Attribution, Incrementality, and MMM

Marketing does not need one source of truth.

It needs a system for making different kinds of decisions without pretending one measurement method can answer all of them.

Attribution tells an operator which observed touchpoints received credit. Incrementality estimates what happened because marketing ran. Marketing mix modeling estimates how channels and other factors relate to business outcomes over time and supports higher-level allocation.

Those are not three vendors competing to own the dashboard. They are three lenses.

The failure begins when leadership asks one lens to do another's job. A platform attribution report is used to claim causality. One lift test becomes the permanent value of a channel. An MMM budget recommendation is pushed directly into tomorrow's campaign. Contradictions are treated as errors to average away.

A serious measurement system does the opposite. It assigns each method a decision, a cadence, an owner, and a known limit. Then it uses disagreement as a reason to investigate.

The Search for One Number Creates False Confidence

The appeal of a single marketing truth is obvious. Finance wants a return. Growth wants to know where to move budget. The board wants a concise answer. Multiple numbers feel like weak measurement.

But marketing decisions occur at different speeds and levels.

A media buyer deciding which creative to pause this afternoon does not need a quarterly econometric model. A CFO approving a $5 million channel allocation should not rely on a seven-day platform attribution window. A brand deciding whether paid search creates demand or harvests it needs a counterfactual, not another rule for distributing credit.

Each method simplifies reality in a different way:

• Attribution observes journeys and applies credit rules.
• Incrementality constructs or estimates a no-marketing comparison.
• MMM models aggregate relationships across time, channels, and external variables.

The goal is not to eliminate assumptions. That is impossible. The goal is to make the assumptions appropriate to the decision and visible to the people funding it.

This builds on the three-layer measurement stack for a post-cookie world. The next step is governance: deciding which evidence is allowed to support which claim.

Attribution Is an Operational Navigation System

Attribution connects observed conversions to marketing touchpoints according to a defined rule. Last click, first click, data-driven attribution, view-through attribution, and multi-touch models all assign credit differently.

That makes attribution useful for:

• Diagnosing campaign and creative movement quickly
• Comparing outcomes under one consistent reporting rule
• Finding tracking breaks and unusual changes
• Feeding conversion or value signals back to platforms
• Generating hypotheses for budget and customer-journey decisions

It is fast, granular, and always available. Those are real strengths.

It is also the layer that decides what the platforms optimize toward, which is why picking the wrong conversion event costs more than assigning credit to the wrong channel. Attribution reports a mistake once. A biddable event repeats it every day until someone changes it.

Its limit is causal interpretation. Attribution does not observe the same person both exposed and unexposed. It does not know with certainty whether a conversion would have happened anyway. A branded search click may receive credit for demand created by a podcast, an organic recommendation, a retail visit, or prior customer experience. A paid social view may be genuinely influential or merely present.

Platform attribution carries an additional conflict: the company selling the media defines the measurement environment. That does not make the report fraudulent. It makes it a platform claim that should be reconciled with the company's revenue and customer data. Teams that skip the reconciliation are usually the ones surprised by how far platform-reported results sit from the P&L.

Use attribution to steer. Do not use it alone to prove.

Incrementality Estimates the Counterfactual

Incrementality asks a different question:

What changed because this marketing activity ran, compared with what would likely have happened without it?

The strongest designs create treatment and control groups through randomization. Depending on the channel and business, that may involve user-level holdouts, geographic experiments, conversion-lift studies, matched markets, or other causal designs.

Incrementality is useful for:

• Testing whether a channel creates net-new outcomes
• Measuring cannibalization of organic or direct demand
• Calibrating platform-reported performance
• Estimating incremental CAC or ROAS
• Evaluating major changes in spend, audience, or strategy

It also has limits.

Experiments require enough signal. Holding out media has an opportunity cost. Contamination can occur when people move across geographies, devices, or channels. A result applies to the tested spend, creative, audience, market, and time—not automatically to every future condition.

Access to that evidence got cheaper in 2025. Google says a study that "once might have cost upwards of $100,000" can now be run for $5,000, and that it is moving Conversion Lift from frequentist to Bayesian estimation specifically so studies can resolve on smaller budgets and fewer conversions. See Google's incrementality update and its Conversion Lift methodology.

Read the eligibility detail before planning around the headline. Google documents a $5,000 minimum spend alongside conversion minimums—roughly 150 conversions in the treatment group and 65 in the control—and it reports a study only once the estimated certainty of lift clears 50%. That threshold means "more likely than not," which is a long way from proven. Cheaper to run is not the same as powered to answer your question.

