MMM vs. Multi-Touch Attribution vs. Incrementality Testing

For about fifteen years, digital marketing measurement had a default answer: track every click, stitch the touchpoints together, and give each channel its share of the credit. That default is gone. Apple's App Tracking Transparency gutted mobile signal in 2021, Safari and Firefox have blocked third-party cookies for years, consent banners remove a meaningful slice of European traffic from analytics entirely, and Google spent half a decade threatening to deprecate third-party cookies in Chrome before settling on a user-choice model in 2024. Click-level measurement didn't die — but it stopped being a complete picture.

Then AI assistants arrived as a genuine research channel. When a prospective customer asks ChatGPT, Perplexity, or Gemini “what's the best CRM for a 10-person agency?” and later Googles the brand it was told about, that entire research phase happens outside anything a tracking pixel can see. The journey's first — and often most decisive — touchpoint never enters the attribution graph. Analytics dutifully reports a “branded search” or “direct” conversion, and everyone congratulates the wrong channel.

This is why the industry conversation has shifted from “which attribution model is right?” to “which measurement method is right for which decision?” There are three families of methods, and they are not interchangeable — they answer different questions, at different speeds, with different blind spots:

• Multi-touch attribution (MTA) — bottom-up, user-level: reconstruct each customer's path from tracked touchpoints and divide credit among them.
• Marketing mix modeling (MMM) — top-down, aggregate: use statistics on weekly spend and outcome data to estimate what each channel contributed, no user tracking required.
• Incrementality testing — experimental: run controlled experiments (hold out a region or audience) to measure what marketing caused, rather than what it touched.

This guide compares all three in depth, shows how AI-mediated journeys change the math for each, and ends with a practical framework agencies can use to match methods to client budgets.

MTA tries to observe the actual path of each individual customer: the display impression, the two paid social clicks, the email open, the branded search, the conversion. It stitches these events together using cookies, device identifiers, click IDs, and login data, then applies an attribution model (last-click, position-based, data-driven, and so on — compared in detail in our attribution models guide) to split conversion credit across the touches.

When the tracking works, MTA is uniquely granular. It can tell you which campaign, ad set, creative, or keyword participated in conversions — daily, at the level where media buyers actually make decisions. That granularity is why in-platform attribution (Google Ads, Meta) and third-party attribution tools (compared in our attribution software guide) remain the day-to-day operating system of most agencies.

The problem is the word “when.” MTA only sees touchpoints it can technically observe and legally record. In 2026 that excludes: most view-through influence, anything on Apple devices post-ATT, users who decline consent, almost everything that happens inside walled gardens, word of mouth, podcasts and other offline media — and now the fastest-growing blind spot of all: recommendations made inside AI assistants, which pass either no referrer or a referrer only for the minority of sessions where the user actually clicks a cited link.

MMM is the opposite philosophy. Forget individuals entirely: take two to three years of weekly aggregate data — spend per channel, revenue or conversions, pricing, promotions, seasonality, even weather — and fit a statistical model that estimates how much each input contributed to the outcome. Good MMMs also estimate saturation curves (the point where extra spend stops paying) and adstock (how long a channel's effect lingers after the spend).

MMM is the oldest method here — consumer goods companies have used it since the 1960s to measure TV — but it has been reborn in the privacy era for one simple reason: it needs no user-level tracking at all. No cookies, no consent dependency, no pixel. Open-source libraries — Meta's Robyn and Google's Meridian being the most prominent — plus a wave of MMM SaaS vendors have collapsed the cost from six-figure annual consulting engagements to something a data-literate agency can operate in-house and refresh monthly.

The trade-offs are real, though. MMM outputs channel-level answers (“paid social contributed roughly 18% of revenue, and is near saturation”), not campaign-level ones. It needs history — a brand with 8 months of data and flat spend gives the model almost nothing to learn from, because MMM learns from variation: channels that never change budget are statistically invisible. And like any regression on observational data, it can be confidently wrong when two inputs move together (brand search spend and demand, famously).

Both MTA and MMM are ultimately observational — they infer from data that already happened. Incrementality testing runs an actual experiment: expose one group to marketing, hold another group back, and measure the difference in outcomes. The difference is the incremental effect — sales that happened because of the marketing, not merely alongside it.

