Open a marketing dashboard and you will find a number next to every channel. Paid search: 31%. Organic: 24%. Email: 18%. Social: 14%. Direct: 13%. The percentages sum cleanly to 100. They are rendered with the same precision, in the same row, at the same visual weight.

Nothing about the number itself indicates that it may have been produced by a fundamentally different act of measurement than the number beside it, or that establishing real confidence in it would require a different kind of test than establishing confidence in another.

This is not a complaint about any specific tool. It is a structural observation about what digital measurement systems are built to do, and what a given output can and cannot, by itself, tell the person reading it.

Two different questions, one shared form of display

Attribution answers a specific question: given a defined set of rules, how should credit for this conversion be distributed across the touchpoints we were able to observe?

That is a real question, and attribution answers it precisely. Last-click assigns the full amount to the final interaction. Data-driven models use more of the observed conversion path, weighting touchpoints according to patterns detected across historical data rather than crediting a single final step. Position-based, linear, and time-decay models distribute credit according to their own logic. Each produces a defensible, internally consistent number, according to the rule it was built to apply.

But none of these methods asks – or can answer – a second question: would this conversion have happened anyway, without this touchpoint? That is a causal question. A controlled comparison can answer it by withholding exposure from a comparable group and measuring the difference. This is what incrementality testing does. It does not distribute a fixed pool of credit across observed events. It measures a difference between what happened and what would have happened without the intervention.

These are not two ways of measuring the same thing. They are two different instruments pointed at two different questions. One measures assignment. The other measures difference.

The structural condition worth naming is not that attribution is wrong. It is that digital reporting can render an attribution output with a degree of numerical precision that does not itself reveal the evidentiary basis behind the number – whether it was assigned by rule, or established by comparison.

Why this is a structural condition, not a competence gap

The condition has a structural basis. Attribution systems are generally built into existing digital reporting infrastructure and are more readily available on an ongoing basis, drawing on data the organization already collects as part of normal operation. Causal methods typically require something additional: a defined test design, a holdout or control condition, sufficient volume to detect a real effect, and time to run. Where randomization is difficult – long B2B cycles, brand campaigns, low-volume channels, complex channel interaction – a clean causal test may be hard to construct at all.

This produces an asymmetry, though not a universal or absolute one: the more readily available number tends to be present continuously, across whichever channels the organization already tracks. The harder-to-establish number tends to be available more selectively, where the conditions for a valid test exist and where the effort to run one has been made.

Nothing in the reporting layer is obligated to represent this asymmetry. A dashboard cell showing an attributed percentage and a dashboard cell showing a confirmed lift result can appear in the same table, with the same formatting, at the same apparent confidence. The distinction between an assigned number and a proven one is not a property visible in the display. It is a property of how the number was produced – and that history does not travel with the number onto the screen.

This is not an argument against attribution

It would be easy to overcorrect here, and the evidence does not support overcorrecting. Attribution is not obsolete. Data-driven attribution represents a more sophisticated method than simple last-click assignment – it draws on more of the observed path rather than crediting a single final step. It remains available on an ongoing basis in a way that few causal methods can match, and most day-to-day decisions – pacing, channel-mix monitoring, campaign-level adjustment – do not require causal proof at all. Attribution is a reasonable operational tool for that layer of decision-making.

Incrementality testing is not a universal substitute, either. It cannot easily be run on every channel at once. It struggles with low-volume channels, long consideration cycles, and situations where a clean control group cannot be constructed. It answers a narrower question, well, under a bounded set of conditions – not a broader question, everywhere.

Marketing Mix Modeling occupies a third position: an aggregate method that can estimate portfolio-level contribution, but one that measures correlation unless it is calibrated against causal experiments. It complements the other two methods rather than replacing either.

None of these three methods is sufficient alone. That is the point. The organization does not necessarily have a broken measurement system. It may simply have three different instruments, suited to three different kinds of questions, none of which announces, on the screen, which question it was actually built to answer.

The actual executive condition

The condition worth naming for leadership is narrower than “your attribution is inaccurate” and narrower than “you need to run more experiments.” It is this:

Before a number drives a consequential decision, does leadership know whether that number was assigned by a rule, or established by a comparison?

That question may never be asked unless the reporting process requires someone to distinguish the basis of the number. The number arrives, correctly formatted, in the right column, next to the right channel name. It looks like every other number on the page.

This is not primarily a data-quality problem, and it is not a governance failure in the way that term is usually used. It is a structural feature of how digital measurement systems present their output: the interface is built to summarize a result, not to annotate the evidentiary act that produced it. Whether that annotation happens at all depends on whether someone thought to ask for it.

What changes when the distinction is visible

Nothing about the underlying measurement needs to change for this condition to matter. The attribution model does not need to be replaced. Any existing incrementality program does not need to expand. What changes is simpler and prior to any of that: whether the organization can look at a number about to inform a consequential decision and correctly identify what kind of number it is.

A dashboard percentage that has never been tested against a control group is not less useful for being an assignment rather than a proof – but a decision that treats it as proof, when it is only an assignment, is making a category error that the number itself gave no indication of.

The distinction is not a call to distrust measurement. It is a call to know, before relying on a specific number for a specific decision, which of the two questions that number was actually built to answer.