Attribution Models
By the end of this lesson, you should be able to: explain what a rule-based attribution model actually is, say why the choice between first-touch and last-touch matters less than everyone thinks, apply the Shapley value from lesson 11 to credit allocation, and name the one thing that separates any of this from a causal claim.
An experiment you can never run at work
Attribution asks: a user saw a paid search ad, then a referral link, then converted. Who gets the credit?
The reason this argument never ends is that nobody knows the answer. You can't observe the world where that user didn't see the ad.
So here we plant one. Sixty thousand synthetic paths, where each channel has a known real effect on conversion, set in advance:
| Channel | True lift per touch | Appears on |
|---|---|---|
| Organic | 0.00 | 55% of paths |
| Paid search | 0.10 | 30% |
| Paid social | 0.01 | 62% |
| Referral | 0.16 | 16% |
| Partnership | 0.22 | 7% |
Two channels are built to be traps. Paid social is on more paths than anything else and moves almost nobody. Partnership is rare and enormously effective. Organic is pure discovery: people who were going to convert pass through it on the way.
Now run the standard models and see which recovers the truth.
Five models, one answer, and it is wrong
| Channel | First | Last | Linear | Decay | Position | Shapley | TRUTH |
|---|---|---|---|---|---|---|---|
| Organic | 27.3% | 27.7% | 27.5% | 27.6% | 27.5% | 3.1% | 0.0% |
| Paid search | 19.9% | 20.4% | 20.1% | 20.2% | 20.1% | 31.8% | 38.8% |
| Paid social | 32.9% | 32.2% | 32.6% | 32.4% | 32.5% | 3.7% | 8.0% |
| Referral | 13.0% | 12.7% | 12.9% | 12.9% | 13.0% | 34.7% | 33.1% |
| Partnership | 6.8% | 6.9% | 6.9% | 6.9% | 6.9% | 26.6% | 20.0% |
Read the five rule-based columns across. They agree with each other to within half a percentage point.
| Model | Credit misallocated |
|---|---|
| First touch | 52.2 pp |
| Last touch | 51.9 pp |
| Linear | 52.1 pp |
| Time decay | 52.0 pp |
| Position based | 52.0 pp |
| Shapley | 11.3 pp |
Every rule-based model misallocates about 52 points of credit out of 100. They are wrong in the same way and by the same amount.
So the argument between first-touch and last-touch isn't a real argument. Teams spend quarters on it. The gap between any two of them is under half a point; the gap between all of them and the truth is fifty.