Every Attribution Model Is Wrong, Some Are Useful

A client once asked me to find the attribution model that would tell them the 'real' contribution of each channel. I had to tell them that question doesn't have an answer, because attribution isn't measuring a fact that exists in the world — it's applying a rule to split credit for something that, in reality, wasn't split at all. The customer saw five touchpoints and converted. There is no ground truth percentage that says channel three deserves 23% of the credit.

Every model — last click, first click, linear, time decay, data-driven — is a different set of assumptions about how credit should be distributed. None of them are observations. All of them are decisions, and the decision that matters isn't which one is accurate, it's which one's biases match the decisions you're actually going to make with the output.

The bias each model actually encodes

• Last-click rewards the channel that closes, which overvalues retargeting and branded search and undervalues anything upper-funnel.
• First-click rewards discovery, which is great for justifying awareness spend and terrible for optimizing your actual media mix.
• Linear pretends every touch matters equally, which is rarely true but is at least honest about being a compromise.
• Data-driven models are the least biased mathematically, but they're a black box you can't explain to a CFO asking why budget moved.

The question isn't which attribution model is correct. It's which one is wrong in a way that still leads you to a defensible budget decision.

What I actually tell clients to do

Pick a model that matches the decision cadence. If you're making weekly bid adjustments, last-click or a short time-decay window is fine because the stakes per decision are small and reversible. If you're making an annual channel mix decision worth real budget, that's when a cleaner data-driven model or even an incrementality test earns its cost — because the decision is big enough and infrequent enough to justify the extra rigor.

What I push back on hardest is model-shopping — running the same conversion data through three models until one makes the channel someone already likes look good, then presenting that one as 'the' attribution model. That's not analysis, it's picking the lie you prefer. Pick one model, document why its biases are acceptable for the decisions you're making, and stop re-litigating it every quarter a number looks inconvenient.