Google's Ads DevCast published its sixth episode on May 28, 2026, dedicating the full running time to Meridian - the company's open-source Marketing Mix Modeling framework - and to a suite of three new product surfaces announced the week before at Google Marketing Live 2026. The episode features Katie Munro, product manager for Meridian, and Jeff Li, a developer relations engineer who has worked on the product for roughly a year. Host Cory Liseno poses what turns out to be the episode's central question early on: why should a CMO or CFO trust an advertising network to tell them how to allocate a marketing budget?

The full episode is available on YouTube:

What Meridian is - and what problem it solves

At its core, Meridian is a Marketing Mix Modeling (MMM) framework hosted on GitHub and built using Python. It ingests aggregated historical media data across every channel - digital, linear TV, billboards, print - and combines that with non-advertising factors such as pricing changes or seasonality. The output is a unified return-on-investment view across channels, designed to help marketers understand the causal impact of their spending rather than the correlational picture that attribution tools tend to produce.

The distinction matters. According to the episode, attribution often credits an ad even if a customer would have made the purchase regardless. Meridian separates "incremental sales" - conversions driven because an ad was shown - from "baseline sales" - conversions that would have occurred from brand loyalty, word of mouth, or time of year. The framework became globally available in January 2025 after testing with hundreds of brands, and it operates entirely on aggregated data, meaning it does not rely on cookies or individual user-level tracking.

That privacy-safe architecture is not incidental. The shift away from third-party cookies has made deterministic attribution harder to sustain across full customer journeys. A 2024 analysis on PPC Land noted that MMM was experiencing a resurgence precisely because it offers a way to assess marketing effectiveness in a privacy-first environment. Meridian was built to make that methodology more accessible and more rigorous at the same time.

The core math: Hill curves and adstock

The technical foundation of Meridian rests on two mathematical constructs that Jeff Li describes in the episode as "universally accepted in the world of advertising."

Hill curves model diminishing returns. The underlying assumption is that the effectiveness of an ad declines as spending increases. A channel can be saturated - at which point additional spend generates smaller and smaller returns per dollar. Meridian uses Hill curves to calculate where headroom exists (channels where more spending would still yield strong returns) and where saturation has already set in (channels where incremental investment will underperform).

Adstock models the lagged and decaying effect of advertising over time. A car commercial seen on a Tuesday may not result in a purchase until several weeks later. Adstock functions capture that delay mathematically, accounting for the lingering influence of an ad long after initial exposure. The combination of Hill curves and adstock is what allows Meridian to tell a marketer not just how much to spend, but when the effects of a campaign are likely to materialise.

According to the episode, these two constructs power the budget headroom analysis at the heart of the tool. Jeff Li gives a concrete illustration: "The next $100,000 you spend on a display campaign is only going to give you half the return of the previous $100,000." That kind of signal - machine-calculated, drawing on historical spend curves - is what the tool is designed to surface. Katie Munro adds that Meridian might reveal "that your brand spend is driving 15% lift in the effectiveness of your direct response campaigns down the line," making cuts to brand budgets look efficient in the short term while quietly degrading direct response performance months later.

Earlier updates in September 2025 extended these mathematical foundations further, introducing binomial adstock decay functions capable of measuring longer-term upper-funnel effects, and adding channel-level contribution priors that allow users to incorporate domain knowledge directly into the model.

The trust problem - and how open source addresses it

The episode addresses a tension that is difficult to ignore. Google operates one of the world's largest advertising platforms. An MMM tool that recommends how to allocate budget across channels - including Google's own channels - is therefore in the position of, as Cory Liseno puts it, "grading our own homework."

According to Katie Munro in the episode, the decision to release Meridian as open source was deliberate precisely because of that perception: "We aren't asking for blind trust. That's why we made Meridian open source. We know that anything coming from Google will have that extra perception, and we wanted to give users the opportunity to look under the hood and validate for themselves that they believe in the methodology as we do, and that there's nothing nefarious going on here."

Jeff Li reinforces the point: "Unlike proprietary models where you just put your data in and cross your fingers and hope that the output is unbiased, Meridian lives on GitHub. Your data science team can audit every single line of the code."

