What is attribution modeling?

The common models

• Last click. All credit to the final interaction. Simple, auditable, and structurally biased toward the channels that harvest demand rather than create it.
• First click. All credit to the first. The mirror bias — flatters discovery, ignores what closed.
• Linear. Equal credit to every touchpoint. Assumes a nine-touch path and a one-touch path are equally informative, which is rarely true.
• Time decay. More credit to touchpoints nearer the conversion. Reasonable in short cycles, wrong in long ones where the early influence is the real one.
• Position-based. Weighted to first and last, remainder split. A compromise, not a finding.
• Data-driven. Weights learned from converting and non-converting paths. The most defensible, and only as good as the data it sees — which now excludes most cross-device and all unclicked exposure.

Why the choice changes the answer

Run the same log through last click and through a data-driven model and the channel ranking will differ, often sharply. Upper-funnel channels that introduce a brand look worthless under last click and material under almost anything else. Retargeting looks superb under last click and often near-zero under incrementality.

This is the single most consequential fact about attribution modelling, and it is routinely forgotten in budget meetings: the model is an input to the decision, not a neutral readout of what happened.

Where every model fails together

All of them share one assumption — that the touchpoints in the log are the touchpoints that mattered. Any influence that leaves no record is invisible to every model equally.

That category has grown. Unclicked exposure, cross-device journeys, word of mouth, and now recommendations inside AI assistants that the user acts on without clicking. No amount of model sophistication recovers an event that was never logged, which is why the honest use of attribution modelling is for allocating between things you can see, with a separate causal instrument for the question of whether the total is working.

What to actually do

Pick one model, document it, and keep it stable — most of the value is in consistency, because trend is readable even when the level is biased. Then run an incrementality test on any channel large enough to matter, at least annually. Where the two disagree, the holdout is right and the model is telling you about its own assumptions.