Multi-Touch Media Attribution

Last-click gives all credit to the final touchpoint before conversion, which systematically over-rewards bottom-of-funnel channels like branded search and retargeting that often merely close journeys other channels started. Berman's analysis shows last-touch over-incentivizes the last exposure and can actually lower advertiser profit relative to smarter rules. Anderl and colleagues likewise find that Markov-based credit allocations differ substantially from last-click, meaning budgets set by last-click can be badly misallocated. Multi-touch methods spread credit according to each channel's contribution across the whole path, giving a fairer and usually more profitable basis for budgeting.

Both are data-driven, but they formalize 'contribution' differently. Markov-chain attribution models journeys as transitions in a graph and values a channel by its removal effect — how much the conversion probability falls if that channel is deleted — naturally capturing path order and being efficient to compute even with many channels. Shapley-value attribution treats channels as players in a cooperative game and assigns each its average marginal contribution over all coalitions, which guarantees fairness axioms but is combinatorial in the number of channels. Markov is often preferred for scalability and sequence sensitivity; Shapley is preferred when axiomatic fairness is the priority. They frequently produce similar rankings and can be used together as cross-checks.

Not by itself. Attribution assigns observed credit based on the paths customers actually took, so it can be confounded by selection — high-intent users gravitate to certain channels, inflating those channels' apparent contribution — and by untracked exposures. Removal effects and Shapley values are counterfactual within the model's assumptions but are not the same as the incremental lift you would measure by randomly withholding a channel. Best practice is to treat attribution as a fast, granular allocation tool and to calibrate or validate it against randomized experiments such as geo-tests, ghost ads, or user-level holdouts, and against uplift models, before making large budget decisions.