Meridian MMM 2026: Paid Media Measurement Guide

Your Google Ads account reports 400 conversions. GA4 says 310. Your CRM closed 260 deals. You raise budget by 30%, the platform reports 520 conversions — and the CRM still closes 265 deals. That gap is not a tracking bug you can fix with another tag. It is the ceiling of what click attribution can tell you: it splits conversions that already happened between channels, but it never answers the only question that matters for budget decisions — how many of those would not have happened without the ad.

Two things answer that question: experiments and MMM and Meridian — marketing mix modeling, and the open-source Google framework that runs it. In 2026 Meridian went from “a library for data scientists” to a product baked into Analytics 360, and the cost of entry to MMM dropped from “hire a team and wait two quarters” to “one analyst, Python, a couple of weeks.” Here is how it actually works, what data you need, where it lies to you, and when you should not run it at all.

Why click attribution broke

Attribution answers “who gets credit.” Incrementality answers “would this have happened anyway.” Those are different questions, and in 2026 the distance between the answers has widened for concrete reasons:

Consent. In the EU a large share of sessions arrives with no analytics consent at all. Some conversions get modeled back, some are gone — the mechanics are covered in our breakdown of Consent Mode v2 and privacy signal loss.
Cross-device journeys. Saw the video on a TV, searched on a phone, bought on a work laptop. To the platform that is three separate people.
Delayed effect. Google’s own 2026 figure: only about 40% of Demand Gen conversions land within the first 30 days. A standard conversion window simply cannot see the rest.
Brand spillover. Video drives no clicks but drives branded searches — and brand search takes the credit.
Offline. Retail, calls, dealers, marketplaces: there is no pixel to fire.

The more channels you run and the longer your sales cycle, the further platform-reported performance drifts from revenue. Nobody is lying; the measurement method physically cannot see what it cannot see.

What MMM is, and what makes Meridian different

Marketing mix modeling takes aggregate series — spend, impressions, reach by channel, by week, by region — plus control variables (seasonality, pricing, promotions, competitor activity, stock availability) and estimates how much each channel contributed to sales. No user-level identifiers anywhere. That is the whole point: MMM does not degrade when cookies disappear, because it never used them.

Classic MMM was expensive and opaque: an agency black box, a report once a quarter, no way to audit the assumptions. Meridian changes three things:

It is open source. A Python framework (versions 3.11–3.13 supported) you can run in-house, inspect assumption by assumption, and rerun next quarter without renegotiating a contract.
It is Bayesian. You do not get a single number, you get a distribution: not “channel ROI is 3.2” but “channel ROI sits between 2.1 and 4.4 with 90% probability.” For a budget decision, that honesty matters more than a confident point estimate.
It is geo-first. Meridian is built for regional data rather than a single national series. Every region is an observation, so the same calendar period yields far more signal, and effects that drown in national noise become measurable.

One practical note: the model uses MCMC sampling (NUTS), which is compute-heavy. The reference setup is a GPU machine in the class of a T4 with 16 GB of RAM. It will run on a laptop — it will just take hours instead of minutes.

What changed in 2026: MMM and Meridian inside Analytics 360

At Google Marketing Live 2026 Meridian was integrated directly into Analytics 360. What that buys you:

Capability	What it means day to day
Native GA4 signal	Conversion and Google-channel spend series are pulled from your account instead of being rebuilt by hand
External channels in one place	TikTok, Pinterest, Snap and other spend feed the same model, so channels are compared on one scale
Scenario planning	Answers “what happens to revenue if I move 20% of budget from channel A to channel B”
Tier	Paid Analytics 360 only — the integration is not available on free GA4

No Analytics 360 budget? The open-source path is still free; you just assemble the data yourself. The usual route is a Google Ads to BigQuery data transfer, which already gives you daily spend, impressions and conversions by campaign and region — you aggregate to weeks and bolt on the non-Google sources.

Three measurement methods — stack them, don’t pick one

The most common mistake is treating MMM as a replacement for attribution. It is not. The three methods answer different questions, and a mature measurement stack runs all three.

Method	Answers	Speed	Blind spot
Attribution (Google Ads, GA4)	Which campaign, keyword or creative worked	Hours	Cannot separate incremental from organic; blind to offline and view-through
Geo experiments and holdouts	How many conversions a channel actually added	2–6 weeks	Expensive in traffic; you can only test a few things at a time
MMM (Meridian)	How to split budget across channels and how much each can absorb	Weeks to build, then refreshes	Blind below channel level; needs a long history

In practice: MMM sets the strategic frame (how much into search, video, social), experiments calibrate the model and sanity-check its claims, attribution runs tactics inside a channel — bids, keywords, audiences. If you have never run a holdout, start there; it is cheaper and faster than building a model. The math is laid out in our guide to incrementality and geo experiments.

