Agentic Audiences is the component of IAB Tech Lab's AAMP framework that defines how advertising agents exchange audience signals. It standardises an envelope for embeddings: compact vectors that encode who a user is, what they are doing now and how they respond to advertising, plus the metadata, consent fields and retention limits a receiving agent needs. According to IAB Tech Lab, the protocol began as the User Context Protocol (UCP), donated by LiveRamp, and its core specification is still a draft.

TL;DR. The Agentic Audiences repository describes IAB Tech Lab's buyer and seller agents as AAMP's control layer and itself as the data layer beneath them. Its draft specification defines the embedding, its model, its consent and retention terms, and when two vectors can be compared. An OpenRTB community extension already carries these embeddings inside the bid request's existing user.data objects. The open questions for data providers, publishers and buyers are about control: which model produced a vector, what it may be used for, how long it may be kept, and who checks.

What is the agentic-audiences component of AAMP?

The agentic-audiences component is one of seven repositories listed on the AAMP hub on GitHub, described there as "an open protocol for intelligent interoperability across advertising agents". IAB Tech Lab's standards page says it joined the open-source agentic initiative through a donation from LiveRamp, and IAB Tech Lab said a lightweight Commit Group would govern it.

The repository calls itself LiveRamp's initial proposal, open-sourced so the community can agree a standard. The exchange specification is labelled Draft v0.1, the embedding taxonomy is marked as a proposal. It is early, open work that the community is shaping, so teams building against it should pin the version they use and follow how the draft develops.

What does Agentic Audiences standardise?

Agentic Audiences standardises how agents exchange embeddings of three signal types: identity (who the user is), contextual (what the user is doing right now) and reinforcement (how the user responds to advertising). According to IAB Tech Lab, embeddings of 256 to 1,024 dimensions can carry these signals within the response times real-time bidding requires.

The draft exchange specification sets out what travels with each vector. An embedding object states the vector's type, dimension and data type. A model descriptor identifies the model and version, the similarity metric and an embedding space identifier. A consent block references the governing framework, lists permissible uses and sets a time-to-live. A security section recommends mutual TLS between agents and signed envelopes.

The interoperability rule matters most commercially. According to the specification, two vectors are comparable only if they share an embedding space, or a known converter maps between spaces, and their dimension, metric and normalisation are compatible. An embedding is only useful to a counterparty that understands the model behind it.

How does Agentic Audiences relate to existing IAB audience work?

Agentic Audiences builds on existing IAB standards rather than replacing them, in three documented ways.

The first is the bid request. An OpenRTB community extension carries embeddings in BidRequest.user.data using the standard Data and Segment objects. Each data provider contributes one Data object, and the vector and its metadata sit under Segment.ext.aa. IAB Tech Lab's AAMP 2.3 release post says this extension, with a Prebid module, makes Agentic Audiences ready for the existing programmatic workflow.

The second is consent and transparency. When the protocol was donated, IAB Tech Lab said it would apply the Data Transparency Standard, the Transparency and Consent Framework (TCF) and the Global Privacy Protocol to it.

The third is taxonomy. The AAMP reference buyer agent builds an audience plan from three kinds of reference: segments from the IAB Audience Taxonomy, categories from the IAB Content Taxonomy, and references to Agentic Audiences embeddings. Shailley Singh, IAB Tech Lab's COO, gave the reason in the AAMP 2.3 post: "audiences are where programmatic deals actually live or die."

How do AAMP's reference agents agree an audience?

The reference buyer and seller agents treat the audience as a term of the deal. According to the buyer agent's wire-format documentation, an audience plan carries a primary audience plus optional constraints, extensions and exclusions, and any embedding reference must carry a compliance context stating its jurisdiction and consent framework. A seller that cannot honour part of a plan returns a structured rejection naming each unsupported element, and the buyer agent degrades the plan and retries. At booking, both sides log a hash of the plan, which the documentation calls the forensic anchor for later disputes.

In a hypothetical case, a buyer agent sends an Audience Taxonomy segment as primary and an embedding as an extension. The seller does not support extensions and rejects that element by path. The buyer drops it, retries, and both systems log the same plan identifier.

What does Agentic Audiences mean for data providers, publishers and buyers?

For data providers, the model becomes the product. The repository's architecture document envisages external model builders producing compatible embeddings and competing on quality, with each model descriptor telling buyers whether its vectors can be compared with anything else they hold.

For publishers, the minimum input the repository describes is modest: page title and keywords, supplied through meta tags, data layers or server-side. Its Prebid module is labelled an example; vendors are told to set their own Global Vendor List ID and use the maintained module in Prebid.js.

For buyers, the question shifts from whether a user is in a segment to how well a campaign fits a context. The repository includes a reference scoring service designed to run alongside a DSP, but a score means little when buyer and seller models do not match.

What control and privacy questions should each party ask?

These questions are not legal advice; the answers belong with each party's privacy and legal advisers.

Which consent framework and consent string does each embedding reference, and who checked it?
Which permissible uses are declared? The draft defaults to activation scoring only when none are listed.
Where is the time-to-live enforced? The draft says receivers must enforce time-to-live and purposes before storing or training on an embedding.
Could an identity embedding be reversed into personal data? The draft taxonomy says embeddings derived from personal data should not be reversible.
Are envelopes signed, and are message identifiers and hashes logged for audit?

Anthony Katsur, IAB Tech Lab's chief executive, framed the wider problem when the protocol was donated: "the industry will need to collectively agree on what consent means when a consumer's only point of contact with a company is through a single agent-to-agent interaction."

Where does Agentic Audiences sit alongside other protocols?

AdCP, the Ad Context Protocol from AgenticAdvertising.org, handles audience data through its own tasks, including sync_audiences, get_signals and activate_signal. Buyers should ask vendors which protocols they support, and at which version.

An audience plan is one term in a deal, and each side needs the same record of every term. Alkimi runs a neutral marketplace that works within IAB Tech Lab's AAMP framework; its agent is listed in the IAB Tech Lab Agent Registry, and the marketplace is being built to be AdCP compatible. Deals agreed there are recorded on a shared DealSheet both parties hold, with humans keeping approval rights at the deal level.

Frequently asked questions

Is Agentic Audiences the same as UCP?

Yes. Agentic Audiences was formerly called the User Context Protocol (UCP), which LiveRamp donated to IAB Tech Lab's open-source agentic initiative.

Is Agentic Audiences a finished standard?

No. The exchange specification is labelled Draft v0.1, and the repository describes itself as an initial proposal for the community to develop.

Does Agentic Audiences replace the IAB Audience Taxonomy?

No. The AAMP reference buyer agent uses Audience Taxonomy segments, Content Taxonomy categories and Agentic Audiences embeddings side by side in one audience plan.

Is there an Agentic Audiences or AAMP certification?

IAB Tech Lab's published AAMP material describes none. It runs an Agent Registry where companies list their agents, with entries reviewed for validation information.

What Is AAMP's Agentic Audiences Component and How Do Agents Work With Audiences?