"Agentic audiences" is about to enter the vocabulary of every marketing team that buys advertising. Vendors will pitch it. Conference panels will debate it. Very few of the people repeating the phrase will be able to say what it means. This article fixes that in plain English, then connects it to the work you already do on brand visibility.
What just shipped
The IAB Tech Lab, the organization that maintains the technical standards behind online advertising, has released a bundle of related pieces:
- An open standard called Agentic Audiences, donated by the data company LiveRamp. It was previously named the User Context Protocol.
- Version 2.3 of AAMP, short for Agentic Advertising Management Protocols, which describes how software agents should handle advertising tasks end to end. AAMP 2.0 arrived in April 2026 and version 2.3 followed on July 30. IAB describes 2.3 as taking agentic media buying "beyond demos and first live implementations" into production buying at scale, and it adds a vendor approval gate through the IAB Diligence Platform so unapproved vendors are filtered out before agents transact.
- An OpenRTB extension, the piece that lets an audience summary travel inside the standard request format ad systems already use to trade inventory.
- A live module in Prebid, the open-source ad-auction software used by many publishers. In ad tech, "publisher" means any website or app that sells ad space, from news sites to games; it does not refer to news organizations specifically. The module is named agenticAudienceAdapter.
- A new Agentic Audiences Task Force, being launched to maintain the standard going forward. Companies participate through IAB Tech Lab membership.
The plain-English version
Here is the idea in one paragraph. Today, when an advertiser wants to reach a specific group, the ad industry trades lists and labels: cookie IDs, device IDs, segments with names like "in-market for SUVs." Agentic Audiences replaces the label with a summary of meaning called an embedding. An embedding is a list of numbers produced by a machine learning model, arranged so that similar things end up close together. Instead of asking a seller for the SUV-intenders segment, a buyer's software agent sends a mathematical summary of the audience it wants, and the seller's agent scores its own users against that summary.
The design has one job: let a buyer aim at a precise audience without anyone handing over personal data. The seller never hands over a list of people. The buyer never sees individual users. Each side keeps its own data, and the only thing that crosses the line is the mathematical summary.
Think of the standard as an envelope
Anthony Katsur, CEO of the IAB Tech Lab, describes the standard as an envelope. It carries information between agents, and it stays out of the business of deciding what that information means.
The envelope holds three layers of signal, as defined in the spec:
- Identity: who the user is, in hashed or probabilistic form
- Context: what the user is doing right now
- Reinforcement: how the user has responded to ads before, including clicks and conversions
One catch, which the spec itself is candid about: the standard does not say which model has to produce the embeddings. The numbers only mean something inside the model that made them. Think of a temperature reading of 30: on a Fahrenheit thermometer that is a cold day, on a Celsius thermometer it is a hot one. Same number, two different meanings, and the reading only makes sense once you know which scale produced it. Embeddings work the same way. If the buyer's agent builds its summary with one model and the seller's agent scores users with another, the comparison is meaningless. Both sides must agree to use the same model, and the standard leaves that agreement to the companies themselves. Any vendor pitch that skips past that point is skipping past the hard part.
Why not require one model for everyone? A standards body that picked a single model today would freeze the technology at today's quality and hand whichever vendor built that model a permanent advantage. Embedding models improve month to month, while standards change slowly. The IAB Tech Lab standardized the part that needs to be identical on every machine, the envelope and its three signal layers, and left the part that still competes, the math inside, to the companies involved. Shipping containers set the precedent: the world agreed on the size and fittings of the box, never on what goes inside it. The predictable result is that early matches work best where both ends already run the same vendor's software or have negotiated a shared model, and making compatibility broad and easy is now the task force's problem to solve.
What it cannot do
Katsur drew the line himself in the announcement post: "Embeddings solve targeting. IDs still own attribution." An embedding can say that one audience resembles another. It cannot prove that a specific person saw a specific ad and then bought something. Attribution, the part of advertising that ties spend to results, still requires a deterministic chain: a known identity, a logged impression, a recorded conversion.
Anyone who has followed the debate over AI visibility scores will recognize this argument. Aggregate similarity scores are easy to produce and hard to act on. The evidence that matters sits at the level where the machine actually acted: what the model said, when, and to whom. The advertising industry's own standards body just made the same point about its newest technology, which is why Parceit reports query-level evidence in every audit rather than a single fuzzy number.
Why this matters outside advertising
The premise underneath the standard is the same one reshaping search: software agents increasingly mediate how companies reach customers. In advertising, agents now negotiate with other agents. In search, answer engines like ChatGPT decide which brands to mention before a person ever sees a results page, a dynamic documented in [ChatGPT Decides Who to Recommend Before It Searches](https://parceit.com/blog/chatgpt-decides-who-to-recommend-before-it-searches). The same shift is happening in two industries at once, and in both of them the representation of your company is computed and exchanged by machines, out of view.
The IAB Tech Lab is building the plumbing for that machine-to-machine world. It already maintains the Spiders and Bots list that analytics teams use to identify automated traffic, and it has added an Agent Registry for identifying AI agents. When an industry starts issuing registries and envelopes so machines can talk to machines, it is committing to the agent-mediated web. Visibility measurement exists because your brand's reputation now lives inside those exchanges.
What you should actually do
Nothing here requires buying anything this quarter. Four moves are worth making.
- Learn the vocabulary now: agentic audiences, embedding, AAMP, OpenRTB extension. These words will likely show up in vendor decks within months. When a pitch arrives, ask one question: which embedding model do both sides use? If the vendor cannot answer, the pitch is decoration.
- Ask your ad-tech vendors where they stand. If your agency or demand-side platform has joined the new task force, you will hear about standards changes early, which is a reasonable proxy for how seriously they take the agent shift.
- Watch the Agentic Audiences Task Force itself. Participation runs through IAB Tech Lab membership, and the standard will evolve there. Parceit is tracking it and will report changes that matter to visibility teams.
- Connect it to your visibility work. Your audience data and your brand representation are flowing into machine-to-machine channels on parallel tracks. Measurement on both tracks needs evidence at the level where the machine acted: the query, the impression, the bid.
The bottom line
Agentic Audiences is an ad-industry standard for letting software agents trade audience summaries without trading people's identities. It is well designed for what it does, and unusually explicit about what it does not do. For marketing teams, the practical value right now is fluency: understand the standard before the jargon lands in your inbox, and hold every fuzzy score, in advertising or in AI visibility, to the same demand for proof.