Attribution models explained: who gets the credit for a sale?

A customer rarely buys after a single visit. They might open an e-mail, click a search ad a week later, find you again through organic search and finally type your address and buy. The sale is worth the same whatever happened, but who gets the credit decides which channels look good and where budget goes. An attribution model is the rule that divides that credit.

One journey, five models

Take a customer who buys something worth 100. Their path, and how long before the purchase each touchpoint happened:

1. Email, 12 days before
2. Paid search, 6 days before
3. Organic search, 3 days before
4. Direct (typed the address), on the day of purchase

This script gives the 100 out under each model.

JOURNEY = [("Email", 12), ("Paid search", 6), ("Organic search", 3), ("Direct", 0)]   # (channel, days before purchase)
VALUE = 100.0


def first_click(j):
    return {c: (VALUE if i == 0 else 0.0) for i, (c, _) in enumerate(j)}


def last_click(j):
    return {c: (VALUE if i == len(j) - 1 else 0.0) for i, (c, _) in enumerate(j)}


def linear(j):
    return {c: VALUE / len(j) for c, _ in j}


def time_decay(j, half_life_days=7):
    weights = [0.5 ** (days / half_life_days) for _, days in j]
    total = sum(weights)
    return {c: VALUE * w / total for (c, _), w in zip(j, weights)}


def position_based(j, first=0.4, last=0.4):
    n = len(j)
    middle = (1 - first - last) / max(n - 2, 1)
    out = {}
    for i, (c, _) in enumerate(j):
        out[c] = VALUE * (first if i == 0 else last if i == n - 1 else middle)
    return out


models = {"First click": first_click, "Last click": last_click, "Linear": linear, "Time decay": time_decay, "Position-based": position_based}
print(f"{'Channel':<16}" + "".join(f"{name:>16}" for name in models))
for channel, _ in JOURNEY:
    print(f"{channel:<16}" + "".join(f"{fn(JOURNEY)[channel]:>16.1f}" for fn in models.values()))

Channel              First click      Last click          Linear      Time decay  Position-based
Email                      100.0             0.0            25.0            11.7            40.0
Paid search                  0.0             0.0            25.0            21.2            10.0
Organic search               0.0             0.0            25.0            28.6            10.0
Direct                       0.0           100.0            25.0            38.5            40.0

Every column adds up to 100. Only the rule changed, and the same sale makes e-mail look worth everything, or nothing.

The models

First click. All the credit goes to the first touchpoint. It rewards the channels that introduce people to you and ignores everything that convinced them.

Last click. All the credit goes to the final touchpoint. It is simple, and it flatters whatever comes last: branded search, direct visits and retargeting. It undervalues the channels that started the journey. Here "Direct" gets the whole sale, though it is unlikely to be what created the interest.

Linear. Equal credit to every touchpoint. It acknowledges the whole path but treats a decisive click and a casual one as the same.

Time decay. Touchpoints closer to the conversion get more credit. In the script credit halves for every 7 days of distance, so the e-mail 12 days out gets 11.7 and the direct visit gets 38.5. It suits short buying cycles, where recent contact matters most, and it underrates the long early nurture of a slow one.

Position-based (U-shaped). A fixed share, here 40%, to the first and last touchpoints, and the remaining 20% shared among the middle ones. It values both the introduction and the close.

Data-driven. Instead of a fixed rule, a statistical model estimates each touchpoint's contribution from many real paths, including those that did not convert. It can be the most informative when you have enough data, but it is harder to explain and to audit.

What every model gets wrong

• They only see what is tracked. A podcast mention, a conversation or an offline ad leaves no touchpoint. Browser privacy features, consent choices and ad blockers hide others.
• Credit is not cause. A touchpoint that appears on the path may have changed nothing: the customer was going to buy anyway. Only a controlled experiment, such as a holdout group that does not see the channel, measures the effect of a channel.
• The lookback window matters. A 7-day window and a 90-day window produce different answers from the same data.
• People use several devices. Without a way to join them, one person looks like several, each with a short path.

How to use this well

1. Compare, do not crown. Run two or three models side by side. A channel that looks good under all of them is probably good. One that changes completely from model to model is a question mark, not an answer.
2. Match the model to the decision. Judging a brand awareness campaign by last-click tends to make it look poor. Judging a retargeting campaign by first-click tends to make it look poor.
3. Write down the rule you use. Teams argue about numbers that differ only because each used a different model or window.
4. Test the big claims. When a lot of money rides on a channel, a holdout or geographic test shows what attribution cannot.