	Unveiled; limited beta
January 29, 2025	Global open-source launch; 20+ certified partners
September 30, 2025	Major update: non-media variables (pricing/promotions), channel-level contribution priors, binomial adstock decay, mROI priors; partners expanded to 30+ (Publicis Media, dentsu, Monks, Adswerve, KINESSO, Accenture)
February 19, 2026	Scenario Planner launched
May 5, 2026	GeoX, Studio, and Data Manager updates announced pre-GML
May 20, 2026	Analytics 360 integration announced at Google Marketing Live 2026
August 5, 2026	Scenario Planner open beta documentation published

(ppc.land, 2026-05-20 — most complete timeline source found)

Key differentiators

1. Google Query Volume (GQV) data

Meridian can ingest Google Query Volume data — a search demand signal — as a control variable for paid search channels. This separates the effect of organic demand from incremental paid search impact, a limitation of most other MMM tools. GQV is sourced via the MMM Data Platform (request form, 3-business-day response) and is indexed by advertiser/country pair. Raw query counts are not provided. (developers.google.com/meridian/docs/faqs, 2026-06-30)

2. Reach and frequency for YouTube

Meridian natively models reach (unduplicated audience) and frequency (average exposures) for YouTube campaigns — essential for upper-funnel video measurement. Most MMMs use spend or impressions as the media input; reach/frequency separates coverage from intensity. (developers.google.com/meridian; Forrester, 2025-01-29)

3. GeoX calibration loop

Geo Holdout Testing#Google Meridian GeoX|GeoX converts geo-experiment results into Bayesian priors that calibrate the MMM. This bridges the experimental/observational measurement gap — lift studies provide ground truth, MMM provides scale and budget planning; GeoX connects them. Announced May 2026; testing begins later in 2026. (developers.google.com/meridian/geox, as-of 2026-05-14)

4. Analytics 360 integration

Announced at Google Marketing Live 2026: unified interface combining first-party and cross-channel data, causal performance measurement, and predictive scenario planning inside Google Analytics 360. (ppc.land, 2026-05-20)

Scenario Planner

Launched February 2026 (open beta August 2026). A no-code dashboard structured in three tabs — Analyze (model diagnostics), Plan (channel spend constraints), Optimize (allocation recommendations). Intended to translate MMM outputs for CMO/CFO audiences.

Key mechanic, per developer relations presenter Jeff (youtube.com, 2026-08-05):

"If a channel makes up 22% of your spend share but drives 34% of your outcome — that's a heavy hitter."

The optimizer "moves dollars away from channels that are flatlining at the top of the [saturation] curve and pouring them into channels that are still in the steep, high-growth phase." Per-channel minimum and maximum spend constraints prevent unrealistic allocations.

Data security note: sharing a Scenario Planner dashboard by "Save and Share" passes the owner's credentials to viewers, giving them access to the owner's underlying data. "Temporary Dashboard link" is the recommended secure sharing method. (youtube.com/watch?v=X3ksrQ1U1SI, ppc.land, 2026-08-05)

Certified partner ecosystem

30+ certified partners as of September 2025, including Publicis Media, dentsu, Monks, Adswerve, KINESSO, and Accenture. (ppc.land, 2026-05-20; martech.org, 2025-09-30)

Named enterprise adopters on the official site: Shopify, Vestiaire Collective, ASOS, Finder, Reserved. (developers.google.com/meridian, as-of 2026-08-14)

Case study: Alaska Airlines migrated from a hybrid attribution stack to Meridian via Adswerve specifically because Meridian can tap YouTube and Google Search data that the previous stack could not access. (adexchanger.com, 2026-07-14)

Limitations and known constraints

Technical constraints (per official documentation)

No time-varying coefficients: channel effectiveness is assumed constant over the modelled period. Time-varying effects are approximated by manually splitting a channel into separate channels for each period — Google cautions this increases parameter count and can cause convergence issues. (developers.google.com/meridian/docs/faqs, 2026-06-30)
No channel synergies: cross-channel interaction effects cannot be modelled. (developers.google.com/meridian/docs/faqs)
Geo-first design: hierarchical model is built for geographic disaggregation; other categorical hierarchies (product, sales channel, campaign) are discouraged because SUTVA and scaling assumptions break. (developers.google.com/meridian/docs/faqs)
Causal inference, not forecasting: Meridian estimates incremental outcome under hypothetical future budgets but cannot forecast raw outcome. (developers.google.com/meridian/docs/faqs)
Qualified Future Conversions circularity risk: a planned integration with Gemini-generated QFC signals (predicted future sales from Google Ads data) has raised "prior contamination" concerns — a Google-derived signal feeding a model that also uses Google Ads spend data. No methodology published yet. (ppc.land, 2026-05-20)

Performance constraints (per PyMC Labs benchmark, September 2025)