The data you actually need

This is the underestimated part. Roughly 80% of a first MMM build is data plumbing, not modeling. The minimum panel:

Input	Granularity	Depth (guideline)
Business outcome (revenue, orders, qualified leads)	Week × region	2 years minimum, 3 is better
Media spend per channel	Week × region	Same period, no gaps
Impressions / reach / frequency	Week × region	Mandatory for video and reach channels
Controls: price, promo, seasonality, holidays, stock	Week	Same period
Past experiment results	Per channel	Used to set priors

Treat the depth numbers as guidelines, not law. Long sales cycles and heavy seasonality demand more history; a wide geo split lets you get away with less, because each region adds observations. The real requirement is variation: if you spent exactly the same amount every week in every region for a year, no amount of math will tell you what happens at a different budget.

Building your first model: seven steps

Pick the dependent variable. Not “conversions from the ad account” — a business outcome: revenue or confirmed orders from CRM/ERP. If you model leads, model qualified leads, or the model will optimize your budget toward junk.
Assemble the panel. One row = region × week. Channels in columns, controls beside them. Fill zeros only where spend genuinely was zero.
Reconcile the geo dimension. Regions in the CRM and in ad platforms must map to the same list. This is usually the most tedious step: one system uses states, another DMAs, another cities.
Set priors. This is the Bayesian advantage: a geo test result for a channel goes into the model as prior knowledge, so the model starts from a measured fact instead of a guess.
Fit and check convergence. Watch MCMC diagnostics (R-hat) and holdout error. A model that fits the past perfectly but misses the held-out period is worthless.
Read the saturation curves. The real output of MMM is not “channel ROI” but curve shape — the point where a channel stops returning. That is what tells you where the next 10% of budget belongs.
Validate with an experiment. Model says video is undervalued by 40%? Run a geo test on video and see if it holds. If it doesn’t, fix the model, not reality.

An MMM that has never been calibrated against an experiment is a well-dressed opinion, not a measurement. A Bayesian model returns a confident number even where the data was thin — only a real test separates a sound inference from a guess.

Five mistakes that kill a first MMM

Too many channels, too short a series. Fifteen channels over 80 weeks and the model starts shuffling credit between correlated columns. Group small channels together.
Ignoring price and promo. If you ran a 20% discount in the same weeks you scaled traffic, the model hands the entire lift to media.
Flat budgets everywhere. No variation, no inference. Deliberately vary regional budgets — that is not wasted money, it is an investment in measurability.
Using MMM for tactics. It will never tell you which keyword to pause or which creative wins. That is what account reports and Experiment Center tests are for.
Running it once. A model built once is stale within a quarter. Put refreshes on the calendar and watch how saturation curves move.

New 2026 metrics: QFC and attributed brand searches

Alongside Meridian, Google shipped in-account metrics that attack part of the same problem:

Qualified Future Conversions (QFC) — a Gemini-powered predictive metric estimating conversions that will occur within 180 days of an ad interaction, reported as three columns: qualified future conversions, cost per qualified future conversion, and qualified future conversion rate. It launched as a restricted global pilot with wider beta expected later in 2026. The rationale is blunt: for Demand Gen and video, most of the effect sits outside the standard window, which makes those campaigns look unprofitable when they are not.
Attributed Brand Searches — an always-on metric for branded searches influenced by campaigns, primarily YouTube. Previously this required a formal brand lift study; now the signal runs continuously.
Campaign-type attribution — reporting that isolates Demand Gen’s contribution from Performance Max and Search instead of blending them through de-duplication. That is directly useful when you are untangling cannibalization between PMax and Search and cannot tell who is taking credit from whom.

One caveat worth repeating: QFC is a prediction, not a fact. Use it as an additional signal to be reconciled with closed revenue — not as the sole justification for multiplying a budget.

What the output actually looks like

To make this concrete, here is a stylized quarterly summary across five channels: spend, modeled average ROI with its credible interval, and marginal ROI — the return on the next dollar.

Channel	Quarterly spend	Average ROI (interval)	Marginal ROI	Action
Brand search	$60,000	8.4 (6.9–10.1)	1.1	Do not scale — saturated; the high average only reflects demand that already exists
Non-brand search	$210,000	3.1 (2.6–3.7)	2.4	Scale moderately while marginal ROI stays above your payback threshold
Performance Max	$180,000	2.8 (1.9–3.9)	2.1	Wide interval — test before deciding, don’t scale on the point estimate
YouTube	$70,000	1.9 (0.8–3.4)	1.8	Growth candidate: far from saturation, but the interval demands a test first
Native	$40,000	0.7 (0.2–1.4)	0.6	Cut or change placements

Three readings a good planner takes from a table like this — and three they refuse to take:

Decide on marginal, not average ROI. Brand search at 8.4 looks like the best channel you own, but the next dollar into it returns 1.1, because you already capture nearly all brand demand. Budget decisions live on the margin.
Read interval width as confidence. YouTube at 0.8–3.4 is the model honestly saying “we don’t know precisely.” That channel gets a test, not a 3x budget.
Compare channels on one scale. Putting every channel on a single metric is precisely the value MMM adds over comparing platform reports, each with its own attribution logic.
Do not push a channel-level conclusion down to a specific campaign — the model cannot see campaigns.
Do not act on a point estimate with the interval stripped off; that hides the uncertainty that should govern the decision.
Do not assume returns hold when you move budget: saturation curves are non-linear, and doubling spend almost never doubles results.