Lower access thresholds put experimentation within reach of more advertisers. They do not turn every study into independent truth. A platform-run lift test is still designed, delivered, and measured inside the platform. It is valuable evidence with a disclosed source.

MMM Supports Portfolio Allocation

Marketing mix modeling uses aggregate time-series data to estimate the relationship between media, business outcomes, and other drivers such as seasonality, pricing, promotions, or economic factors.

MMM is useful for:

• Comparing channels on one business outcome
• Estimating saturation and diminishing returns
• Planning quarterly or annual budget ranges
• Evaluating channels that are hard to track at the user level
• Separating baseline demand and non-media factors from media effects
• Running allocation scenarios

It is particularly helpful when customer-level tracking is incomplete or inappropriate. It does not need to reconstruct every individual journey.

Its limitations are different from attribution and experimentation. MMM needs enough historical variation to identify relationships. Correlated channel spending can make separation difficult. Results depend on model specification, priors, data quality, selected controls, and the stability of the underlying business. A model can be mathematically polished and still encode a weak business story.

Google made its open-source Meridian model broadly available in 2025 and describes support for reach and frequency, prior knowledge, experiments, and budget optimization in the Meridian release. Its technical documentation makes the assumptions inspectable, including media transformations and prior distributions in the model specification.

Open source improves transparency. It does not remove the need for statistical judgment or business review.

Use MMM to plan the portfolio. Its output is a range for how weight should sit across channels over a quarter or a year. Do not ask it which ad to pause at 2 p.m.

Which Lens Should Answer Which Decision?

This decision matrix is the core operating artifact.

Decision	Primary lens	Supporting lens	Cadence	Main failure mode
Pause or expand a creative	Attribution	Business conversion quality	Daily or weekly	Mistaking credited conversions for caused conversions
Change an optimization event	Attribution	Cohort quality and lift test	Weekly or monthly	Training toward volume that does not become profit
Decide whether brand search is incremental	Experiment	Attribution for diagnosis	Periodic	Contamination or insufficient power
Approve a major channel increase	Experiment or calibrated MMM	Attribution for execution	Monthly or quarterly	Assuming the current return holds at higher spend
Set annual channel ranges	MMM	Experiments as calibration	Quarterly or annual	Poor controls, correlated spend, stale history
Explain total revenue movement	MMM plus finance bridge	Attribution and experiments	Monthly or quarterly	Ignoring pricing, product, sales, or seasonality
Diagnose an abrupt tracking drop	Attribution and source data	Experiment results as context	Immediate	Treating a measurement failure as demand loss

The “primary lens” is the evidence most suited to the claim. The supporting lens adds context or calibration. It should not be averaged mechanically with the primary lens.

One row deserves extra attention. Approving a major channel increase is a question about what the next dollar buys, not what the average dollar bought. Historical return is the wrong evidence for an expansion decision no matter which lens produced it.

The matrix should be adapted to the business. A low-volume enterprise company may rely more heavily on pipeline stages and longer matched-market tests. A high-volume ecommerce company may run more frequent user- or geo-level experiments. The principle remains: measurement architecture follows the decision.

What Should You Do When the Lenses Disagree?

Suppose platform attribution reports a 5.0 revenue ROAS, a lift test estimates 1.8 incremental ROAS, and MMM estimates a channel ROI range centered around 2.4.

The wrong response is:

(5.0 + 1.8 + 2.4) / 3 = 3.07, so the truth is 3.07

The numbers do not measure the same object under the same assumptions. Averaging them creates a number with no coherent meaning.

Ask why they disagree.

Platform attribution may include conversions that would have happened organically. The lift test may capture a specific geography, creative set, and spend level during a short period. MMM may estimate longer-term effects but struggle to distinguish channels that always move together.

Use a reconciliation protocol:

1. Align the outcome. Are all methods measuring net revenue, contribution, qualified pipeline, or something else? Platform revenue and contribution after every variable cost are not the same object, and three lenses pointed at three different objects will never agree.
2. Align the horizon. Are they evaluating seven days, six weeks, or twelve months?
3. Align the population. New customers, all customers, one geography, or the whole business?
4. Align the spend level. Average historical spend or the current marginal budget?
5. Inspect contamination and overlap. Did other channels change during the experiment?
6. Inspect model assumptions. Which priors, controls, attribution windows, and lag effects matter?
7. Choose the decision standard. Which evidence is strongest for the specific capital decision?

Disagreement can reveal brand-demand harvesting, customer-quality differences, delayed effects, cross-channel reinforcement, or a tracking defect. That is information. Do not erase it.

