Somebody sees a Meta ad, ignores it, searches your name a week later, clicks a Google ad, leaves, comes back through an email, and buys. One sale. Four interactions. Attribution models are the rules that decide who gets the credit, and every one of them is an opinion wearing a percentage.
The useful framing is that you are not choosing a true answer. You are choosing which distortion you would rather live with, then applying it consistently.
This is for you if two channels are both claiming the same sale and somebody has asked you to explain it.
What attribution models are actually doing
They divide one outcome between several causes, using a rule you picked in advance.
That is worth sitting with, because it explains why the debate never resolves. The sale happened once. The question of how much of it belongs to the Meta impression is not a measurement question at all, it is a judgement about influence that nobody can observe directly.
So a model does not discover the answer. It imposes one, consistently, so that your numbers are comparable week to week.
Consistency is the actual benefit, and it is a real one. A rule applied the same way every month gives you a trend you can trust, even when the level is wrong.
The six attribution models, and what each one flatters
Every platform’s list is a variation on these.
Last click. All credit to the final interaction before the sale. Flatters search, brand terms and retargeting. Undervalues everything that created the demand.
First click. All credit to the first interaction. Flatters awareness and top-of-funnel. Ignores everything that closed the sale.
Linear. Credit split evenly across every touch. Fair-looking, and it treats a scrolled-past impression as equal to the click that converted.
Time decay. More credit the closer to the sale. Reasonable for short cycles, punishing for long ones where the early work does most of the persuading.
Position based. Usually 40% first, 40% last, 20% spread across the middle. A compromise that suits businesses where discovery and closing are both genuinely hard.
Data driven. The platform’s own model, built from patterns in converting and non-converting paths. Better than a fixed rule, and limited to what that platform can see.
That last limitation is the one people miss. A data driven model inside Google redistributes credit among Google interactions. It is not adjudicating between Google and Meta, because it cannot see Meta at all.
Why attribution models matter less than the argument about them
Here is the part that saves you a lot of meetings.
Changing attribution models changes your reported numbers. It changes nothing about how many people bought. If switching model makes paid social look 20% better, paid social did not improve by 20%. The rule for dividing credit changed.
I have watched teams spend weeks on this while their conversion tracking was under-counting a fifth of their enquiries. The allocation was being argued over to two decimal places while the input was wrong.
Get the counting right first. Attribution models are how you slice the pie, and slicing is the wrong thing to optimise when you are not sure how big the pie is.
What no attribution models can do
None of them can tell you what would have happened anyway.
Somebody who was always going to buy, who saw your ad, clicked it and bought, is recorded as a conversion you caused under every one of the six. The models differ on which advert gets the credit. They agree entirely that the advert deserves credit, which is the assumption that actually needs testing.
The scale of that assumption is documented. Gordon, Zettelmeyer, Bhargava and Chapsky compared observational attribution against randomised experiments across 15 Facebook experiments and 500 million user-experiment observations, and found estimates off by a factor of three in half the studies. In one checkout study the observational method reported a 1,306% lift against a randomised benchmark of 2.4%.
Every attribution model in common use is an observational method. That finding applies to all six.
The trap when you switch between attribution models
Switching looks free. It is not, and the cost lands in a place nobody checks.
Most platforms restate history when you change the rule. Your old months get recalculated under the new logic, so the chart stays smooth and nothing announces that the basis moved. That sounds helpful and it quietly destroys your ability to say what happened.
The report you sent the board in March described the world under one rule. Pull the same period today, after a switch, and the numbers are different. Neither version is a lie. They answer different questions, and only one of them was in the room in March.
So if you do change, export the old figures first and keep them somewhere the platform cannot rewrite. A spreadsheet is fine. It takes ten minutes and it is the only way to preserve what you actually told people.
The second half of the trap is timing. A switch made mid-quarter guarantees an argument, because performance will appear to move for a reason that has nothing to do with performance. Change at a period boundary, tell everyone before rather than after, and expect the first month to be unreadable.
That last point is worth saying plainly. Attribution models are a reporting decision with political consequences inside a business, and treating them as a purely technical setting is how marketing teams end up defending a change nobody understood.
How to choose between attribution models and stop thinking about it
Three questions, answered once.
How long is your sales cycle? Short cycles suit last click or time decay. Long ones need something that credits the early work, or you will defund the thing that generates demand.
How many channels genuinely contribute? If you run one channel, this whole topic is nearly irrelevant to you and last click is fine.
Who reads the report? If it goes to finance, comparability matters more than sophistication. A simple model held steady beats a clever model you changed in June.
Then write down which you chose and why, and leave it. The written note is the part people skip, and it is what stops somebody re-litigating the decision every quarter.
What to change this week
Three steps.
Find out which model each of your platforms is currently using. Many people have never checked, and defaults change without announcement.
Then confirm nobody is adding two platforms’ conversions together in a report that reaches management. That is the error that actually costs credibility.
Then leave the model alone until you have a specific reason, and write the reason down when you do.
What conversion numbers actually measure covers the counting that has to be right before any of this matters, platform reported conversions explains why each platform claims what it claims, and why your reported cost per lead is wrong is the same argument applied to lead generation. What I do across paid, SEO and the website treats these as one measurement problem rather than three. If your channels are arguing over the same sales, book a teardown.