A buyer's agent that wants to run a campaign across a news site, a podcast network and a connected TV app currently needs three different integrations, three sets of specifications and three sales conversations. The Ad Context Protocol replaces that with one set of callable tasks - find products, propose a buy, sync creative, report delivery - that any compliant seller agent answers in the same way. It sits above the impression auction rather than inside it, and since its third major release it also covers who a brand is, what an agent is allowed to spend and how a brand can appear inside a chatbot conversation.
What is AdCP?
AdCP, the Ad Context Protocol, is an open-source specification published by AgenticAdvertising.org (AAO) that defines how AI agents discover inventory, plan campaigns, execute buys, generate creative, apply governance and report results across advertising platforms. It launched on 15 October 2025 with six founding companies: Scope3, Yahoo, PubMatic, Swivel, Triton Digital and Optable. The specification is licensed under Apache 2.0 with an express patent grant.
Technically, AdCP is an advertising vocabulary layered over general agent plumbing. Agents speak to each other over Anthropic's Model Context Protocol (MCP) or the agent-to-agent (A2A) protocol; AdCP defines what they say when the subject is media. At the time of writing the published documentation resolves to version 3.1.21.
What changed in AdCP 3.0?
AAO published AdCP 3.0 on 21 April 2026, six months after launch, and described it as the moment agentic advertising "moved from experiment to infrastructure". The release is a change of scope rather than a feature update. Version 2 covered three areas: media buying, signals and creative. Version 3 organises the protocol into six domains that an agent can declare support for: media buy, creative, signals, governance, brand and sponsored intelligence.
The main additions are:
- brand.json. A machine-readable file at /.well-known/brand.json in which a brand declares its identity, brand hierarchy, visual guidelines and the operators it authorises to buy on its behalf. It is the buy-side counterpart to adagents.json, the file publishers use to declare which agents may sell their inventory.
- Governance. Agents can check a campaign plan against budget authority, brand suitability and regulatory policy before spend is committed. When a governance agent is attached to a plan, sellers must call check_governance and reject a spend commitment without a valid, signed approval. AAO says thirteen regulatory policies ship pre-seeded in the registry, covering US, EU and platform-level rules.
- Sponsored Intelligence. A protocol for brand sessions inside AI assistants, discussed below. It ships as experimental.
- Twenty channels. The channel model grew from nine to twenty, adding linear TV, radio, streaming audio, gaming, influencer, cinema, email, print and sponsored intelligence, along with forecasting, guaranteed buys and a formal accounts model for billing.
- A trust layer for real money. Every state-changing request must carry an idempotency key so retries cannot double-book spend, webhooks must be signed, and request signing under RFC 9421 is mandatory for agents claiming the AdCP Verified status.
Version 2 was fully deprecated on 1 August 2026. Any publisher or vendor that implemented AdCP in 2025 on the earlier schema is now running an unsupported version.
How does an AdCP transaction work?
A buyer agent discovers a seller agent, asks it what it supports through get_adcp_capabilities, then searches its inventory in natural language through get_products. The seller answers with structured products: placements, formats, pricing and forecasts. The buyer proposes a media buy, the seller accepts, negotiates or routes it to a human for approval, and creative and reporting flow through the same channel. Responses can be asynchronous, so a proposal that needs a sales team's sign-off can sit for hours or days without breaking the workflow.
The sell-side reference implementation is the Prebid Sales Agent, which Prebid.org took over in January 2026. The specification and the reference code are deliberately held by different organisations: AAO owns the standard, Prebid maintains the software.
Is AdCP a replacement for OpenRTB?
No, and AAO now says so in the protocol's own documentation. AdCP operates at the campaign and deal layer and hands off to OpenRTB infrastructure where impression-level execution is needed. The useful way to think about the split is that OpenRTB standardised the auction for standard units, and AdCP standardises the conversation before it: what is for sale, on what terms, under which rules. That is why its early traction has come in direct and programmatic guaranteed deals, and in inventory that never fitted the bidstream at all, such as broadcast spots, newsletter placements and podcast reads.
Who governs AdCP?
AgenticAdvertising.org is structured as an industry consortium with four voting classes - brands, agencies, publishers and technology providers - each with ten board seats and equal voting weight. Directors serve staggered three-year terms. The first elected board was due to be seated at AAO's first annual general meeting on 6 May 2026, replacing a four-director interim board.
Governance is the part a publisher should check rather than assume. An open issue in the AdCP repository notes that, months after the meeting, the public governance pages and the charter still describe the interim board and the election as a future event, and asks for the elected directors to be published. It also asks for the copy on Scope3 concentration and recusal to be updated. Scope3's chief executive, Brian O'Kelley, is the protocol's most visible advocate, and whether a standard incubated by one company becomes genuinely neutral is the question every advertising standard has faced. The answer here depends on a board roster that was not yet public at the time of writing.