The Bayesian statistical approach used in Meridian adds another layer of auditability. According to Katie Munro, the Bayesian framework means teams can bring their own domain knowledge into the model as "priors" - setting boundaries on what they believe about their channels based on real-world evidence from incrementality experiments - rather than relying solely on correlations or a single platform's attribution numbers. The model then runs entirely within the team's own environment.

PPC Land's April 2026 examination of Meta's Robyn MMM explored the same structural tension - the conflict of interest that exists when a platform builds the tool that guides budget allocation. The Meridian episode covers the same ground from Google's perspective, with open sourcing presented as the mechanism for resolving it.

The measurement trifecta

Neither Munro nor Li argues that Meridian replaces other measurement tools. The episode is explicit that Meridian occupies one specific role within what the participants call a "modern measurement trifecta." Attribution - via tools like Google Analytics 4 - handles day-to-day tactical campaign optimisation. Incrementality experiments test specific hypotheses against real-world controls. Meridian itself operates at the macro level, informing long-term channel budgeting decisions over periods of months rather than days.

According to Katie Munro, the three layers work in sequence: "You'll still use attribution for your day-to-day campaign tweaks, and you'll still use real-world incrementality experiments to test hypotheses, validating the results of your model and feeding them back in to calibrate it and make it smarter over time." The implication is that Meridian without a functioning attribution layer and without incrementality testing is a weaker tool - the calibration loop requires inputs from all three sources.

New products from GML 2026

The pre-GML announcement on May 5, 2026 introduced three new Meridian-related products, all discussed in depth in the episode. The May 28 DevCast provides the clearest technical description of each to date. The episode also follows Google's May 20 announcement that Meridian has been integrated directly into Google Analytics 360.

Meridian in Google Analytics 360

The GA360 integration is positioned as the most accessible entry point for teams that lack dedicated data science resources. According to Katie Munro, "We're putting Meridian's core methodology into GA 360 to create a turnkey based budgeting solution designed specifically for marketers." The work of connecting data sources and building the model happens on Google's infrastructure; the user-facing interface exposes budgeting tools that surface cross-channel ROI insights without requiring any model configuration.

The integration draws on the behavioural data that GA360 customers have already collected about customer journeys across their sites and apps, reducing the data preparation burden that has historically made MMM implementations time-consuming. Jeff Li describes the practical outcome: getting "cross-channel ROI insights and budget optimisation recommendations with greater speed and basically less heavy lifting" - removing the need for a months-long modelling project.

For teams that are not data scientists but need defensible channel-level budget recommendations, this path eliminates the implementation barrier entirely.

Meridian Studio

Meridian Studio is described in the episode as "Google's new enterprise platform built on Google Cloud that empowers teams to build and manage Meridian models at scale now with greater speed and easier flexibility." It is an additive offering to the open-source library, which according to Katie Munro Google "remains fully committed to and will continue to invest in heavily in terms of methodology, innovations, and improvements."

BigQuery underpins the data storage layer, making it easier to configure, collaborate on, and iterate models at enterprise scale. Looker Studio (referred to as "Data Studio" in the episode) handles the visualisation of model outputs, allowing marketing teams to run scenario analyses and budget optimisations directly against the Meridian results.

Participation in 2026 will be limited during an initial pilot period. According to the episode, teams can stay informed about Studio availability by subscribing to the Meridian website newsletter.

Agentic capabilities inside Meridian Studio

One of the more technically specific sections of the episode covers the native agentic AI functionality being built into Meridian Studio. Jeff Li describes a three-stage workflow where agents assist at each phase of the modelling lifecycle.

During the data preparation phase - which, according to Jeff Li, "often consumes the majority of a project's lifecycle" - an integrated agent monitors data quality checks. "If a data slice fails a quality check, the integrated agent will actually summarise the issue directly for you instantly and provide suggested fixes," he explains. Users can then address the issue manually, ask follow-up questions through a conversational interface, or let the agent resolve the problem automatically.

Once a model converges, the agent shifts to validation support, generating a comprehensive output report with a model health summary and proactively suggesting improvements. The final phase applies the same agentic layer to strategy: interpreting complex results, running optimisations, and generating budget plans that connect directly to the Scenario Planner for presentation to senior stakeholders.

