Attribution models compared

An attribution model does not measure anything. It applies a rule for distributing credit among the contacts that preceded a sale. Every rule embeds an assumption, and the assumption is always partly wrong.

The models and their biases

Why last click persists

It is the default in most tools, it is simple to explain, and it produces confident numbers. It is also the model that most reliably leads to defunding the channels that create demand, because those channels are almost never last.

The recognisable failure pattern: a business shifts budget towards the channels last click favours, results improve briefly, then decline for two years as the demand pool empties.

The measurement gaps no model fixes

All models work only on contacts they can see. In practice they cannot see: cross-device journeys where someone browses on a phone and buys on a laptop; the substantial share of users who decline tracking consent; offline contacts entirely, so print, radio, outdoor, word of mouth; and anything happening inside a platform without a click.

That means the model is dividing credit among perhaps half of the real contacts. It is a partial map, not a measurement.

The practical approach

1. Pick position-based if your volume is modest. It is a reasonable compromise and its biases are mild.
2. Use data-driven if you have several hundred conversions a month.
3. Never compare across models. A channel's numbers will change dramatically; that is the model, not the channel.
4. Add an unmodelled figure. Total spend divided by total customers, which no attribution rule can distort.
5. Test the channels the model dislikes. Run a holdout before defunding anything.

What to tell the board

Report the attributed numbers with a stated model and a stated limitation, and report total cost per customer alongside. The second number is cruder and harder to argue with, and it prevents the discussion from becoming a debate about the tool.