The organizational half: who owns the model

Technically an MMM comes together in a few weeks. It usually fails on process, not math — built once, presented in a meeting, forgotten. Settle four questions before you start:

Who owns it. One accountable person for data, refreshes and interpretation. No owner, or three owners, and the model dies after one quarter.
How conclusions reach the plan. Agree upfront that quarterly budget planning starts from saturation curves rather than last year’s channel splits. Otherwise it stays a nice deck.
What happens when it contradicts the ad account. A disagreement is a hypothesis, not an argument — decide in advance who runs the experiment that settles it.
What experiment cadence you maintain. One or two geo tests per quarter on the largest channels is a sensible baseline: it calibrates the model and insures you against systematic error.

Also agree on time horizons. MMM speaks in quarters; a buying team lives in weekly reports. If those horizons are not explicitly separated inside the company, every model conclusion gets overruled by weekly noise and yesterday’s CPA.

When you should not run MMM

The section most MMM articles skip. Do not build a model if:

Your monthly spend is small. Below a few tens of thousands of dollars a month, channel effects sit inside the noise and the effort will not pay back. Run geo holdouts and fix your tracking instead.
You run one channel. A media mix model without a mix is an expensive way to describe seasonality. An on/off experiment is enough.
You have no history. A three-month-old account has no series to model. Build the history and meanwhile fix the basics — starting with duplicate conversions and overcounting.
Your CRM data is dirty. A model trained on a wrong outcome variable gives you a confidently wrong answer. Data first, model second.

In all of those cases, more money comes from cleaning up the account itself — structure, bidding, and a hard number for your maximum CPA derived from LTV. If you want an outside read on that, our Google Ads agency team does exactly this kind of account review.

Implementation checklist

Lock the business metric you will model, and verify it is clean.
Build a region × week panel covering at least two years.
Map CRM and ad-platform regions to a single reference list.
Run at least one geo experiment before modeling — it becomes your prior.
Stand up Meridian (open source or via Analytics 360) and fit a baseline model.
Check convergence and holdout error before believing anything.
Read saturation curves and draft a reallocation scenario.
Validate the model’s headline claim with a dedicated experiment.
Schedule a quarterly refresh.
Never change in-channel tactics based on MMM output — that lives in the ad account.

FAQ: MMM and Meridian

Is Meridian free?

The framework is — it is open-source Python. You pay for compute (a GPU machine) and analyst time. The Meridian integration inside Analytics 360 requires the paid Analytics 360 tier; it is not available on free GA4.

How much data do I need?

Two years of weekly data as a guideline, three is better. But variation matters more than length: a long flat-spend series will not produce useful answers. With a geo split you need less calendar history because each region adds observations.

Can I model national data without regions?

Yes, Meridian supports it, but precision drops. Geo-level data is the model’s main source of statistical power, so use regions whenever you have them.

Does MMM replace GA4 and Google Ads reporting?

No. MMM works at channel and week level; it cannot see keywords, ads or audiences. Tactics still live in the ad account and GA4.

How accurate are the results?

A Bayesian model returns an interval, not a point. Read it as “channel ROI is very likely between X and Y.” A very wide interval means the data was insufficient and you should not act on it yet.

How is MMM different from incrementality tests?

A test measures one channel in one period with almost no assumptions — it is close to a true experiment. MMM covers every channel at once but rests on modeling assumptions. That is why they belong together: the test calibrates the model.

What if MMM contradicts my Google Ads report?

Expected, not alarming — they answer different questions. The gap usually means part of the reported conversions were not incremental (brand traffic that would have converted anyway). Resolve the disagreement with an experiment, not by arguing with a report.

Do I need a data engineer?

For the panel build, almost always — at least at the start. Fitting the model is a handful of commands; reconciling spend, regions and revenue into one coherent series is engineering work.

What are Qualified Future Conversions in plain terms?

Google’s prediction of conversions that will happen within 180 days of an ad interaction. It exists because effects are delayed: by Google’s own number, only about 40% of Demand Gen conversions land in the first 30 days.

How often should the model be refreshed?

Quarterly is a sensible default. Refresh out of cycle after major changes: new markets, a different product mix, or a sharp shift in budget structure.

Where do I start with no budget for any of this?

One geo holdout on your largest channel plus a tracking cleanup. That delivers most of the value for a fraction of the effort. MMM earns its keep once you run many channels at real scale.

If your bottleneck is not measurement but keeping campaigns live at all, look at our Google Ads agency accounts and advertiser verification service — no measurement model helps an account that cannot spend.

MMM and Meridian in 2026: Measuring Paid Media When Click Attribution Lies