Attribution modelling is the framework you use to decide how marketing credit gets distributed across the touchpoints in a customer's path to purchase. Where attribution is the question ("who gets the credit?"), modelling is the rule you apply when answering it.
I tell clients to pick a model deliberately rather than inheriting whatever their analytics tool defaults to, because the default in most tools is last-click — and last-click is the model that most distorts where money should go. The scenario plays out the same way every time: a content effort drives discovery, email or paid search closes the sale, and the dashboard tells you "email is working, content isn't." Cut the content budget on that signal and conversions slowly fall over the next two quarters, with no obvious cause in the dashboard.
In practice, the modelling choice is a trade-off between simplicity and accuracy. Rule-based models (linear, time-decay, position-based) are easy to explain to a non-technical leadership team and easy to audit. Algorithmic models — Google's Data-Driven Attribution, Adobe's Algorithmic, or a Markov-chain model run in a tool like R or Python — are more accurate when you have enough conversion volume, but harder to defend when somebody asks "why did this channel get more credit this month?".
For most service businesses with under a few hundred conversions a month, a position-based model gives most of the value without the complexity. Above that volume, data-driven attribution starts pulling its weight. The wrong choice isn't picking the "wrong" model — it's never picking one at all.