When MMM and MTA Disagree: Which Number Governs Budget

Two measurement systems, two answers, and a budget meeting where nobody can agree which number is real. This is the routine end state for SaaS teams that run both marketing mix modeling (MMM) and multi-touch attribution (MTA): the attribution tool says paid search drives pipeline, the mix model says it barely moves the needle, and the room splits along whichever number favors their channel. The instinct is to pick a winner and standardize on it. That instinct is wrong. The disagreement between MMM and MTA is not a bug to eliminate — it is information about your funnel that a single number would hide.

The reason they disagree is structural, not fixable, and once you understand it the conflict becomes usable. MTA and MMM measure at different altitudes with different data, so they systematically credit different channels, and knowing which method is right about which layer is what turns two contradictory dashboards into one coherent budget decision. This post explains why the disagreement is predictable, how to assign each method the decisions it is actually good at, and how to use incrementality testing as the tiebreaker when a specific channel is in dispute. It builds on the broader problem of attribution model accuracy and error rates — here applied to the case where two models fight.

Why the disagreement is structural, not a bug

MTA works bottom-up from tracked user touches, so it can only credit what leaves a trail, and it naturally over-weights channels that sit close to the conversion. MMM works top-down from aggregate spend-and- outcome time series, so it credits channels by their statistical contribution to total revenue regardless of whether they left a clickable touch. These are not competing attempts at the same measurement; they are measurements of different things. Expecting them to agree is like expecting a thermometer and a barometer to report the same number — both are correct instruments pointed at different properties of the same weather.

Because the methods differ structurally, their disagreements fall into predictable patterns. MTA overvalues demand-capture channels: branded search is the classic case, because it "converts beautifully and creates almost nothing," as one measurement teardown puts it (Improvado). Someone else created the demand; branded search just intercepted it at the moment of purchase and MTA hands it the credit. MMM, conversely, credits upper-funnel and offline activity — brand campaigns, podcasts, events — that MTA cannot see because those touches never became a tracked click. Recognizing the pattern lets you predict, before you even look, which method will inflate which channel.

Assign each method the decisions it is good at

The reconciliation starts by refusing to make MMM and MTA compete for the same job. They operate at different grains, so give each the decisions that match its grain: MMM for budget allocation across channels, MTA for optimization within a channel. MMM sees cross-channel and saturation effects and answers the quarterly question of how much total budget each channel deserves. MTA operates at the campaign, creative, and keyword level MMM cannot resolve, and answers the weekly question of which specific campaigns inside a channel to scale or kill. Neither can do the other's job well, so a channel-level MMM verdict and a campaign-level MTA verdict are not in conflict — they govern different decisions.

This division is the mainstream position among measurement practitioners: run MMM for quarterly budget allocation, MTA for daily and weekly campaign optimization, and reconcile the two with experiments (Measured). Once you frame it as "two methods, two altitudes, here is which one governs which decision," the budget meeting changes character. A disagreement you can explain is a far stronger position in front of a CFO than a single number you cannot defend — when finance asks why the figure differs from last quarter, altitude is the answer. This is the same clarity that comes from a well-chosen north-star metric framework: the goal is not one number, it is the right number for each decision.

Incrementality as the tiebreaker

When MMM and MTA disagree about a specific channel and the budget stakes are real, neither correlation-based method can settle it — you need an experiment. Incrementality testing measures cause directly by withholding a channel from a comparable group and comparing outcomes, which gives you experimental proof of lift rather than a modeled inference. A geo-based holdout — turn a channel off in a set of matched regions and measure the conversion difference against control regions — is the standard technique, and its verdict does not depend on either MMM's or MTA's assumptions. That independence is exactly why it can referee them.

The practical protocol is to reserve incrementality tests for the disputes that matter. You cannot test every channel every quarter, so use the MMM-versus-MTA disagreement as the trigger: when the two methods tell different stories about a channel carrying meaningful spend, that channel earns a holdout. The Open Haus measurement teardown frames incrementality as the layer that "validates the model and resolves disputes when MMM and MTA tell different stories" (Haus). The result also feeds back: an incrementality test that contradicts MTA on branded search is evidence to down-weight MTA's credit there permanently, the same logic behind incrementality testing for top-of-funnel signal decay.

The reconciliation workflow in practice

Assemble the three methods into a standing workflow rather than treating each as a separate report. Weekly, MTA drives in-channel optimization — which campaigns to push, which to pause. Quarterly, MMM sets the cross-channel allocation and you compare its channel verdicts against the MTA picture. Where they align, you have high confidence and no further work. Where they diverge on a material channel, you queue an incrementality test, and its result updates both your trust in the models and, over time, the priors you bring to the next allocation. The output is not a single reconciled number but a governed process for producing defensible ones.

Documenting the reconciliation is as important as running it, because the value shows up in the budget conversation. Keep a short record of which method governs which decision, which channels have been incrementality- tested and what they returned, and how disagreements were resolved. When allocation changes, that record is your justification. This is the same discipline as maintaining a single source of truth for metric definitions — the reconciliation only holds if everyone agrees on the rules before the numbers arrive, not after one of them looks inconvenient.

What small and mid-stage SaaS teams should actually do

Most SaaS companies below significant scale do not need MMM at all, and forcing one produces a confident-looking model built on too little data. MMM needs a long history of spend and outcomes across enough channels for the statistics to separate one channel's effect from another's; a startup running two paid channels for eight months cannot support that. For these teams the honest stack is disciplined MTA plus periodic incrementality tests on the largest channels — the MTA tells you what to optimize, and the holdout checks whether the credit is real. That is enough to avoid the worst allocation mistakes without pretending to a rigor the data cannot bear.

Add MMM when spend and channel diversity grow to the point that cross-channel and saturation effects genuinely change allocation — when you are spending across search, social, display, and offline, and the question "are we over-invested in the channel that just captures existing demand?" has real money riding on it. Until then, the combination of clean attribution and targeted experiments is not a downgrade from MMM; it is the correct tool for your scale, the same way an early-stage team should skip formal attribution modeling until volume justifies it. Match the method to the decision and the data you actually have, and the disagreement between models stops being a crisis and becomes a routine, resolvable input to budget.