Your Attribution Reports Are Probably Wrong

The four levels of marketing measurement, and what they do and do not tell you.

The proliferation of AI in marketing has given teams the ability to shift more attention away from copy and content creation and toward data, reporting, and optimization.

The problem many teams now find themselves facing, however, is a technical, data-driven talent gap.

The secondary, and often compounding, problem is the rise and adoption of out-of-the-box CRM and platform reporting.

The talent gap is fairly apparent. Many marketers understand basic KPIs like CTR, ROAS, and CAC. Far fewer understand the data actually driving those metrics, and fewer still understand the structural methodologies required to make sure that data is being interpreted accurately.

Let's break this down through a simple maturity model for marketing measurement.

The Four Levels of Marketing Measurement

Level 1: Platform Metrics

These are the basic, face-value metrics: clicks, opens, likes, shares, CTR, engagement, and platform-reported ROAS.

This is the bread and butter of marketing reporting and, historically, the directional data most teams have based decisions on.

These metrics are useful. They can tell you whether one ad received more engagement than another, whether an email subject line generated more opens, or whether a campaign produced activity inside a particular platform.

They can't reliably tell you whether that activity created incremental business value. A platform attributing a conversion to an interaction is not the same as proving the conversion would not have happened without it. That distinction becomes more important as you move up the maturity model.

Level 2: Multi-Touch Attribution

Level 2 is multi-touch attribution, or MTA. Instead of one platform reporting on its own performance, you are trying to track a single customer across channels and assign credit to each touchpoint along the way.

Attribution remains one of the most widely used and frequently consulted forms of marketing measurement. It's also where the ground is currently shifting underneath teams.

MTA depends on being able to identify and track one person across devices, platforms, channels, and time. That tracking layer has become increasingly incomplete, fragmented, and dependent on modeled recovery. Apple's App Tracking Transparency, cookie restrictions, consent requirements, device fragmentation, and closed advertising ecosystems have removed a meaningful portion of the deterministic tracking MTA depends on.

Let's be clear: this isn't eventually. It's happening now.

And that's not to say attribution is useless. It can still help teams understand customer journeys, optimize channels, and make directional decisions, especially when the company has strong first-party data and authenticated user activity.

But the crux here is that the level many teams rely on most heavily is also the most structurally exposed. Attribution can tell you which touchpoints appeared along the path. It can't reliably tell you which touchpoints changed the outcome.

Level 3: Incrementality Testing

Level 3 is incrementality testing. This is where you begin differentiating between causation and correlation.

You turn a channel off in one market and leave it running in another, then compare the results. You hold back a portion of an audience from a campaign and measure what happens without the exposure. You create treatment and control groups to determine whether marketing actually produced an incremental result.

When treatment and control groups are properly constructed and randomized, this is not just close to an experiment. It is one. More importantly, it gives you the cleanest direct evidence on this list that a channel caused a result instead of simply appearing somewhere near one.

The tradeoff is cost, speed, and practicality. You cannot run a perfect experiment across every channel, campaign, and audience at all times. Some channels are difficult to isolate. Some tests require enough volume to generate a meaningful answer. Others mean intentionally withholding marketing from people you would otherwise want to reach.

So you do not use incrementality testing to answer every question. You use it to answer the biggest, most expensive, or most uncertain ones. Then you use those answers to check your work elsewhere.

Level 4: Marketing Mix Modeling

Level 4 is Marketing Mix Modeling, or MMM. MMM uses aggregate data (spend, sales, seasonality, pricing, promotions, economic conditions, and other external factors) to estimate, under a defined set of assumptions, how much each channel contributed to a business outcome. It does this without tracking a single individual across their journey.

Because MMM does not depend on cookies, device IDs, or personally identifiable user paths, it is far less exposed to the privacy changes currently weakening Level 2.

Historically, MMM required a specialized data science team, significant infrastructure, and often a six-figure vendor contract. That has changed. Google's Meridian has joined Meta's established Robyn project, giving teams access to sophisticated open-source MMM frameworks that were previously far less accessible.

The software is increasingly free. Building something trustworthy with it is not.

