How Shapley value works
The idea comes from co-operative game theory, where several players produce a result together and the problem is dividing the winnings fairly. The Shapley method asks, for each player, how much the group’s result improves when that player joins — then averages that marginal contribution across every possible order in which the players could have joined.
Applied to marketing, the players are channels or touchpoints and the result is a conversion. Instead of asking which advert was last, the method asks how much the chance of conversion rises when paid search is added to a journey that already contains organic and email, and repeats that question for every combination. Credit ends up spread according to how much each channel genuinely improved the odds, not according to where it sat in the sequence. Google’s data-driven attribution draws on this family of methods.
Why Shapley value matters
Rule-based models are arbitrary. Last click hands everything to the final touch, first click to the earliest, and a position-based split is simply a number somebody chose. None of them reflects how the combination actually performed.
Shapley value replaces the arbitrary rule with a property worth having: a channel that never changes the outcome receives no credit, two channels that contribute equally receive equal credit, and the shares always add to the whole conversion rather than exceeding it. That last property alone makes it more defensible than adding up what several platforms each claim.
Where Shapley value goes wrong
It is still correlational. The method observes which combinations of touchpoints preceded conversions; it cannot tell you what would have happened if a channel had been switched off. Remarketing appears in almost every converting journey precisely because it targets people already close to buying, so it can score well without adding anything. Only an incrementality test answers that.
It is also demanding. The calculation grows quickly with the number of channels, so real implementations approximate rather than evaluating every ordering exactly. It needs a large number of complete journeys before the estimates settle, which rules out most small advertisers. And it can only see the touchpoints you recorded: anything lost to consent refusal, cookie expiry, a device switch or a closed platform is invisible, so the credit is redistributed among the channels that happened to be measurable.
How to act on it
Use it to correct obviously unfair reporting rather than to set budgets by itself. If an upper-funnel channel is being written off because last-click gives it nothing, a Shapley-style model is a fairer basis for the argument, and it will usually shift credit towards discovery and away from the final touch.
Then check the conclusion the only way it can be checked: hold the channel back and see what happens. Read the model as one view among several, alongside your own sales records, and be honest when your data volume is too thin to support it. For most businesses, clean and consistent conversion tracking plus a blended view of results will do more good than a sophisticated model fed with incomplete journeys.