What is Last-click attribution?

Last-click attribution credits a conversion entirely to the final ad interaction before it. Everything earlier gets nothing.

A person sees a video ad, searches the brand a week later, clicks a search ad and buys. Last-click gives the whole conversion to the search ad. The video that created the search gets zero, and in most reports does not appear at all.

The rule is simple enough to state in a sentence, which is a large part of why it persisted as the industry default long after its weaknesses were well understood.

It is worth being fair to the model, because its properties are real and several of the alternatives lack them.

Deterministic. The same data produces the same answer every time. There is no model to drift, no version to change underneath a report.

Auditable. A last-click figure can be recomputed from a log by somebody who wants to check it. A modelled figure cannot.

Comparable. Two systems applying the rule to the same journeys agree, which is not true of proprietary models.

Intelligible. Everybody in a business can understand what it means, which matters more in practice than it should.

In an era when most journeys were short, observable and largely on one device, those properties mattered and the bias was tolerable. Both halves of that sentence have since changed.

The bias is not random, and knowing its direction is most of what makes the numbers usable.

Lower-funnel activity wins, always. Branded search, retargeting, and any channel a person passes through on the way to a decision already made. These are present at the end by construction.

Demand creation loses, always. The video, the display campaign, the sponsorship — earliest in the sequence, frequently outside the window, often producing no click at all.

Intermediaries are overpaid. Comparison sites, coupon extensions and anything else that inserts itself between a decided buyer and a checkout collect the credit for a sale they did not cause.

An account allocating strictly on last-click therefore drifts toward the bottom of the funnel. The drift is gradual and each step is justified by the report, which is what makes it hard to argue against until the effect becomes visible in the top line.

The pattern is consistent enough to describe in advance.

Budget moves from upper-funnel activity, which reports poorly, toward branded search and retargeting, which report superbly. The account's blended reported efficiency improves. Somebody is congratulated.

Six or twelve months later, the volume of people searching the brand stops growing, because nothing is creating new ones. Retargeting pools shrink, because fewer new visitors are arriving to fill them. The campaigns that report best now have less to work with, and their efficiency starts falling for reasons that appear to be about the auction.

The activity that would have fixed it was cut earlier, on the strength of a model that could never have seen its contribution.

As a reconciliation view. Because it is deterministic, it can be checked. Running it alongside a modelled figure and watching the gap is a genuine control: a widening gap means the model is crediting earlier touchpoints more than it did, which is a measurement change rather than a performance one.

For short, simple journeys. Where a purchase genuinely involves one interaction — an impulse product, a direct-response offer with no consideration — the model's bias has little to distort.

As a conservative floor. A channel that pays for itself on last-click is paying for itself on the least generous basis available, which is a useful thing to know about a channel somebody is arguing to cut.

A data-driven model distributes credit across the sequence and is closer to how buying works, at the cost of being unauditable. That trade is usually worth taking for allocation.

The blended ratio — total revenue over total spend — is immune to every attribution argument and cannot allocate at all, which makes it the right control rather than a replacement.

And for the questions where the answer decides real money, a controlled test that withholds advertising from a comparable group is the only instrument that measures cause rather than distributing credit.

The working arrangement: allocate on a data-driven view, reconcile against last-click, control with the blended figure, and test the expensive questions. Each covers a weakness of the others, and an account running only the first of the four is the one this page is a warning about.

Analytics tools frequently apply a variant rather than the plain rule, and the difference matters when reconciling figures.

Last non-direct click skips direct traffic and credits the last identifiable source. The reasoning is that somebody typing a URL was sent there by something, and crediting "direct" explains nothing. It shifts credit back toward whatever campaign preceded the direct visit.

Last paid click ignores organic and other unpaid touchpoints, which is closer to what an ad platform reports and further from what the business experienced.

Last click within channel applies the rule inside a channel's own reporting, which is why a platform's figure and an analytics figure for the same campaign differ even when both describe themselves as last-click.

The practical point is that "last-click" names a family rather than one measurement, and two systems both claiming to use it will disagree. Establishing which variant each report uses is the first step in any reconciliation, and it resolves a surprising share of the arguments.

Inherited accounts carry the model's fingerprints and they are recognisable.

Budget concentrated in branded search and retargeting. Upper-funnel campaigns either absent or repeatedly cut and restarted. A history of tests that were stopped early because they reported poorly. Excellent blended cost per acquisition and flat new-customer growth.

That last pairing is the signature. An account acquiring customers at a good price and not acquiring more of them has usually optimised itself into harvesting, and the report that justified each step will show nothing wrong at all.

Last-click gives every conversion to the final ad interaction before it; a data-driven model distributes the credit across the touchpoints using patterns learned from journeys that converted and journeys that did not. The case for the rule is everything the model gives up. It is deterministic, so the same data produces the same answer twice. It is auditable, so anybody who wants to check a figure can recompute it from a log. It is comparable, so two systems applying it agree. And it is intelligible to people outside the advertising team, which matters more in practice than it should. The price is a bias that never varies: everything goes to whatever was present at the moment of decision, so branded search and retargeting are rewarded and the advertising that created the demand they harvest is rewarded with nothing. An account allocating strictly on this rule drifts toward capture and away from creation, defensibly, one justified decision at a time. Its enduring use is as a reconciliation view: a figure that can be checked, sitting beside one that cannot.