According to Katie Munro, the design intent is clear: "The user is still maintaining all control in that case and authorising the agent to act as its data science assistant on behalf of them, so that they can spend more time on the high-level strategy and decision making." The agents act, but humans approve.

The Scenario Planner itself launched in February 2026 as a no-code interface sitting on top of the Meridian model, designed to make budget planning outputs accessible to people without coding skills. Meridian Studio's agentic layer extends that accessibility further up the modelling stack.

Meridian GeoX

Meridian GeoX is introduced in the episode as a new open-source geo-based incrementality solution. Its function is to serve as the calibration engine for the broader Meridian MMM - running geographic experiments to establish causal evidence about which channels are genuinely driving incremental growth, then feeding those results back into the main model as priors.

The methodology compares test markets against statistically matched control regions. By holding some regions back from media exposure, GeoX isolates the incremental lift of a media spend rather than simply observing correlation. Both single-cell and multi-cell experimental designs are supported, and the open-source code base means teams can audit the experimental methodology in the same way they can audit the core model.

According to Katie Munro, "Closed beta testing has officially kicked off and the product is in the hands of customers," with general availability targeted for later in 2026. The episode also describes a feedback loop running in both directions: Meridian will proactively suggest which channels would benefit most from a GeoX experiment, making the decision about where to run experiments data-driven rather than intuitive.

The same transparency principle applies here. Jeff Li notes that GeoX "provides a rigorous, defensible proof of performance across any channel that you can confidently present to your CMO or CFO," citing the move from correlation to causation as a significant step for measurement credibility.

Why this matters for marketers

The Ads DevCast episode matters for the marketing community for reasons that go beyond the product specifics. For the past several years, the measurement industry has been navigating a structural shift: the gradual erosion of deterministic user-level tracking has left attribution models covering less of the customer journey than they once did. MMM has gained traction as an aggregate, privacy-safe alternative, but adoption has been limited by the technical barrier of building and maintaining models.

The three new Meridian surfaces address different parts of that barrier. GA360 integration targets teams with no modelling capacity. Meridian Studio addresses enterprise teams that have the expertise but need better tooling for scale and collaboration. GeoX addresses a methodological gap - the need for causal evidence that can ground the model in real-world experiment results rather than historical correlations alone.

Google's original Meridian announcement in March 2024 positioned the framework as a response to the deprecation of third-party cookies and the complexity of modern cross-channel measurement. Two years later, the product family has expanded substantially, and the DevCast episode is the most detailed public explanation of how each layer connects to the others.

Timeline

Summary

Who: Google's Advertising and Measurement Developer Relations team, specifically product manager Katie Munro and developer relations engineer Jeff Li, speaking on the Ads DevCast podcast hosted by Cory Liseno.

What: Episode 6 of Ads DevCast, published May 28, 2026, provides a technical walkthrough of Meridian - Google's open-source Marketing Mix Modeling framework - covering its core mathematical components (Hill curves and adstock), the rationale for open sourcing the codebase, and three new products announced at Google Marketing Live 2026: Meridian in Google Analytics 360, Meridian Studio with agentic AI capabilities, and Meridian GeoX, an open-source geographic incrementality solution entering closed beta.

When: The episode premiered on May 28, 2026, following Google Marketing Live 2026 on May 20 and the pre-GML measurement announcement on May 5, 2026. Meridian GeoX is targeting general availability later in 2026; Meridian Studio participation in 2026 will be limited during an initial pilot period.

Where: The episode is published on the Google Advertising and Measurement Developers YouTube channel and as a podcast. Meridian's open-source code is hosted on GitHub, Meridian Studio runs on Google Cloud with BigQuery for data storage, and the GA360 integration sits within the existing Google Analytics 360 interface.

Why: The accelerating shift away from third-party cookies has eroded the reliability of deterministic attribution, pushing marketers toward aggregate, privacy-safe measurement methods. Meridian addresses that gap with a Bayesian causal inference framework that operates on aggregated data. The new product surfaces are designed to bring that methodology to teams with varying levels of technical capability - from enterprise data science teams running hundreds of models on Google Cloud, to marketing directors who need budget guidance without building models themselves.