How Could Shapley Values Pay Publishers for Agent Traffic?

Parag Agrawal says Parallel uses Shapley values to estimate what a page is worth to an agent and what the content owner should be paid. Shapley value is a game-theory split of a result that is larger because several sources were used together.

Remove one source, rerun the agent, and see what quality is lost. If a cent of extra compute, or a better model, restores that quality, the missing source is worth about a cent to the end customer. Computing the full value for one content owner can cost several dollars, more than the dollar you might pay them, so Parallel estimates it from scenarios instead of simulating every subset. Unique, differentiated data should be paid more. A banker in an expensive job should pay more for that read than his retired dad, because the read is part of higher-value work. By his calculation, this math could send meaningful dollars to a wide range of content owners in 12 to 24 months.