When Does a DTC Brand Need Incrementality Testing or MMM?

A DTC brand needs incrementality testing as soon as one channel takes a large share of the budget and nobody can say how many of its sales would have happened anyway. It needs marketing mix modeling later, once it has about two years of weekly data across several channels and decisions that span all of them. For most brands between $5 million and $50 million in revenue, that order is test first, model second.

This guide sets out what each method answers, the data each one needs, and how to run a holdout test without a modeling budget. Platform details are as of 25 October 2026.

What is the difference between incrementality testing and marketing mix modeling?

Incrementality testing measures what one channel or campaign added by comparing a group that saw the ads with one that did not, while marketing mix modeling estimates the contribution of every channel at once from historical spend and sales.

Both exist because platform reports measure something else. Google's own researchers wrote that click metrics do not indicate how the consumer would have behaved in the absence of the advertising. An ad can be clicked by someone who was going to buy anyway.

The gap can be large. In 12 Facebook ad studies, academic and Facebook researchers found that non-experimental measurement methods often failed to match the results of true experiments, even when they used thousands of behavioral variables. The studies date from before iOS 14 and compared statistical methods rather than platform dashboards, and the gap between methods is the point.

When does a DTC brand need marketing mix modeling?

A DTC brand needs marketing mix modeling when it spends across several channels, has about two years of clean weekly data, and is making budget decisions that a single test cannot settle.

The data requirement is the real gate. Google's open-source model, Meridian, asks for a minimum of two years of weekly data for a model built on regional data and more for a national one. Meta's open-source Robyn sets the same two-year weekly minimum. Neither publishes a revenue or spend threshold.

Variation matters as much as history. For US brands, Meridian suggests modeling the top 50 to 100 media markets by population. Its documentation shows that two years of national weekly data was too little for a model with many channels, and notes that insufficient variation in the media spend weakens national models. A brand that has spent the same amount on the same channels every week gives a model little to learn from.

So the answer to "does a $20 million brand need MMM?" depends on its data. A brand with two years of weekly spend across four or more channels and sales by region can run one usefully [CB/I view]. A brand with national data only, few channels and flat spend will get wide error bars and a slide deck, not a decision.

Meridian's code is free, but preparing data and running it takes analyst time or a partner. Commercial vendors do not publish prices, and third-party estimates vary widely.

Start with an incrementality test

Most brands at this size should test their largest channel first, because the result is useful on its own and later calibrates any model.

Both open-source models are built to absorb test results. Robyn's guidance strongly recommends calibrating a model with experimental results treated as ground truth, and Meridian accepts experiments as inputs.

Platform lift tests are the easiest route when a brand qualifies.

• Google Ads Conversion Lift needs at least 1,000 observed conversions and a $5,000 minimum budget, and is set up through a Google account representative. It now covers Search and Performance Max as well as video and Demand Gen.
• Meta Conversion Lift is, in Meta's words, currently limited, with access through a Meta representative.

For many brands between $5 million and $20 million, the 1,000-conversion requirement is the binding constraint, more than the budget [CB/I view].

How to run a holdout test without an MMM budget

A brand without lift-test access or modeling budget can run a randomized geographic holdout, switching ads off in some regions and comparing sales with the regions where ads stayed on.

The method comes from a 2011 Google research paper by Vaver and Koehler, and Meta's free GeoLift tool builds on the same idea with a synthetic control.

1. Choose regions, such as US states or Nielsen designated market areas, where the brand can target ads and see sales from Shopify orders.
2. Group regions by size, then randomly assign each group to test or control. The paper found grouping by size before assigning can narrow the result's confidence interval by 10% or more.
3. Check the design on past data first, to see what size of effect the test could detect.
4. Run a pretest period, then a test period of at least one full purchase cycle, with ads off in the holdout regions.
5. Compare revenue in test and control regions against the pretest gap, and divide the difference by the spend to get incremental return on ad spend.

Avoid the shortcut of comparing one city with another. The same Facebook study simulated matched-market tests where the true lift was 33%, and found estimates could land anywhere from -2% to 80%. Many randomized regions and repeated tests give a reading a leadership team can act on.

What to do with the result

Compare the incremental return with what the platform reports, and adjust targets by the gap.

Common Thread Collective's Q1 2026 benchmarks, built from its proprietary geo holdouts and models, show how far apart the two can be. Google brand search reported a median 19.07x return but an estimated 5.72x incremental return, while Meta acquisition campaigns came out higher on an incremental basis than in-platform. Those are one agency's proprietary estimates, so use your own test, not theirs.

Between tests, track the ratio of total revenue to total ad spend each week. Orders that begin in AI assistants such as ChatGPT arrive without an ad click, and a holdout on total revenue captures them too. Margin-based targets are worked through in what is a good ROAS for a DTC brand, and which metrics a DTC brand should review every week covers the weekly view.

Frequently asked questions

Does a $20 million DTC brand need marketing mix modeling?

Only if it has about two years of weekly data, several channels and real variation in spend. Otherwise a lift test on the largest channel answers more for less.

How long should an incrementality test run?

Long enough to cover at least one full purchase cycle. Google allows Conversion Lift studies as short as 7 days but typically recommends more than 14, aiming for 90% certainty. Check a geo test's design on past data before starting.

Can we trust platform-reported ROAS?

Use it to compare campaigns within a platform, not to decide budgets across channels. Platform reports count conversions they touched, not the sales they caused.