What is AdCP (Advertising Context Protocol)?

What AdCP is for

Programmatic advertising already has a protocol for machines talking to machines: OpenRTB. It assumes a page, a slot and a bid request that describes both. None of those survive the move to a conversation, and OpenRTB has no vocabulary for an autonomous buyer that can negotiate rather than simply bid.

AdCP is the attempt to fill that gap. It describes advertising as a set of operations an agent can perform — find inventory that matches an intent, ask what it costs, agree terms, place the buy, report what happened — rather than as a bid on a slot. It is frequently described as OpenRTB for the AI era, which is accurate about the ambition if not about the shape.

The practical claim is interoperability. Without a shared protocol, every buyer integrates separately with every AI surface, and the surfaces with the most users win by default because they are the only ones worth the integration cost. That outcome is what AI walled gardens describes.

How AdCP relates to MCP

AdCP does not invent its own transport. It builds on MCP (Model Context Protocol), the open standard for connecting AI applications to external tools and data.

The division is clean. MCP defines how an agent calls a tool at all — how it discovers what is available, passes arguments and receives results. AdCP defines the advertising-specific tools: what an inventory query looks like, what a media package contains, how a transaction is confirmed, what a delivery report must include. One is the wiring, the other is the vocabulary spoken over it.

What AdCP specifies

The specification covers the stages a media buy actually passes through, expressed so an agent on either side can drive them.

• Discovery. An agent describes what it wants to reach — an intent, a topic, an audience condition — and receives the inventory that matches it.
• Negotiation. Terms, pricing and constraints are exchanged as structured data rather than as a rate card and an email thread.
• Transaction. The buy is confirmed, with the obligations of both sides recorded.
• Delivery and reporting. What ran, where, and what it produced, returned in a form the buying agent can act on without a human reading a dashboard.
• Sponsored intelligence. A layer for paid experiences delivered inside the response rather than beside it — see sponsored intelligence (SI).

Why a protocol matters more here than it did on the web

On the open web, a buyer who could not integrate with a publisher could still reach that publisher's audience elsewhere. Audiences were addressable across sites, so no single property was essential.

Inside AI assistants there is no cross-surface identity to fall back on and no long tail of interchangeable inventory. Reaching a user in a conversation means reaching them on the assistant they are actually using. If buying that assistant requires a private integration, then access is a commercial negotiation rather than a market, and the surfaces set the terms.

AdCP is the counter-proposal: make access a protocol rather than a relationship. Whether it succeeds depends less on the specification than on whether the largest surfaces adopt it, which is the open question of the AI open web.