Establish an Evidence Hierarchy Without Worshiping It

Randomized experiments generally provide stronger causal evidence than observational attribution. That does not mean every experiment is better than every model or that experimentation can answer every question.

Use an evidence hierarchy as a starting point:

1. Randomized experiment on the relevant population and spend level
2. Well-designed quasi-experiment or matched-market analysis
3. MMM calibrated with relevant experiments and business controls
4. Consistent observational analysis across source data
5. Platform attribution
6. Platform benchmark or case study

Then add relevance.

An excellent experiment from eighteen months ago may be less decision-useful than a carefully reviewed current model after the product, pricing, and channel mix changed. A vendor case study can show product direction but should not become your forecast. A platform lift study can be strong within its scope and still need reconciliation with bankable outcomes.

Confidence should depend on design quality, relevance, recency, and agreement with business data. No method gets permanent authority.

Create a Measurement Charter

The board does not need every methodological detail. It does need to know what can be claimed.

Create a one-page charter with these fields:

Field	Required definition
Business outcome	Net revenue, gross profit, contribution, qualified pipeline, or another approved outcome
Attribution role	Which daily and weekly decisions attribution may support
Experiment roadmap	Which high-value uncertainties will be tested and when
MMM role	Which planning decisions the model may support
Source labeling	How platform, company, experimental, and modeled results are distinguished
Confidence scale	What high, medium, and low confidence mean
Decision rights	Who may scale, hold, revise, or stop spend
Reconciliation cadence	When contradictory evidence is reviewed
Model and test expiration	When evidence must be refreshed

The charter prevents measurement shopping. A team cannot switch from finance revenue to platform-reported revenue because one makes a campaign look better. A board cannot demand experimental certainty for every small creative decision. Everyone knows what evidence is fit for purpose.

It also clarifies ownership. Growth owns the questions and operating actions. Data or analytics owns instrumentation and method quality. Finance owns recognized outcomes and economic definitions, which is why growth leaders who can read the P&L they are judged against negotiate better charters than those who cannot. Leadership owns the risk tolerance. Agencies and platforms supply evidence; they do not adjudicate the business case.

Run the System at Three Speeds

A three-lens system needs three cadences.

Weekly: operate

Review attribution movement, spend, conversion quality, tracking health, and exceptions. Make reversible campaign decisions. Do not relitigate the entire channel portfolio every Monday.

Monthly: reconcile

Compare platform-attributed outcomes with CRM, commerce, and finance records. Review recent cohorts. Check experiment progress and model assumptions. Identify discrepancies that could change budget.

Quarterly: allocate

Review completed lift tests, refresh or inspect MMM, examine saturation, update allowable CAC and contribution requirements, and set channel ranges. Write the assumptions behind every material change.

The speeds should connect. Weekly attribution generates hypotheses for experiments. Experiments calibrate models. Models identify expensive uncertainties worth testing. Finance data keeps all three tied to outcomes the business recognizes.

This is a system, not a maturity ladder where MMM replaces attribution or experiments make dashboards obsolete.

Report Confidence Before Precision

A board measurement page can fit in six lines:

1. Business outcome: contribution generated and change versus plan
2. Attributed view: where platforms and analytics assigned observed outcomes, including the channel splits reported inside automated campaigns
3. Causal view: completed and active experiments, lift range, and limitations
4. Modeled view: channel contribution range and saturation direction
5. Reconciliation: where the lenses agree or conflict
6. Decision: scale, hold, revise, or stop—with confidence and next evidence date

Use ranges where the method produces uncertainty. Name vendor-reported figures as vendor-reported. Separate modeled outcomes from booked revenue. Record when a test or model no longer reflects the current business.

This is not hedging. It is more accountable than placing an unjustified decimal point on a weak estimate.

Choose the Decision, Then Choose the Lens

Attribution, incrementality, and MMM do not compete for one measurement throne.

Attribution helps the team operate at speed. Incrementality estimates causal lift for a defined intervention. MMM helps leadership allocate a portfolio across longer horizons. Each fails when asked to impersonate the others.

Start with the decision. Choose the primary lens. Add supporting evidence. Label the source and uncertainty. Investigate disagreement. Then make the action explicit.

If you need to decide which campaign signal changed today, use attribution. If you need to know whether the channel created outcomes that would not otherwise exist, run an experiment. If you need to set next quarter's ranges across a complex portfolio, use a calibrated model. If the decision is material, reconcile all three with contribution the business actually recognizes.

The goal is not one perfect number. It is fewer expensive decisions made with the wrong evidence.

If you need an owner-operator to build that measurement charter, apply to work with us.