Memory: Meridian requires ~5 GB RAM for the smallest synthetic dataset; PyMC-Marketing uses 900 MB–2.5 GB for comparable datasets.
Speed: PyMC-Marketing is 2–20× faster in posterior sampling (NumPyro: 110 ESS/s vs Meridian's 5.66 ESS/s at startup scale).
Accuracy: Channel contribution recovery SRMSE of 0.70 (Meridian) vs 0.41 (PyMC-Marketing) at startup scale — roughly twice the attribution error.
Enterprise-scale convergence: Meridian failed to converge on a 50-geo, 30-channel synthetic dataset; PyMC-Marketing completed it.
Automatic Knot Selection trade-off: AKS (v1.2.1) improves in-sample R² and MAPE but increases residual autocorrelation and "leakage of media signal into the baseline," degrading causal attribution accuracy. (pymc-labs.com, 2025-12)

Conflict-of-interest debate

The dominant critical framing in trade press is the "Trojan Horse" argument:

"There's a huge amount of power in setting where everybody starts in a solution." — Mutinex CEO Henry Innis (adexchanger.com, 2026-07-14)

AdExchanger (July 2026) reported: Google salespeople have "hard KPIs tied to how many advertisers they get on Meridian"; numerous agencies and measurement vendors have built proprietary offerings on Meridian source code "without advertisers knowing"; and Innis proposed disclosure requirements analogous to post-Enron financial reporting, arguing "measurement is most trustworthy when it isn't provided by the same players who stand to benefit from the results." (adexchanger.com, 2026-07-14)

Nick Stoltz (Measured CSO): Google is pushing Meridian partly to upgrade Google Analytics, which is "still trying to shake its reliance on click-based attribution." (adexchanger.com, 2026-07-14)

Mike Ryan (Smarter Ecommerce) described Meridian as having mapped "the entire Googleverse of media and data in loving, intricate detail." (adexchanger.com, 2026-07-14)

Meridian best fit (per Forrester, 2025-01-29):

Heavy Google spend (YouTube/Search integration has most value)
In-house data science, data engineering, and visualisation resources
History of geo-level marketing and sales data
Existing experience with incrementality testing and ROI calculation
NOT recommended for: teams lacking data science resources, or organisations where executive buy-in is the primary constraint ("the education aspect remains unaddressed")

For DTC brands with a Shopify site and strong social presence (no TV, no offline), "a DTC brand might be fine with Meridian and/or a lightweight MMM product from a third-party provider like Northbeam or Triple Whale" — per Prescient AI CEO Mike True, omnichannel advertisers need more. (adexchanger.com, 2026-07-14)

Contradictions

Meridian's backend — TensorFlow Probability (common misconception: PyMC/Stan) A common internet misconception (and one reflected in several third-party blog posts) is that Meridian is built on PyMC or Stan, likely because all three are Bayesian Python libraries. The PyMC Labs benchmark (pymc-labs.com, 2025-09-08) and the GitHub source confirm Meridian is built on TensorFlow Probability (tensorflow_probability). PyMC-Marketing is a separate, competing library. The vault's Media Mix Modeling (MMM) page correctly states only "Bayesian framework" without naming a backend.

Experimental calibration — unsolved vs addressed Recast (getrecast.com, 2024-03-11) listed incorporating experimental results as a significant limitation: "unclear how that should happen practically." The GeoX announcement (developers.google.com/meridian/geox, 2026-05-14) directly addresses this by converting geo-experiment results into calibrating priors. However, GeoX "will begin testing later in 2026" and is not yet GA — the limitation is architecturally addressed but not yet in widespread practice.

Conflict-of-interest risk — absent vs structural Some sources (Keen Decision Systems citing search results, 2026-02): "no evidence of bias has yet surfaced." AdExchanger (2026-07-14): Mutinex CEO argues structural conflict of interest is inherent and analogous to pre-Enron financial self-reporting. Both agree no proven bias exists; they disagree sharply on whether the structural risk is material.

Automatic Knot Selection — improvement vs degradation AKS (v1.2.1) improves in-sample fit metrics (R², MAPE) — straightforward better-model signal. PyMC Labs follow-up (December 2025): AKS increases residual autocorrelation (Durbin-Watson falls further from ideal ~2.0) and causes "leakage of media signal into the baseline," degrading causal attribution accuracy. Net effect depends on whether the user's priority is predictive fit or causal inference.

Key terms

Term	Meaning
GQV	Google Query Volume — indexed search demand signal used as paid-search confounder
GeoX	Meridian's geo-experiment calibration extension
AKS	Automatic Knot Selection — algorithm for automated spline baseline fitting
mROI	Marginal ROI — incremental outcome from the next dollar of spend (Meridian's label for miROAS)
MMM Data Platform	Google's tool for requesting GQV and YouTube R&F data for Meridian
Studio	Enterprise platform for managing high-volume Meridian models on Google Cloud
ESS/s	Effective Sample Size per second — MCMC sampler efficiency metric
TFP	TensorFlow Probability — the backend library Meridian is built on

Next frontier concepts (dangling links from this run)

PyMC-Marketing — PyMC Labs' competing open-source MMM library; 2–20× faster than Meridian per benchmark; no page yet
Meta Robyn — Meta's open-source MMM; reportedly being wound down per AdExchanger (2026-07-14); no page yet
Qualified Future Conversions (QFC) — Gemini-powered Google Ads metric predicting future sales; circularity risk when used as Meridian prior; no page yet
Adstock — carryover/lag transformation; referenced across MMM cluster; no dedicated page
Saturation — Hill function diminishing-returns transformation; no dedicated page