Last-click attribution is dead. The teams defending media budgets at the C-suite level in 2026 are running on Marketing Mix Models. This hub is the complete operating reference — the architecture argument, the 12-point readiness checklist, the FAQ, and the proof-layer thesis that makes MMM more than a vendor deliverable.
Last-click attribution gave a generation of marketers a number to defend in a budget meeting. It also gave them a number that systematically under-credits upper-funnel investment, over-credits whichever channel happens to be closest to the conversion event, and breaks the moment a privacy regulation, browser change, or consent-mode adjustment shifts the underlying signal. Marketing Mix Modeling is the discipline that handles attribution differently — at the channel level, from aggregated data, with statistical confidence intervals you can show your CFO.
Meridian is Google's open-source Bayesian MMM. The reason it matters isn't that MMM is new — large enterprises have been running mix models in some form since the 1960s — it's that an auditable, open-source, modern MMM removes the two failure modes that quietly killed the previous generation: vendor lock-in on a black-box methodology, and a refresh cadence too slow to make in-flight decisions on.
Three reasons. First: Meridian is open source, which means the modeling assumptions are auditable. When a skeptical CFO asks "why does the model say YouTube is incremental," the answer can be "here is the prior distribution, here is the saturation curve, here are the parameters we chose, here is the sensitivity analysis." A vendor MMM cannot offer that conversation.
Second: Meridian is Bayesian. Outputs come with credible intervals, not point estimates. A media plan defended with "YouTube contributed somewhere between $4.2M and $5.8M of incremental revenue at 90% credibility" survives finance scrutiny in a way "YouTube contributed $4.9M" never will.
Third: Meridian was designed against the actual constraints of 2026 marketing data — privacy-preserving aggregated inputs, consent-mode reality, retail-media data formats, geo-experiment lift estimates as priors. Most legacy vendor MMMs are still adapting decades-old methodologies to a world where the underlying data stopped looking like 2015.
The architecture
The five inputs that decide whether your MMM compounds.
Every MMM that produces decisions a CFO actually uses is engineering five inputs underneath the model. Every project stuck at "interesting but not actionable" is missing at least two.
Input Layer 01
Channel-level spend, time-series
Weekly or daily spend by channel, at consistent granularity, going back at least 104 weeks. This is the spine of the model. Patchy historical data means patchy posterior distributions, and the model will quietly tell you exactly that — in the form of uselessly wide credible intervals.
Input Layer 02
Clean business-outcome series
Revenue, conversions, member-years, funded accounts — whichever business metric the model is being asked to attribute to. Backfilled, deduplicated, and at the same cadence as the spend data. The model is only as honest as the outcome series.
Input Layer 03
External variables
Seasonality, pricing changes, promotional periods, competitor activity, macroeconomic shocks. Without these, the model attributes everything they actually caused to whichever marketing channel happened to be on at the time.
Input Layer 04
Prior information from experiments
Geo-holdout incrementality studies, conversion-lift tests, and other causal experiments — used as Bayesian priors. This is where Meridian quietly outperforms older MMMs: the experimental evidence gets folded into the model, rather than sitting on a separate slide nobody references.
Input Layer 05
A governance cadence
A clear refresh schedule, an owner for each input series, and a defined point in the planning cycle when the model output drives an allocation decision. The model is a tool; the cadence is the capability. Most failed MMM projects failed at the cadence layer, not the modeling layer.
The diagnostic
The 12-point Meridian readiness checklist.
Run this before commissioning a Meridian implementation — or before letting a vendor sell you one. Score each yes / partial / no. Anything below 9/12 means the modeling work will land as a dashboard nobody trusts; fix the readiness layer first.
The 12-point Meridian readiness checklist
Designed to be run in a 2-hour cross-functional session with analytics, finance, and the channel leads.
01. Channel-level spend history. 104+ weeks of clean spend, by channel, at consistent granularity. Gaps and re-categorizations identified and documented.
