You've paid for the tracking, built the dashboards and sat through the model presentation. The budget meeting still ends with everyone backing their own number. It's a miserable return on the effort, and an expensive place to get stuck. Before commissioning the next piece of measurement, ask what it will help you decide. A growth leader allocating next year's budget, a channel manager investigating a conversion drop and a finance director challenging brand search need answers to different questions.
Attribution, marketing mix modelling and incrementality testing can each earn their place. The right choice depends on what's at stake, the data you've got, how soon you need to decide and the cost of being wrong. Sometimes combining methods is worth the effort. Sometimes the business already has enough evidence to act. You should be able to explain what the next piece of measurement will change before asking someone to pay for it.
Attribution: useful visibility, with boundaries
Attribution assigns credit for conversions to marketing interactions according to a set of rules or a model. It helps teams understand recorded customer paths, monitor performance and operate campaigns. If a campaign's reported conversions fall abruptly, attribution and tracking data can help identify whether the movement is concentrated in a device, market, landing page or conversion event. That operational visibility is valuable even when it can't establish the causal effect of the advertising.
The limitations follow from what's observed and how credit is assigned. Tracking gaps, consent choices, cross-device behaviour, platform boundaries and attribution windows can change the picture. A customer may interact with several channels before purchasing, and multiple platforms can claim the same outcome. A sophisticated allocation of credit still needs careful interpretation when the business asks how many sales would disappear if the advertising stopped.
MMM: a wider view that still needs challenging
Marketing mix modelling relates aggregate business outcomes to marketing activity and other relevant drivers over time, sometimes across geographies. It can support broader allocation questions, including how response may change at different investment levels. Depending on the design and available data, it can examine channels whose effects are difficult to follow through individual customer journeys. Its usefulness depends on whether the model represents the business well enough for the decision being made.
Spend and revenue alone don't make a model decision-ready. Price, promotions, seasonality, distribution, demand and other factors can move alongside advertising. Closely correlated channel activity can make individual effects difficult to separate. Google's Meridian documentation explicitly describes the causal assumptions required for MMM and notes that a central assumption can't be tested directly from the observed data.[1] A well-fitted model can therefore still be misleading if its causal structure or inputs are unsuitable. Examine uncertainty, stability and the spend ranges supported by the evidence before using a model to recommend a reallocation.
Incrementality: what changed because of the advertising?
An incrementality experiment compares outcomes under different advertising conditions to estimate what the activity changed. A well-designed randomised holdout, where feasible, can provide stronger causal evidence than a simple before-and-after comparison. Geography-based designs can also support causal evaluation, although the strength of the inference depends on the design and its assumptions. The result applies to a particular intervention, audience or market, period and level of investment.
Tests require practical discipline. Spillover between groups, differences in exposure, insufficient sample size and concurrent changes can weaken the conclusion. A small business may be unable to detect a commercially plausible effect within an acceptable time or budget. Even a convincing result doesn't automatically generalise to every future spend level or season. The useful output includes the estimated lift, uncertainty and conditions under which leadership can apply it.
Start with the question that keeps coming back
| Decision | Useful starting point | What to establish before acting |
|---|---|---|
| Investigate a sudden fall in recorded conversions | Attribution, tracking checks and business records | Whether the change is measurement, customer behaviour or both |
| Allocate a material budget across channels and markets | Suitable MMM alongside commercial economics and other evidence | Data suitability, causal assumptions, response uncertainty and actionable spend ranges |
| Assess whether a defined activity creates additional sales | A feasible incrementality experiment | A credible comparison, sufficient power and a relevant commercial outcome |
| Decide how much to spend tomorrow | Operational signals plus economic guardrails | Reporting delay, signal quality, margin, capacity and what remains causally uncertain |
These are starting points rather than a compulsory sequence. A suspected tracking fault rarely justifies commissioning an MMM first. A large allocation decision can warrant broader evidence than a single platform report. Check whether a proposed test can detect a meaningful effect before committing the budget. An inconclusive result after weeks of disruption is a painful way to discover the design couldn't answer the question.
When the answers clash, investigate the clash
Suppose an aggregate model recommends more paid social investment while platform attribution looks weak. The discrepancy could reflect under-recorded customer journeys, a model that has assigned unrelated demand to advertising, or differences in time horizon and outcome definition. Averaging the two answers might calm the meeting. It wouldn't explain the disagreement. Investigate the mechanism, review the customer economics and identify which additional evidence would materially change the proposed investment.
An experiment may help challenge or calibrate a model where the scope is relevant. Meridian's calibration guidance explains how experimental evidence can inform model priors, while recognising that the translation involves judgement.[2] A result from one period and intervention still needs to be interpreted before applying it elsewhere. Triangulation is useful when the methods address a shared decision and their differences are understood; collecting three sets of numbers on its own provides little assurance.
After all that work, someone should be able to make a better decision. I'd expect a recommendation with an action, supporting evidence, assumptions, expected commercial consequence, confidence, owner and review date. If stronger measurement is needed, explain what it will unlock. If a signal needs fixing, fix it. If the evidence is already sufficient for a reversible decision, give the team permission to move! Measurement loses credibility when the only recommendation it ever produces is more measurement.
If you've invested in measurement and still can't get to a decision, let's look at where the process is breaking down. GV Marketing helps growth leaders turn conflicting evidence into a recommendation their teams can act on. Where the underlying question needs untangling first, a paid diagnostic can establish the priority and next step.
Sources
[1] Google Meridian: About MMM as a causal inference methodology
Read next: how to allocate your marketing budget when the reports disagree