MMM vs MTA vs Incrementality: Three Different Questions
Marketing mix modelling estimates channel-level contribution from aggregate historical data. Multi-touch attribution splits credit for sales that already happened across the touches in each path. Incrementality testing measures whether a campaign caused a sale that wouldn't have happened otherwise. They're not competing options — each answers a genuinely different question, and most brands eventually need more than one.
What this looks like in a real account
The question each method actually answers
It's easy to talk about these three as if they're rival approaches to the same problem, chosen based on sophistication or budget. They're not rivals — they're answering different questions, and confusing which one you're asking is the root of most measurement disagreements. MMM answers "how much did each channel, in aggregate, contribute to sales over this period." MTA answers "which touches, within a path that already converted, should get credit for that conversion." Incrementality testing answers "did this specific spend cause sales that wouldn't have happened without it." Only the third one is a genuinely causal question — the other two describe patterns in data that already exists.
Data requirements, side by side
MMM needs aggregate spend and sales data over a long history — typically 18-24 months minimum — with real variation across channels to model against; it doesn't need shopper-level data at all. MTA needs shopper-level, cross-channel path data, which on Amazon means Amazon Marketing Cloud specifically, since no single ad product's native reporting stitches paths across channels. Incrementality testing needs either a randomised audience holdout or matched geographic markets, held constant for a defined test period — it's the most operationally demanding of the three, because it requires actually withholding spend from part of your addressable audience, not just analysing data you already have.
Why none of the three replaces the other two
MMM can tell you DSP contributed an estimated share of sales over the last year, but it can't tell you which specific audience or creative within DSP did the work, and it can't distinguish correlation from causation on its own — a channel that happens to scale up during your strongest sales months will look more contributory than it actually is unless the model carefully separates seasonality from spend effect. MTA can tell you a specific path pattern — DSP display followed by branded search — is common among converting customers, but it can't tell you whether removing the DSP touch would have cost you the sale, only that it was present. Incrementality testing directly answers the causal question but only for the specific audience, channel or market it was designed to test — it doesn't automatically generalise to every other channel or the whole account.
A worked illustration of how they'd disagree, and why that's useful
Say MMM estimates DSP contributed 22% of sales over the last year. A path-to-conversion MTA query shows DSP appears in 35% of converting paths. A geo-lift incrementality test on DSP specifically shows a 9.5% lift in sales in markets where DSP ran, versus matched markets where it didn't. All three numbers are legitimate and none of them contradicts the others — they're measuring different things at different altitudes. The MMM figure describes DSP's estimated aggregate contribution across a year of variation; the MTA figure describes how often DSP shows up in a path; the incrementality figure describes the actual causal lift from a specific test. If you need one number for a board deck, the incrementality figure is the most defensible, because it's the only one of the three answering the causal question a board is actually asking.
The common mistake, including ours
The mistake is presenting whichever of the three numbers happens to be most favourable for a given argument, without naming which question it's actually answering. We've watched — and, early on, done ourselves — a media plan get defended with an MTA path-frequency number when the actual question being asked was a causal one, because the MTA number was available and the incrementality test hadn't been run yet. It wasn't a lie; it was answering a different, easier-to-produce question and letting the audience assume it answered the harder one. We now name the method explicitly next to any number in this territory — "path frequency," "modelled contribution," "measured lift" — so the audience knows exactly what kind of claim they're looking at.
Why all three still leave a gap
Even running MMM, MTA and incrementality testing together doesn't produce a single, unified number — and that's worth accepting rather than fighting. Each method is scoped to what it can actually measure, and forcing them into one blended figure usually means quietly picking whichever assumptions make the three methods agree, rather than reporting the honest, sometimes uncomfortable spread between them. The more useful habit is reporting all three side by side with their method named, and letting a reader see where they agree — which is where confidence is genuinely highest — and where they diverge, which is where the next test should be aimed.
How to sequence the three in practice
Start with MTA-style path analysis in AMC — it's the cheapest to build and gives you a map of your account's actual structure. Use that map to identify which channels or audiences look worth a closer causal test. Run incrementality tests on those specific candidates, prioritising your largest spend lines first. Build MMM once you have enough spend history and variation for it to be reliable, and use it as a periodic, top-down sanity check against what the bottom-up incrementality tests have already shown you — not as the first or only tool in the stack.
| Method | Question answered | Data needed |
|---|---|---|
| Marketing mix modelling | Aggregate channel contribution over time | 18-24+ months of channel-level spend and sales history |
| Multi-touch attribution | Credit split across touches in converted paths | Shopper-level cross-channel path data (AMC on Amazon) |
| Incrementality testing | Did the spend cause sales that wouldn't have happened | A held-out audience or matched control markets |
Which one you should actually pick
Any brand with an analyst can build MTA path analysis in AMC and run incrementality tests without a vendor — MMM is the one of the three that genuinely benefits from dedicated data science capability at scale. reMKTR runs all three in sequence on managed accounts, naming which question each number answers before it reaches a client, as part of the same discipline behind Full Circle's $500M+ in managed Amazon spend across 100+ brands.
Shortlist on the job, not the feature grid. Pull your search-term report for the last 90 days and total the spend against terms that produced no orders — 33.6% on the account above. Then ask each vendor on your list what they would do about it in week one, and see who answers with a process rather than a screenshot.
Common questions
Which of the three should I invest in first?
Start with MTA-style path analysis in AMC to map your account's structure, then use incrementality testing on the channels or audiences that map flags as worth a closer causal check. Build MMM once you have enough spend history for it to be reliable.
Can MMM and incrementality testing disagree without either being wrong?
Yes — they measure different things at different scales. MMM gives an aggregate, modelled estimate over a long period; incrementality testing gives a direct, measured result for a specific test. Both can be legitimate at once.
Is incrementality testing always the most trustworthy of the three?
It's the only one directly answering a causal question, which makes it the strongest evidence for a specific budget decision — but it only covers what was actually tested, so it doesn't replace the broader pattern-finding that MTA and MMM provide.
Do I need Amazon Marketing Cloud for all three methods?
You need it for MTA on Amazon specifically, and it's the most practical environment for running Amazon-side incrementality tests. MMM typically draws on your own aggregate spend and sales reporting rather than AMC directly.
We show the method before the number.
Claim the free auditRead next
- Criteo Pricing: The Fee Stack Inside Your BudgetPricing · criteo pricing
- Kenshoo Review: The Name Was Retired in 2021Review · kenshoo review
- Pacvue Pricing: No Public Number — What to AskPricing · pacvue pricing
- Skai vs Pacvue: Contracts, Not Feature GridsHead to head · skai vs pacvue