02. Outcome series at the right cadence. Revenue, conversions, or the business metric you actually want to attribute to, deduplicated and at the same time grain as the spend data.
03. External variables loaded. Seasonality, pricing, promotions, competitor activity, macroeconomic indicators — at least the ones with non-trivial effects on the outcome series.
04. Geo-experiment evidence available. At least one well-designed geo-holdout incrementality test per major channel, formatted to be used as a Bayesian prior.
05. A data engineering owner. One named person responsible for the input pipelines staying clean, not a quarterly contractor relationship.
06. A modeling owner. One named person — internal or vendor — who owns the model structure, the prior specifications, and the sensitivity analysis. Not "the team."
07. Privacy and consent posture documented. How the underlying data was collected, what consent it was collected under, and what the legal and compliance teams have signed off on.
08. A clear modeling question. "What's the ROI of YouTube" is not a modeling question. "What is the credible range of incremental contribution from YouTube to subscription revenue at the current spend level, holding seasonality and pricing constant" is.
09. A refresh cadence on the calendar. When the model gets re-run, by whom, and what decision it feeds into. Without this, the model is decoration.
10. An audit trail for prior choices. Documented rationale for every Bayesian prior. Future-you will thank present-you.
11. A defensible interpretation protocol. How outputs get translated into recommendations, and what guardrails sit around the model-versus-human judgment boundary.
12. Integration with the planning cycle. The model outputs are formally referenced in the annual or quarterly planning meeting, not just in an analytics review nobody attends.
From The Lab
A working Meridian experiment.
Experiment #003 in the Uncommon Move Lab is taking a published MMM and re-running it with vision-model-derived creative features as additional covariates. The hypothesis: a meaningful share of what current MMMs attribute to spend levels is actually attributable to which creative ran when.
The 12-point readiness checklist is built directly into the 2026 Signal Architecture Framework — the same 14-page PDF used by analytics teams at enterprise organizations preparing an MMM commission. Get it free.
The Meridian Readiness Checklist
Embedded inside the full Signal Architecture Framework, alongside the Signal Flow Map and the Cost Per Decision Worksheet.
Five questions about Meridian MMM, answered straight.
Is Meridian a replacement for GA4 or Google Ads attribution?
No — they answer different questions. Platform attribution (GA4, Google Ads) measures what happened inside the tracked customer journey. Meridian MMM measures the incremental impact of channel investment on a business outcome, using aggregated data and statistical priors. The two coexist; the smart move is to use platform attribution for in-flight tactical decisions and MMM for budget allocation decisions.
How much data does Meridian need to produce useful outputs?
Practical floor is roughly 104 weeks of channel-level spend and outcome data, plus the external variables that materially affect the outcome. Less than that and you'll get usable point estimates but uselessly wide credible intervals. The data quality matters far more than the data volume past that threshold.
Can a marketing team run Meridian without a data science team?
Smaller organizations have done it, but it's an uphill climb. The model itself is open source and well-documented; the operational work — data pipelines, prior specification, sensitivity analysis, governance — is what determines whether the project lands. Most teams that succeed without an internal DS function pair with a specialized partner rather than running the full project alone.
How does Meridian handle privacy and consent-mode realities?
It was designed for them. Meridian operates on aggregated, channel-level inputs — not user-level data — so the model is structurally compatible with consent denial, third-party cookie loss, and most enterprise privacy postures. The privacy constraint that broke last-click attribution is the constraint Meridian was built around.
What's the relationship between Meridian and Performance Max measurement?
Meridian provides the cross-channel proof layer that PMax's placement-level reporting cannot. The pattern most enterprises end up adopting: PMax for in-channel execution, geo-holdout experiments for short-cycle incrementality reads, Meridian for long-cycle cross-channel attribution. The Performance Max hub walks through how those three layers fit together.
Connected ideas
Where Meridian fits in the bigger architecture.
Meridian is the proof layer in a larger system. The pieces below are the ones that connect to this hub most directly — the essays underneath, the framework around it, and the channels it most often measures.