Last click attribution is the model everybody has agreed to be embarrassed about, and it is still what I would put on most accounts. It hands all the credit to the final interaction before the sale, which is obviously wrong, and its wrongness is at least predictable.
That is the whole argument in one line. A bias you can describe beats a bias nobody can explain, and most of the sophisticated alternatives fall into the second category.
This is for you if somebody has told you to move away from it and you want to know whether they are right.
What last click attribution actually does
It finds the last interaction before the conversion and gives it everything.
Somebody sees a Meta ad on Monday, ignores it. On Thursday they search your brand name, click a Google ad, and buy. Google gets 100% of that sale. Meta gets nothing, despite being the reason your brand name was in their head.
So the model consistently rewards capture over creation. The channel that closes gets paid, the channel that opened gets ignored.
Nobody defends that as an accurate picture of influence. The question is whether the alternatives describe influence any better, and mostly they do not.
Why last click attribution survives the criticism
Three reasons, and they are practical rather than theoretical.
It is unambiguous. Two people looking at the same data get the same answer, which sounds trivial and matters enormously when marketing and finance are in the same meeting.
It is cheap. No modelling, no vendor, no assumptions anyone has to defend in a review.
And its error runs one way. It undercounts demand creation and overcounts demand capture, always, by roughly the same mechanism. You can hold that in your head while reading any report.
Compare that with a model that redistributes credit using logic you cannot inspect. The output looks more sophisticated. Your ability to say what it means has gone down, not up.
Where last click attribution genuinely misleads you
The damage is real and it is specific, so it is worth naming precisely rather than generally.
It kills brand and awareness budgets. Those campaigns generate demand that surfaces later as branded search. Under this model the branded search takes the credit, so the awareness work reads as a cost with no return, and gets cut.
Then the branded search performance quietly degrades over the following months, because nothing is filling the top of the funnel any more. By then nobody connects the two, and the reports never showed a link.
That pattern is common enough to expect. If you run awareness activity and judge it on last click, you will eventually defund it, and the damage arrives a quarter later than the decision.
The comparison nobody makes fairly
Every model is an observational method, and observational methods have been tested against the real thing.
Gordon, Zettelmeyer, Bhargava and Chapsky compared observational attribution against randomised experiments across 15 Facebook experiments and 500 million user-experiment observations. Estimates were off by a factor of three in half the studies. One checkout study produced a 1,306% lift against a randomised benchmark of 2.4%.
Read that carefully, because it does not say last click attribution is uniquely bad. It says the whole family of methods struggles, including the sophisticated ones people upgrade to.
So the upgrade rarely buys accuracy. It buys a different set of assumptions, usually less visible, sometimes with a licence fee attached.
What to fix before changing the model
Here is the order I actually work in, and the model comes last.
Get the counting right. If a fifth of your conversions never register, no allocation rule can rescue that, and this is far more common than a bad model.
Then get outcome data flowing back, so the platforms optimise towards customers rather than form fills. That single change moves results more than every attribution decision combined.
Then ask whether a different rule would change a decision you are actually about to make. If the answer is no, and it usually is, leave it alone and go and do something that matters.
I have watched teams spend six weeks debating this while their tracking was under-counting. The allocation was argued to two decimals on top of an input that was wrong.
Correcting for the bias without changing anything
You can offset the known weakness without touching a setting, which is what I do on most accounts.
Track branded search volume as a separate line. When awareness activity works, branded search rises, and that movement is your evidence that something upstream is doing its job.
Add a “how did you hear about us” field to your enquiry form. It is unfashionable, self-reported, and it catches influence that no tracking system can see, including word of mouth and offline.
Then judge demand-creation activity on those measures rather than on its last click return. You are not fixing the model. You are declining to let it make a decision it is unqualified to make.
Where last click attribution quietly wins
There is a case for it that rarely gets made, because defending the unfashionable option is nobody’s idea of a good article.
It cannot be gamed internally. When credit is allocated by a rule everyone understands, nobody can argue their channel deserves a bigger share because of a modelling assumption. That sounds like a small thing until you have sat in a budget meeting where two teams are quoting two different allocations of the same revenue.
It also survives staff changes. A model somebody configured, tuned and then left behind becomes an unexplainable black box within a year. A rule you can state in one sentence does not.
And it degrades honestly. When your data gets patchier, and it will, last click attribution simply has less to work with. Sophisticated models keep producing confident-looking allocations from thinner and thinner inputs, and nothing in the output shows you that the ground shifted underneath it.
None of that makes it accurate. It makes it durable, which for most businesses most of the time is the more valuable property.
What to change this week
Three steps.
Find out which model each platform is applying right now, because defaults shift and most people have never checked.
Then, if you run any awareness activity, set up a branded search volume line before you next review budgets. Without it, that spend loses the argument automatically.
Then leave the model alone until you can name a specific decision that would change, and write that decision down when you find one.
The six ways to divide credit covers the alternatives in detail, platform reported conversions explains why each channel claims what it claims, and what conversion numbers actually measure is the counting problem that sits underneath all of it. On Meta specifically the windows do a lot of this work invisibly, which my Meta Ads work treats as part of the setup. If your channels are fighting over the same sales, book a teardown.