Short answer: build a Google Ads measurement stack with three deliberately different jobs. Use attribution for timely conversion-path and campaign diagnostics, use incrementality testing for a defined counterfactual such as a geo holdout, and use media mix modeling (MMM) for longer-horizon channel and budget questions. They can disagree because they measure different things. Define the decision and outcome first, then reconcile the evidence instead of choosing a single dashboard as universal truth.
Google’s 2 September 2026 Ads Decoded episode discusses attribution, periodic incrementality tests and MMM together, including longer purchase journeys and Meridian. This article turns that strategic framing into an operating checklist. It does not report a DMT campaign test, a customer result or a fitted model.
Why the methods disagree
Measurement disagreements are often a question-definition problem. “Which campaign received credit?” is different from “Would the business outcome have happened without the campaign?” and both differ from “How should next quarter’s total budget be allocated across channels?” Attribution, incrementality and MMM answer those questions on different time scales and with different data.
Google’s attribution-model help explains the current Google Ads model landscape, including data-driven attribution and last-click availability. The attribution-report help describes conversion paths and lookback windows. These reports are valuable for operational diagnosis, but an attributed conversion is not presented here as a universal causal estimate.
The three-layer measurement stack
| Layer | Question it answers | Typical output | Use it for |
|---|---|---|---|
| Attribution | Which eligible interactions appear in the conversion path? | Credited conversions, paths, campaign/ad diagnostics and timely reporting. | Creative, query, audience and in-channel optimisation. |
| Incrementality | What changed because the intervention was present rather than held out? | Lift or incremental outcome for a defined test population and period. | Channel funding and causal validation. |
| MMM | How does historical variation relate to outcomes across channels and controls? | Contribution, response curves and scenario/budget guidance. | Aggregate planning and longer-horizon mix decisions. |
The table is an operating framework, not a Google product requirement. A small business may begin with clean outcomes and attribution, add one well-designed test, and only then consider MMM. A larger team may run all three continuously. The correct sequence depends on the decision, data coverage and ability to act on the result.
How to build the stack
1. Write the decision and estimand
Write one sentence before opening a report: “We need to decide whether to keep, reduce or increase channel X for outcome Y over period Z.” Add whether the question is attributed credit, incremental lift or aggregate contribution. This prevents a path report from being used to answer a budget-allocation question it was not designed to answer.
2. Make the outcome trustworthy
Choose a business outcome that the team can reconcile: qualified lead, accepted opportunity, sale, margin or another approved KPI. Record event definitions, deduplication, time zone, delayed conversions, refunds and offline joins. If the outcome is not stable, a more complex model only produces a more elaborate explanation of bad input.
3. Keep attribution useful but bounded
Use Google Ads attribution reports for paths, lag, conversion-window and campaign diagnostics. Record the model and lookback window with every extract. Google’s data-driven attribution help describes how converting and non-converting paths inform credit and gives account-level guidance about data volume; treat its recommendations as guidance, not a universal pass/fail threshold.
Do not compare two channel totals until model, conversion action, window, timezone and counting rules are aligned. If Google Ads, analytics and CRM totals differ, reconcile definitions and delays first. The site’s Google Ads API release guide is useful when a data-pipeline change may affect extraction; an API release alone does not explain a business-performance change.
4. Use incrementality for a causal question
An incrementality test needs a defined intervention, a treatment and a credible comparison, a primary outcome and a pre-specified analysis window. A geo holdout can be appropriate when regions are comparable and interference is manageable; another design may suit a different channel. Protect the test from simultaneous budget, landing-page, pricing or promotion changes that would make the causal question ambiguous.
Decide in advance what result would change the budget decision, how uncertainty will be reported and how the test will be stopped. Do not call a before/after change “incremental” without a counterfactual. A platform lift result can be useful, but its population, estimand and eligibility still need to be documented.
5. Add MMM for aggregate planning
Google’s Meridian data guide describes organising KPI, media and control data. The model specification describes a geo-level hierarchical model and transformations such as carryover and saturation. If you have spend but not exposure, record spend as a proxy and its limitations; cost changes can make spend a weak representation of delivered media.
MMM is not a magic truth layer. Meridian’s model-fit guidance explains that causal estimation is difficult to validate directly and recommends experiments with the same estimand where possible. Review holdouts, residuals, sensitivity, priors, controls, saturation assumptions and data coverage. An excellent fit to historical data can still produce a weak budget recommendation if the causal variation is not credible.
How to reconcile conflicting outputs
- Align definitions. Confirm the same outcome, conversion action, date range, timezone, attribution window and campaign scope.
- Label the estimand. Put “credited”, “incremental” or “modelled contribution” beside every number in the review.
- Check data boundaries. Note click/impression coverage, offline sales, organic demand, seasonality, promotions, channel overlap and missing controls.
- Prefer causal evidence for causal decisions. Use a well-designed experiment to calibrate or challenge a path-based story; do not manufacture a causal conclusion from correlation.
- Record the decision. State which evidence changed the budget, what remains unknown and when the next test or model refresh is due.
This approach also keeps adjacent channel reports readable. The site’s unified attribution guide covers another platform context; the YouTube views versus engaged views guide defines a metric. Neither should be treated as an incrementality result. For the wider tooling layer, see the martech stack guide.
A practical small-team path
- Document the approved KPI, conversion definitions and CRM join.
- Set one stable reporting view with model, window, timezone and data freshness visible.
- Use attribution for within-channel diagnostics, not as the sole budget truth.
- Run one bounded lift or geo test when a material channel decision is at stake.
- Only adopt MMM when historical media, outcome and control data are sufficient for the question, and keep its assumptions inspectable.
- Review the result with finance, marketing and analytics together; do not let the platform being evaluated define every success rule.
For campaign-surface context, the site’s Demand Gen release guide can sit beside the measurement record. It should not be used to infer the performance of a campaign that was not tested.
Likewise, the AI for Google Ads guide covers automation use cases, while this page owns the measurement decision. Keep the implementation and evaluation questions linked but separate.
Frequently asked questions
Which method is the source of truth?
None is universal. Attribution, incrementality and MMM answer different questions. Choose the decision and estimand first, then use the appropriate layer and reconcile conflicts explicitly.
Is data-driven attribution incremental?
Do not make that equivalence. Google documents data-driven attribution as a way to assign credit across paths. Incrementality asks what changed against a counterfactual. They can inform each other but are not interchangeable by definition.
When should a small business use MMM?
When its historical media, outcome and control data can answer a specific aggregate planning question and the team can act on the result. Start with measurement hygiene and a bounded experiment if those prerequisites are not ready; do not invent a minimum spend or data volume.
Can I use spend as the only MMM input?
Spend can be a documented media input, but Google’s Meridian guidance distinguishes media data and controls and warns that proxies have limitations. Record what spend represents, what is missing and how that uncertainty affects the decision.
Bottom line: let attribution explain recent paths, let incrementality test a defined counterfactual and let MMM inform aggregate planning. The quality of the stack comes from clear outcomes, aligned definitions, credible data and recorded uncertainty—not from declaring one dashboard the truth.
