What is last-click attribution?

Last-click attribution assigns 100% of a conversion's credit to the last touchpoint before it — usually the final click, often with direct traffic excluded so the last identifiable marketing touch gets it. It is the default in most analytics platforms, and it is the number the majority of marketing decisions are still made on.

Its virtues are real and worth stating before the criticism. It is unambiguous: every conversion has exactly one final touch, so there is nothing to argue about. It is reproducible, so two analysts get the same answer. And it is universally understood, which matters more in practice than it should — a model everyone reads the same way has coordination value that a better but contested model lacks.

The problem is that it is wrong in one direction rather than randomly. Whatever sits closest to the purchase collects everything: branded search, retargeting, email to existing customers. Whatever sits further up — brand advertising, content, display, anything that created the demand in the first place — collects nothing. A customer who saw a television campaign, searched the brand a week later and clicked a paid link gives all the credit to that paid link, which was capturing demand the campaign created.

The consequence is predictable and expensive. Upper-funnel channels look unprofitable and get cut; lower-funnel channels look excellent and get scaled; six months later the lower-funnel channels perform worse because there is less demand flowing into them, and nothing in the attribution data explains why. The failure loop is slow enough that the cause is rarely connected to the decision.

The alternatives improve on it in different ways and none fixes it entirely. Multi-touch attribution spreads credit across the observed path, which is better and still confined to touchpoints that were recorded. Marketing mix modelling sees all channels including offline. And experiments — geo tests, conversion lift studies — measure incrementality directly rather than allocating credit at all. The honest use of last-click is as an operational monitoring metric, not as the basis for budget decisions.

A retailer reviews a quarter of last-click reporting against geo experiments run on the same channels, and looks at what the attribution-led decisions of the previous year had produced.

The final two rows are the loop closing. Brand video was cut because it produced almost no last-click conversions, which was true and irrelevant — its contribution arrives as people later searching for the brand, and last-click hands that to search. A year on, branded search conversions are down 22%, and the attribution data offers no explanation because the cause is a channel it never valued. This is the characteristic failure mode: the decision looks well-evidenced at the time, and the consequence appears somewhere else, later, with no visible link. Note that the branded search figure is doing two suspect things at once — capturing demand created elsewhere, and capturing demand that needed no advertising at all, since a 0.9 iROAS means most of those customers would have arrived organically. Both are invisible to a model that only asks who was touched last.