On adoption, AAO states that AdCP-enabled sales agents are in production for platforms including Snap, Pinterest, Reddit, Netflix and Vox Media. That is the organisation's own claim, not an audited count. Membership continues to grow: ExchangeWire reported on 8 September 2026 that the media sales software company ADvendio had joined AAO. Google, Amazon, The Trade Desk and Microsoft, which run the demand-side platforms handling most programmatic spend, were absent from the early adopter list.
Why is IAB Tech Lab at odds with AdCP?
Because the two bodies disagree about whether the agentic era needs new standards or extended old ones. AdCP builds natively for agents. IAB Tech Lab's Agentic Roadmap, announced in January 2026, instead extends OpenRTB, AdCOM, OpenDirect, VAST and the privacy frameworks with agentic layers, and in February 2026 its work was grouped under the name AAMP, the Agentic Advertising Management Protocols. Digiday framed the two in February as competing philosophies for the same problem.
The dispute sharpened in September. On 17 September 2026 Hillary Slattery of IAB Tech Lab published a direct rebuttal of adagents.json, AdCP's replacement for ads.txt, calling a new specification "unverified and unverifiable" and arguing that it drops the counterparty check that sellers.json provides. AdCP's answer is that brand.json supplies the other side of the relationship. That argument is live and is covered in more detail in our page on who is authorised to sell ads in AI answers.
For publishers the practical conclusion is that there will be more than one protocol for some time. Prebid, which maintains the AdCP reference code, has said it will stay agnostic between them.
What does AdCP 3.0 still not do?
AAO publishes a known-limitations page, and three items on it matter to publishers. First, AdCP Verified is self-attested: an agent runs the compliance suite and publishes signed output, but AAO does not audit it, so the badge means the agent says it passed, not that anyone checked. Independent audit is on the 3.1 roadmap. Second, the protocol does not carry user consent itself; consent still travels through TCF, GPP and their equivalents. Third, Sponsored Intelligence is experimental, with session handling, identity and consent all subject to change between 3.x releases on six weeks' notice.
What does AdCP mean for publishers?
It is three separate things, and it is worth taking them one at a time.
As a sales channel, AdCP is a way to make inventory legible to buyer agents, including inventory that programmatic never priced well. The cost of entry is standing up a sales agent, mapping products to your ad server and keeping pace with a specification that has shipped a breaking major version in six months.
As an authorisation question, it asks publishers to publish adagents.json alongside, not instead of, ads.txt. With IAB Tech Lab arguing against the file, the low-risk position is to maintain both.
As a route into AI assistants, Sponsored Intelligence is the part most often misread. It defines a session between an AI assistant, which acts as the host, and a brand agent, which the user has agreed to engage. The publisher whose reporting informed the assistant's answer is not a party to that session. Sponsored Intelligence standardises how a brand talks to a user inside a chatbot; it does not route any value to the sites the chatbot read.
That is the boundary of every agentic buying standard so far. AdCP, AAMP and the Prebid Sales Agent all describe how machines buy and sell inventory that a publisher controls. None of them covers the moment an AI agent fetches a publisher's page to build an answer, because at that moment there is no ad slot to sell. That retrieval is the surface blankspace works on, detecting agent traffic at the CDN edge and placing contextual brand facts into what the agent takes away. It is a different problem from the one AdCP solves, and publishers need an answer to both.
Frequently asked questions
Who created AdCP?
AdCP launched on 15 October 2025 with six founding companies: Scope3, Yahoo, PubMatic, Swivel, Triton Digital and Optable. It is now published by AgenticAdvertising.org, an industry consortium with brand, agency, publisher and technology voting classes. Prebid.org maintains the sell-side reference implementation separately.
What is the latest version of AdCP?
AdCP 3.0 was published on 21 April 2026, and the documentation currently resolves to the 3.1 line. Version 2 was fully deprecated on 1 August 2026, so implementations built on the original schema are no longer supported.
What is the difference between AdCP and AAMP?
AdCP is AgenticAdvertising.org's standalone protocol built natively for AI agents. AAMP, the Agentic Advertising Management Protocols, is IAB Tech Lab's programme of extensions to existing standards such as OpenRTB, AdCOM and OpenDirect. They address the same shift, they are not formally aligned, and IAB Tech Lab has publicly criticised parts of AdCP, including adagents.json.
What is AdCP Sponsored Intelligence?
Sponsored Intelligence is the part of AdCP 3.0 that governs brand sessions inside AI assistants. A user expresses interest, grants consent, and the assistant hands the conversation to a brand agent, with an optional handoff to purchase. It is marked experimental and may change between 3.x releases. It involves the assistant and the brand, not the publishers whose content the assistant used.
Should publishers adopt AdCP now?
Publishers with engineering capacity and direct-sold or non-standard inventory can reasonably pilot a sales agent now, since AAO reports production deployments at several large platforms. Publishers without that capacity lose little by waiting for the governance record, independent verification and the standards dispute with IAB Tech Lab to settle. In either case, keep ads.txt and sellers.json current, since they remain the files most buyers check.