What is multi-touch attribution?

Multi-touch attribution allocates a conversion's credit across the several touchpoints a customer encountered rather than giving everything to the final one. Rules-based versions split it evenly, or weight the first and last touches, or decay by recency. Data-driven versions fit a model to converting and non-converting paths and infer weights from the patterns. Both are meaningful improvements on last-click, which is a low bar.

The improvement is genuine: upper-funnel touchpoints receive something rather than nothing, so brand and display channels stop looking worthless, and the reported picture is less distorted than a single-touch rule produces. Data-driven models can also capture real patterns in sequence and timing that no fixed rule would encode.

The limitation is structural rather than a matter of model quality, and it is the thing to understand. Attribution allocates credit among touchpoints it observed; it never estimates what would have happened without them. A customer touched by five channels who would have bought anyway generates five allocations of credit for a conversion nobody caused. No weighting scheme over that path can recover the counterfactual, because the information required is simply not in the data.

Three specific blind spots follow. It cannot see channels outside the click path — television, outdoor, radio, and view-through impressions that never produced a click. It cannot distinguish demand capture from demand creation, so a channel intercepting people who were already coming still scores well. And it depends on cross-device and cross-session identity resolution, which has degraded substantially with privacy changes, so paths are increasingly fragmentary and the model is fitting an incomplete picture.

The workable position is to treat it as a diagnostic rather than a decision rule. It is useful for understanding the sequences customers actually follow and for spotting channels that appear early in paths. Budget decisions should rest on incrementality established by experiment, or on a marketing mix model calibrated against experiments — with MTA describing the journey rather than pricing it.

An advertiser compares three attribution views of the same quarter — last-click, a data-driven MTA model, and geo experiments — across four channels.

MTA is clearly better than last-click here: branded search drops from 34% to 21% and display rises from 3% to 11%, both moves in the direction the experiments confirm. That is real value. But the television row is the structural point — a channel with a measured iROAS of 2.8 receives zero credit under both models, because it produces no clicks and therefore appears in no path. No improvement to the weighting scheme changes that. The last two rows explain why even the channels it can see are imprecisely valued: 61% of paths contain a single touchpoint, so for most conversions there is nothing to allocate across, and only 38% of cross-device journeys were stitched, meaning the multi-touch paths it does model are a biased subset. The sensible use is what the middle rows support — MTA as a corrective on last-click and a description of journeys, with the budget decisions taken from the geo column.