What is Last-click attribution?
Last-click attribution assigns 100% of a conversion's credit to the last traffic source before the purchase. It is the default in most analytics tools because it is simple and needs the least memory of a visitor's history. It reliably over-credits whatever people use to come back — direct traffic, branded search, email.
Why it is the default
Last-click only needs to know where the visitor came from on the visit that converted. No long-lived identifier, no stitching of sessions across weeks, no decision about how to divide credit. It is the cheapest model to compute correctly and the hardest to get wrong mechanically — which is a different thing from being right.
The systematic bias
The channels people use to return to a site they already know take credit for the work done by the channels that introduced them.
A customer reads a comparison article, follows a link, leaves. Two weeks later they type your domain in directly and buy. Last-click reports: direct traffic, one sale. The article that actually did the persuading gets nothing, and the report quietly recommends you invest in “direct” — a channel you cannot buy.
Branded search has the same problem in a more expensive form. People who already know your name search for it, click the ad, and convert. The ad takes the credit for demand it did not create.
Reading it honestly
- Treat direct and branded traffic under last-click as evidence of demand created elsewhere, not as a channel to scale.
- Compare it against first-click on the same conversions. A channel that looks strong in first-click and absent in last-click is a discovery channel, not a failure.
- Watch the trend rather than the level. Last-click is biased consistently, so its movement over time still carries signal even when its absolute ranking does not.
When it is genuinely the right model
Short consideration cycles where the visit that converts is very often the only visit — impulse purchases, small one-off transactions, direct-response campaigns judged on the same day. When there is no journey, there is nothing for the model to mis-assign.