<b>When Shapley and Markov disagree, who's right?</b>

<b>When Shapley and Markov disagree, who's right?</b>

The question: two of the most respected algorithmic attribution methods routinely assign different credit to the same channel. Which do you trust?

What the methods actually do:
— <b>Shapley value</b> (borrowed from cooperative game theory) asks: across every possible ordering of channels in a path, what is each channel's average marginal contribution? It's fair by construction but order-agnostic.
— <b>Markov chains</b> model paths as transition probabilities between states, then compute a 'removal effect' — how much conversion volume collapses if you delete that channel entirely.

Why they diverge: simulation studies (and replications on real conversion logs) show the gap widens for channels that appear mid-path. Markov rewards channels that are structurally hard to route around; Shapley rewards channels that lift the average regardless of position. Upper-funnel display tends to score higher under Markov's removal logic; mid-funnel retargeting often scores higher under Shapley.

The nuance: neither is 'measuring incrementality.' Both are sophisticated ways of dividing observed credit — a descriptive exercise, not a causal one. A channel can score high in both and still be non-incremental if those users would have converted anyway. Correlation in the path data is not causation about the spend.

What to actually do: run both. Treat large disagreements as a flag, not a tiebreaker — they usually point to a channel whose role is positional rather than persuasive. Reserve causal claims for a holdout test.

<b>Bottom line for practitioners:</b> divergence between Shapley and Markov is signal about path structure, not a bug to resolve. Use it to decide where an incrementality experiment is worth the cost.