Attribution vs. MMM vs. incrementality: what each one measures

What is the difference between attribution, marketing mix modeling and incrementality testing, and which one do I need?

Measurement debates often sound like a contest: attribution is dead, MMM is back, only experiments count. The contest is mostly a confusion. Attribution, marketing mix modeling and incrementality tests answer three different questions, with different data, at different speeds. Once you know which question you are asking, the choice of method mostly makes itself.

Three questions, three methods

Attribution asks: which touches came before this sale? It follows individual journeys (clicks, visits, views) and assigns each sale's credit among them with a model. Its strength is speed and detail: per campaign, per ad, per day. Its limit is that it describes what happened before a sale, not what caused it. How the credit is split is covered in attribution models explained.

Marketing mix modeling asks: how much does each channel contribute to total sales over time? It uses aggregate data, typically weekly spend by channel, sales, prices, seasonality and promotions, and estimates each channel's contribution statistically. It needs no cookies and can include channels attribution cannot see, such as TV or out-of-home. Its limits are that it needs a lot of history and it is weak on individual campaigns.

Incrementality testing asks: what did this campaign cause that would not have happened without it? It compares a group exposed to the campaign with a comparable group that was not, through a holdout audience or regions where ads were paused. It is the only one of the three that measures cause directly, and it is slow, narrow and costs the sales you hold back during the test.

A side-by-side

Where each one goes wrong

Attribution credits touches that happened to be there. A retargeting ad shown to people who were going to buy anyway collects credit for sales it did not cause. Platform attribution adds a second problem: each platform credits itself, so several can claim the same order. Checking every claim against your store's orders fixes the second problem; only tests address the first. We measure how much of a platform's claim can be verified in how much of what Meta claims can you verify.

MMM needs years of weekly data with enough variation in spend to separate channels, and correlated channels (you always raise Meta and Google together before a sale) are hard to tell apart. Its outputs are estimates with wide uncertainty, and a model is only as good as its assumptions about how long an ad's effect lasts.

Tests are expensive in lost sales, hard to run on small volumes and answer one question at a time. A test on one channel in November says little about another channel in May.

Tools you can use

Two major platforms publish open-source MMM packages. Google's Meridian is an open-source MMM built by Google, now with Meridian GeoX to calibrate the model with geo experiments. Meta's Robyn is an experimental, open-source MMM package from Meta Marketing Science, designed to be calibrated with ground-truth methods such as geo tests, lift studies or attribution. Both require someone comfortable with statistics and code.

For tests, the ad platforms offer lift studies to eligible advertisers, and a geo test (pausing or increasing spend in some regions and comparing with similar ones) can be run with nothing more than disciplined planning and your store's orders by region.

How they fit together

The methods work best as a system. Attribution runs daily and tells you which campaigns precede sales. Tests, a few times a year, tell you how much of that credit is real for your largest channels. MMM, if you get there, tells you how to split next year's budget across channels, calibrated by the tests. Each one corrects a blind spot of the others.

A practical order for most stores

• First, make attribution honest. Check each platform's claims against your store's orders, remove refunds and fix double counting. Everything else builds on numbers you can trust. The causes of the gap are in why Meta, Google, GA4 and your store never match.
• Then test your biggest bet. Pick the channel with the most spend and run one clean test, for example a geo holdout for four to six weeks.
• Consider MMM later. Once you have two or more years of weekly data, several channels and budget decisions across quarters, an open-source model becomes worth the effort.

Adray sits in the first step: order-level attribution across channels, with each platform's claim kept next to the verified orders. It does not sell media mix models or lift tests, and a store that needs them should use the tools above alongside it. For the concept of multi-touch attribution in depth, see our guide (in Spanish).