The common designs, roughly in order of accessibility:

• Geo holdouts / geo-lift tests: turn a channel off (or up) in some regions and not others, then compare matched markets. Meta's open-source GeoLift package made this rigorous and free. Works for any channel — including ones with no tracking at all.
• Platform conversion-lift studies: Meta, Google, and TikTok can split users into exposed/held-out groups inside their own systems. Convenient and statistically clean, but you are trusting the referee who also plays for one of the teams.
• Audience holdouts: suppress a random slice of a retargeting or email list. The classic way to discover that retargeting “ROAS” was mostly harvesting people who would have bought anyway.

Incrementality is the closest thing marketing has to ground truth, which is why sophisticated teams use it to calibrate the other two methods. Its costs are equally concrete: tests take weeks, require enough conversion volume for statistical power, sacrifice revenue in holdout groups, and answer only one question at a time. You cannot run your whole media plan as a permanent experiment.

The rise of AI assistants as a research layer is not a niche concern. Google now shows AI Overviews on a large share of informational queries, and independent studies through 2025 consistently measured substantially lower click-through to websites when an AI Overview is present. ChatGPT alone reports hundreds of millions of weekly users. The pattern that emerges for brands is consistent: more research happens off-site, fewer clicks carry the intent signal, and the clicks that do arrive look like they came from nowhere — direct traffic or a branded search, with the actual influence hidden upstream. (We cover how to capture what is visible in our guide to tracking AI traffic.)

What this does to MTA

MTA suffers most, because its entire premise is observing the journey. The AI-assistant touch is usually invisible, so credit flows to whatever tracked touch happens to come last — typically branded search or direct. The practical symptom many teams see: branded search and direct conversions grow while attributed performance of prospecting channels stagnates, and last-click logic then argues for cutting the very activity (PR, content, reviews, community presence) that feeds the AI recommendations. MTA remains essential for tactical work, but treating its output as a causal map of demand creation is now more misleading than it was in 2019.

What this does to MMM

MMM is structurally indifferent to how a journey happened — it only needs aggregate inputs and outputs. If AI-driven discovery drives revenue, MMM will register the effect, but it will park it in the baseline or attribute it to correlated channels unless you model it deliberately. Forward-looking teams are starting to feed MMMs AI-era input variables: AI-referral sessions, AI crawler activity, branded search volume, and share-of-voice metrics from AI visibility monitoring (compared in our AI visibility tools guide). It's early, and data histories are short — but MMM is the natural home for measuring a channel you cannot track at user level.

What this does to incrementality

Least affected — and arguably more valuable than ever. An experiment doesn't care whether the influence path was trackable; it just measures outcomes in exposed versus held-out groups. Geo tests work even when the “channel” is something as fuzzy as a digital-PR push aimed at improving how often a brand is cited by AI assistants in one market. When tracking fails, experiments become the arbiter.

The consensus that has emerged among measurement practitioners — echoed by Google's and Meta's own measurement guidance — is triangulation: run the methods in parallel, use each for the decisions it's built for, and reconcile disagreements with experiments.

1. 1MMM sets the strategy. Quarterly or monthly: channel budgets, saturation checks, offline/online balance.
2. 2Incrementality calibrates. A rolling program — even just one geo or holdout test per quarter on the biggest or most doubted line item. Feed results back as priors/constraints into the MMM and as sanity checks on attribution numbers.
3. 3Attribution runs the day-to-day. Creative, audience, bid, and budget-pacing decisions inside channels — treated as a relative signal (“A outperforms B”), not an absolute truth (“A caused 214 sales”).

When the methods disagree — and they will — the disagreement itself is information. Attribution says retargeting is your best channel while a holdout shows near-zero lift? You've just learned your attribution is counting harvested conversions. MMM says TV works but attribution can't see it? That's expected; trust the method that can see the channel.

Most agency clients cannot afford — and do not need — the full stack. A practical way to scope measurement to client size:

For agencies specifically, there's a commercial angle worth naming: measurement design is becoming a sellable service line, not overhead. The agencies winning pitches in 2026 are increasingly the ones that walk in with a point of view on triangulation and AI-era blind spots — not the ones with the prettiest last-click dashboard.