Last-click attribution has been lying to your CFO for years. When a prospect searches your brand name, clicks a paid ad, and converts, that keyword claims 100% of the credit. The CFO sees a 4:1 return and approves the budget. But did the ad cause the sale, or did the prospect's preexisting intent cause both the search and the purchase? Google's Meridian, the open-source marketing mix modeling framework made available to all marketers in January 2025, finally gives PPC teams a defensible way to answer that question.
The timing matters. Privacy regulations keep tightening, third-party cookies are disappearing, and customer journeys now span dozens of touchpoints before anyone clicks "buy." Relying solely on user-level tracking is no longer a viable long-term strategy. That reality has driven a resurgence in marketing mix modeling, and Meridian has emerged as the framework most relevant to performance marketers.
The Search Causality Problem
Traditional MMMs were built for TV and print. They used national-level data aggregated by week or month, which failed to capture the rapid fluctuations of digital performance media. Worse, they couldn't distinguish between correlation and causation in search. When query volume spikes and paid search conversions spike with it, a legacy model might credit the ads for lift that organic demand already guaranteed.
Meridian tackles this head-on with two features that matter most to PPC teams. First, it runs calculations at the geo level rather than the national level. By modeling variations in spend and performance across regions, it can isolate incremental impact with far more statistical power. Second, it incorporates Google Query Volume as a control variable. By separating organic search demand from paid search performance, the model can estimate what your ads actually caused rather than what they happened to coincide with.
Google's own documentation recommends treating brand-specific keyword campaigns and generic keyword campaigns as separate channels in the model. Brand campaigns are often modeled using clicks because they're intended to drive direct traffic. Generic campaigns are often modeled using impressions because impressions can influence even when they don't result in a click. This granularity matters when you're defending budget allocation to a CFO who wants to know why you're bidding on your own brand name.
Open Source Means Auditable
The open-source aspect deserves more attention than it usually gets. Brands have been right to worry about Google grading its own homework. With Meridian, you can inspect and modify the underlying code. You can see exactly how the model handles adstock decay, saturation curves, and prior assumptions. If your data science team disagrees with a default, they can change it.
Meridian is fully Bayesian, which means you encode what you already believe about a channel as a prior, and the model updates that belief against the data. Teams with real commercial knowledge get to use it. Teams without it spend the review meeting defending numbers somebody typed in because the notebook wanted a value. The Bayesian approach also propagates uncertainty through every parameter estimate, so you get credible intervals rather than false precision.
From Model Output to Budget Decision
The gap between having MMM insights and acting on them has always been the problem. Google launched Scenario Planner in February 2026 specifically to close it. The no-code interface lets marketers test different budget allocations and see projected ROI without writing Python or waiting for a data science queue.
According to a Harvard Business Review Analytic Services report, nearly 40% of organizations struggle to connect MMM outputs to real-world business decisions. Scenario Planner addresses this by turning complex model outputs into interactive simulations. You can shift spend across channels, set constraints, and compare conservative versus aggressive scenarios side by side.
Google's analysis of Meridian projects in APAC found that advertisers achieved a 14% increase in media-driven revenue or conversions by optimizing budget allocations based on Meridian results. That's the kind of number that survives a board meeting.
What You Need Before You Start
Meridian requires at least two years of weekly data across several channels, plus experiments to calibrate the model. The data requirements are not trivial: spend, impressions or clicks, conversions, and control variables like pricing, promotions, and seasonality. If you're already pushing Google Ads data to BigQuery, the integration is straightforward. If you're not, the data engineering work comes first.
Running a model has become the easy part. Getting a CFO to move spend on the strength of one is the same conversation it always was. The model gives you the math. You still need to explain the assumptions, show the sensitivity analysis, and propose a pilot that limits downside risk.
A Two-Week Pilot Plan
Start with a single question: what is the incremental contribution of brand search? Pull two years of weekly data by geo. Include Google Query Volume as a control. Run the model with default priors, then run it again with priors informed by any incrementality tests you've already conducted. Compare the credible intervals. If they're wide, you need more data or better calibration. If they're tight and the ROI estimate differs materially from your last-click reports, you have a conversation worth having with Finance.
The goal is not to replace your attribution system overnight. The goal is to build a parallel measurement capability that can answer the questions attribution cannot. When the CFO asks why you're spending on brand keywords when organic would capture the same traffic, you want a model that can estimate the counterfactual, not a platform report that assumes every click was incremental.
Meridian won't make the politics of budget allocation disappear. But it gives you a framework that Finance can audit, assumptions they can challenge, and outputs they can stress-test. That's the difference between a marketing expense and a revenue-predictable engine.
