Last-click did not come from anyone asking how advertising works. It came from ad servers needing to assign each conversion to exactly one touch so the ledger balanced. Somebody had to be credited, the rule had to be simple, and the most recent click was the easiest thing to defend in a room.
That is an accounting decision. It was a reasonable one. The trouble began when the output started being read as a finding.
Credit and cause are different claims
Credit is an assignment. Given a set of touches and a purchase, a rule decides who gets the entry. Cause is a counterfactual: it asks what would have happened had the touch not occurred.
Last-click contains no counterfactual anywhere in it. It observes ordering and reports the touch nearest the purchase. That is a true statement about sequence and it says nothing about influence. A rule that always credits the last thing before the outcome would also conclude that walking through the door causes people to buy groceries.
The bias runs in one direction
If it were merely imprecise, it would scatter error in every direction and roughly average out. It does not. It rewards proximity, and proximity is not distributed evenly across your media.
The channels that sit closest to a purchase are the ones that intercept people who had already decided. Branded search, retargeting, coupon and affiliate placements, email to an existing customer. Each of them is positioned where the decision has already been made, and each collects the credit for it.
The channels that created the decision sit further back. Video, connected TV, social prospecting, out of home, anything working on people who did not know you existed last month. Their contribution often lands outside the attribution window entirely, and much of it never produces a click at all.
So the rule does not simply fail to see demand creation. It takes the credit that belongs to demand creation and hands it to whatever intercepted the demand on its way to converting.
Branded search is the cleanest example. Someone types your company name into a search box and an ad appears above the organic listing they were going to click anyway. Last-click awards that conversion in full. The counterfactual is easy to picture, which is why brand terms are so often the first thing anyone tests.
It funds the thing it can see
The consequence compounds, because budget follows reported performance.
Money moves toward the channels showing the strongest last-click return, which are the harvest channels. More harvest capacity gets pointed at the same pool of existing intent. The channels that were refilling that pool lose funding, because they cannot demonstrate a return under the rule being used to judge them.
For a while the reported numbers improve, since a larger share of a shrinking pool still looks efficient. The pattern to watch for is efficiency metrics climbing while total revenue flattens. That combination is not a paradox. It is what harvesting looks like when nothing is planting.
Fractional models change the arithmetic, not the epistemics
The usual response is to move to a multi-touch model that distributes credit across the path rather than dumping it on the final click. It is a better bookkeeping rule and it does soften the proximity bias.
It does not change what kind of claim is being made. Every touch-based model divides observed credit among observed touches according to a rule somebody chose. None of them observes the counterfactual, because the counterfactual is not in the data. Redistributing credit more thoughtfully produces a more sophisticated allocation, and an allocation is still not evidence.
The practical position has also weakened. These models depend on stitching a person's touches together across sites, apps and devices, and that stitching has become steadily less complete. Fewer of the touches are visible, so a rule that was already making an assumption is now making it with a partial view.
What it is good for
Last-click is a fast, cheap, consistent operational signal. For pacing spend, rotating creative, catching a broken landing page, or spotting a campaign that stopped delivering, it is genuinely useful and nothing about the above argues otherwise.
The failure is one of scope. It is being used to answer a question it was never built for, in the room where budgets get set, in front of someone who is going to ask whether the revenue would have arrived anyway.
That question is a counterfactual, and no allocation rule contains one.
What replaces it
Not a better attribution model. The reasonable next question is what to use instead, and the honest answer is that the replacement is not another rule for dividing credit. It is a different kind of evidence, produced by three things doing separate jobs.
Experiments produce the counterfactual. They are the only place one can come from, because they create a group that did not receive the media and then observe what that group did. They are narrow, periodic and expensive, and within their scope they settle the question. Geo designs are the usual instrument, and there are others where geography is not available.
A model carries the answer between them. Experiments cannot run continuously on every channel, and they say nothing about the parts of the business that never touch an ad server: organic demand, brand, offline, pricing, seasonality, what competitors did. A media mix model covers all of it and runs all the time. Its credibility comes from being calibrated against the experiments rather than from how well it fits history, which is a distinction worth insisting on.
Attribution stays, demoted. It remains the fastest signal you have for whether a campaign is delivering today. It is operational telemetry, and the change is not that you stop looking at it. The change is that it stops being the thing budget decisions are argued from.
The common error is treating touch-based attribution and mix modeling as rival answers to one question. They are not answering the same question. One allocates credit among observed touches, one estimates contribution across the whole business, and only the experiment observes a world where the spend did not happen. The reason this arrives as a system rather than as a piece of software is that no single tool can hold all three. The reporting disagreements that started this are a symptom of expecting one to.