Two concepts are worth understanding plainly before trusting this level.

The first is adstock. Adstock is how long a channel's effect continues after the initial spend. Paid search may produce a fairly immediate response and fade quickly. Brand advertising, sponsorships, or television may influence demand over a much longer period.

The second is saturation. Saturation is the point where additional spend stops producing proportional additional results. Spending twice as much does not mean you will generate twice the outcome. Eventually, the most responsive audience has already been reached, frequency increases, and marginal performance begins to decline.

Model every channel using the same assumptions, and the output can confidently mislead you.

Before anyone should trust a model with real budget, a few basic health checks are required. MAPE, or mean absolute percentage error, helps show how far the model's predictions were from observed outcomes. R² shows how much of the observed variation the model fits. Where the available data allows it, a holdout or forward-validation test checks whether the model performs against a period or subset it did not use during training.

Those checks matter. But they test predictive fit, not whether the model correctly identified what caused the outcome. Causal confidence also depends on model convergence, uncertainty, sensitivity to assumptions, the treatment of confounding variables, business plausibility, and, where possible, calibration against incrementality experiments.

Skip that work and you do not have a trustworthy model. You have a guess with better formatting.

The Real Measurement Gap Isn't Math. It's Coordination.

We talk about these measurement frameworks as if they are purely data science problems. They aren't. They are coordination problems.

You can download the most sophisticated open-source code or buy the flashiest analytics suite on the market. But if Product, Sales, Finance, Data, and Marketing are working from different definitions of the customer journey, conversion, revenue, and success, no model can resolve that disagreement for you.

Moving from Level 2 to Level 4 is not a matter of installing another dashboard. It means reconciling channel spend with actual business outcomes. It means defining the decisions the model needs to support, choosing the right control variables, testing adstock and saturation assumptions, validating the output, and then building a repeatable decision process around it.

It also means getting the organization to agree on which numbers will be trusted when the systems inevitably disagree. That last part is often harder than the model itself.

Common assumptions, clear ownership, defined decision rules, and an agreed process for what happens next. The model is one component. The operating system around it is where the value is created.

Where This Leaves You

The mistake would be assuming Level 4 replaces everything below it. It does not.

The better answer is triangulation. MMM sets the strategic budget view. Incrementality testing calibrates and validates it. Attribution still helps inform weekly, in-channel decisions. Platform metrics still help practitioners optimize individual campaigns, ads, emails, and experiences.

Each level answers a different question. The problem begins when a business asks one level to prove something it was never designed to prove.

Many teams still rely heavily on Level 2 reporting for budget decisions, even as the infrastructure supporting it has become less complete.

Getting to Level 4 is not really a tooling problem anymore. It is a data infrastructure, talent, and organizational trust problem.

Are your spend and revenue data clean enough to model? Are your teams aligned on the business outcome that actually matters? Can your organization distinguish predictive fit from causal proof? And are you willing to act on a model when it contradicts the dashboard everyone already understands?

Before investing in another reporting tool, marketing leaders should ask three questions:

1. Which budget decisions are currently being made using attribution data?
2. Which of those conclusions have been validated through an experiment or an independent model?
3. Where would a wrong answer cost the business the most money?

Those answers will tell you more about your measurement maturity than the number of dashboards you have.

One More Thing Before You Assume Level 4 Is the Ceiling

It probably won't be for much longer.

A growing number of vendors are beginning to introduce agent-assisted MMM: systems that can accelerate data preparation, maintain models, run scenarios, and generate optimization recommendations.

The credible version of this is not a fully autonomous AI agent freely moving millions of dollars across channels. It is governed autonomy. Spend caps. Approval gates. Defined decision rules. Transparent assumptions. An audit trail. A human still holding the leash.

Over time, these systems will likely connect measurement more directly to execution, allowing approved budget changes to happen faster as new evidence enters the model.

It is still emerging, and some of the current language is ahead of the operational reality. But the direction is fairly clear. Marketing measurement is moving from reporting what happened, to estimating what caused it, to recommending what should happen next. Eventually, it will begin acting on those recommendations within defined limits.

More on that soon.