Could agent advertising hide inside platform take rates?

Beyond transaction fees, Jaipuria identifies "one other opportunity, which is around advertising (i.e., specifically trying to drive incremental demand for merchants)," where "a merchant might pay the agent (in addition to the customer directly) to have their offer considered when someone asks for a hotel or a pair of shoes." But he argues this revenue may never surface as a recognizable ad unit: "it is possible that these don't look like ads at all and in an agentic world, could just be paid through higher take rates." He sizes a rough ceiling using existing platforms' ad-revenue-to-GMV ratios, noting "Amazon is the most mature and the best monetized ads business" and that "the ~2% range seems reasonable and achievable for most different kinds of platforms."

His reasoning is explicitly provisional rather than a settled estimate: "there is a lot to be figured out in terms of whether there will be ads at all, and how they work and whether they can be done tastefully." He then names two concrete risks that follow from the mechanism itself. One is a user-experience risk: "proactive ads to suggest things to buy may not be received well by users." The other is an incentive risk, stated directly: "the agent receiving money to choose an item that may not be best for the user magnifies the principal-agent problem" — the economics term for what happens when the party acting on someone's behalf has a financial incentive that diverges from that person's own interest.

This is the companion to How could an AI agent's revenue per transaction grow over time?, supplying the fourth monetization channel his taxonomy names but does not fold into that blended take-rate projection — precisely because, in his account, it may not be separable from the take rate at all. It also sharpens Can an AI agent serve both merchant and user interests fairly? by supplying the general economic term, "principal-agent problem," for the conflict Azhar illustrates narrowly through Muse's Expedia partnership, and by naming a variant Azhar doesn't: an agent paid specifically to favor one offer over another, rather than a flat referral fee on whatever the user already chose. It also complicates Do consumers actually want to delegate shopping decisions to AI?: Accenture's respondents report trusting an agent to purchase on their behalf, but that survey does not ask whether trust would hold if respondents knew the agent's recommendation could be paid for directly by the merchant being compared.

Jaipuria is explicit that this part of the piece is unresolved — "there is a lot to be figured out" — so the ~2% ad-revenue-to-GMV figure is a rough ceiling drawn from other platforms' existing ad economics, not a measurement of any agent's actual advertising revenue, and no agent is reported in the excerpt as currently running this model. The piece does not say whether disclosure or structural separation would mitigate the principal-agent risk it names; it only establishes that advertising folded invisibly into a take rate would be harder for a user to detect or discount than a labeled ad would be.