<b>What is every attribution model secretly assuming when it hands a channel credit?</b>

<b>What is every attribution model secretly assuming when it hands a channel credit?</b>

This is the foundational question under all the others, and answering it honestly reorganizes how you read every report you own.

<b>The hidden assumption</b>

Every rule-based and most data-driven attribution models assume that <i>presence in the path implies contribution to the outcome</i>. Last-click, time-decay, U-shaped, Shapley, Markov — all of them divide credit among touchpoints that <i>co-occurred</i> with conversions. None of them, on their own, establish that the touchpoint <i>changed</i> whether the conversion happened.

<b>Why this matters</b>

The counterfactual — would this person have converted anyway? — is invisible to path data. A channel that only ever appears in front of already-decided buyers will accumulate enormous attribution credit while causing nothing. This isn't a flaw in one model; it's the boundary condition of the entire observational approach. Correlation is precisely what attribution measures, and causation is precisely what it cannot.

<b>The nuance</b>

This is not a counsel of despair. Attribution is still useful as a <i>fast, granular allocation heuristic</i> and as a hypothesis generator. The error is epistemic: treating a credit-splitting rule as a causal verdict. Only experiments and well-identified quasi-experiments cross the line from "was present" to "made a difference."

<b>What to actually do</b>

— Read every attribution number as a claim about correlation, then ask which experiment would confirm it causally.
— Reserve the word "caused" for results backed by a counterfactual — a holdout, a geo test, an RD or DiD design.
— Build a cadence where attribution proposes and incrementality disposes.

Bottom line for practitioners: attribution answers "who was there?" not "who made the difference?" Hold the two questions apart, and never let a credit-allocation rule quietly promote itself to a statement about cause.