Attribution models: who gets the credit for a conversion

Someone reads an article found through a search engine, a few days later sees a social media post, then receives a newsletter and finally types the site's address directly and buys. Which channel produced the sale? The question sounds simple but has no objective answer. Every attribution model is a conventional rule for sharing credit among touchpoints.

The path and its touchpoints

Every visit that precedes a conversion is a touchpoint. In the example the path is:

1. organic search;
2. social;
3. email;
4. direct visit, with purchase.

Models differ in how they divide the value of the conversion across these four steps.

The most common models

The same purchase can therefore be credited entirely to organic search or entirely to direct traffic, depending on the model chosen.

What each model tells you

• Last click rewards channels that close: email, retargeting, searches for the brand name. It tends to undervalue whatever introduced the site.
• Last non-direct click fixes an obvious flaw: anyone who types the address already knows it from somewhere, so a direct visit is rarely the real cause.
• First click rewards discovery channels: organic search on general questions, shared articles, online word of mouth.
• Linear is neutral but naive: it treats a three-second visit the same as a twenty-minute one.
• Position-based recognises both discovery and closing.
• Time decay suits short cycles, where the last interactions genuinely carry more weight.

There are also data-driven models, in which an algorithm estimates each channel's contribution by comparing paths that convert with paths that do not. They are interesting, but they need large volumes and behave like black boxes: it is hard to explain why a channel received a particular weight.

Limits no model solves

Any model works only on the touchpoints it can see. In practice it misses many:

• different devices: someone who reads on a phone and buys on a computer appears as two people unless they sign in;
• deleted or declined cookies: the path breaks and starts again from scratch;
• untrackable channels: a spoken recommendation, a podcast, a private message;
• attribution window: touches older than the window, say thirty days, are ignored.

Then there is a conceptual limit: attribution distributes credit, it does not measure causation. The fact that a channel appears often in buyers' paths does not mean those sales would vanish without it. To measure a channel's real effect you need experiments, for example pausing a campaign in some regions and comparing them with similar ones.

Using them sensibly

In my work I suggest a pragmatic approach:

1. choose a reference model and use it consistently in periodic reports;
2. compare it now and then with first click, so as not to forget discovery channels;
3. look at the most frequent paths, not just the credited percentages: they often say more;
4. when deciding whether to cut an important channel, do not rely on attribution alone and, if possible, run a test.

For small sites with few conversions a month, the most reliable solution is often the simplest: ask in the contact form how the person heard about the service. The answers are imprecise, but they capture exactly the channels the tools cannot see.

In short

• An attribution model is a convention for dividing credit among touchpoints.
• Last click rewards channels that close; first click rewards those that bring discovery.
• Models work only on visible touches and miss devices, cookies and offline channels.
• Attribution does not measure causation: that requires experiments.
• Use a consistent reference model and compare it periodically with other perspectives.