Incrementality and MMM With Google Meridian: A Guide

Why does last-click attribution mislead budget decisions?

Last-click attribution gives all the credit to the final touch before a conversion. That systematically overrates channels that catch people at the end (branded search, retargeting) and underrates the ones that create demand earlier (video, social, podcasts). It also cannot see sales the marketing did not cause, so it over-credits marketing in general.

The result is budget arguments that never end, because each channel owner has a report that says their channel works. Incrementality measurement asks a different question: what would have happened without this spend? That is the question a budget owner actually needs answered.

What is Meridian?

Meridian is a marketing mix modeling (MMM) framework that Google released as open source and made available to everyone in January 2025. It is licensed under Apache-2.0. A marketing mix model uses aggregated data over time (spend, impressions, clicks, and outcomes such as revenue or conversions) to estimate how much each channel contributed. Because it works on aggregates, it needs no cookies or user-level tracking.

Meridian uses Bayesian causal inference, which means it combines prior knowledge with the data. It can handle large-scale geo-level data as well as national data, models channels by reach and frequency rather than impressions alone, and includes a budget optimiser that searches for the allocation that maximises return for a given budget or hits a target ROI.

How do geo holdouts calibrate the model?

A model is only as trustworthy as what anchors it. The strongest anchor is an experiment. In a geo holdout you pause or reduce spend on a channel in a set of comparable regions and keep it running elsewhere, then compare outcomes. The difference is that channel's incremental effect in the test period.

Meridian is designed to take results like this as priors: its documentation describes it as integrating incrementality experiment results as priors regardless of the channel or the experiment. In practice you run a test on your biggest or most disputed channel, feed the result in, and the model's estimates for that channel move towards what the test showed instead of floating on correlations.

• Pick test and control regions that looked alike before the test.
• Run long enough to cover your sales cycle.
• Change one channel at a time so the result has one cause.

Why report ROI as a range?

A Bayesian model gives a distribution, not a single number, and that is a feature. Saying a channel returns somewhere between 1.4 and 2.1 per dollar, with most likely 1.7, is more honest than saying 1.7, and it tells the budget owner how confident to be. Narrow ranges justify bold moves; wide ranges say run a test first.

This also changes the conversation in the room. When I built spend-to-revenue measurement with a Meridian incrementality read, budget decisions went from last-click arguments to a reallocation plan with ROI ranges, reviewed with executives.

What data do you need to prepare?

Most of the work in a marketing mix model is the data, not the model. Build one weekly table with a row per week (and per region, if you have geo data) and a column for each input. For media, include spend and, where you have them, impressions, clicks, or reach and frequency for video. For the outcome, use the metric the business runs on: revenue, conversions or qualified leads.

Add the things that move sales but are not media, known as controls. Price changes, promotions, product launches, seasonality and big external events all belong here. Without them, the model will credit whichever channel happened to spend more during a busy season.

Check the table before modelling. Look for weeks with missing spend, channels that changed names, and sudden jumps that came from a tracking change rather than real behaviour. A clean, boring dataset gives a model you can defend; a messy one gives confident numbers that are wrong.

• Weekly rows, per region if possible.
• Spend plus impressions, clicks or reach and frequency per channel.
• The outcome metric the business actually uses.
• Controls: price, promotions, launches, seasonality.

How do you present the results to executives?

Lead with the decision, not the model. Executives need three things: where the next dollar earns most, where money is being wasted, and how sure you are. Show ROI ranges per channel on one chart, mark the channels whose ranges are narrow enough to act on, and propose a specific reallocation with an expected effect.

Then propose the next experiment for the channel with the widest range. That keeps measurement honest and turns the model into a cycle of decisions and tests rather than a one-off report.

How do you get started without a data science team?

You need two to three years of weekly data per channel if you can get it, plus your outcome metric and any big non-marketing drivers such as price changes or seasonality. Start in Google's demo notebooks, fit a national model first, check that the channel estimates make sense against what you know, then add geo-level data and your first experiment.

Meridian also ships a no-code Scenario Planner (in open beta) for exploring what-if budgets, and its budget optimiser supports fixed budgets, target ROI and target marginal ROI. Treat the first model as a draft. Its job is to tell you which experiment to run next.