Agentic innovation in advertising is moving at breakneck speed. With the AAMP (Agentic Advertising Management Protocols) framework 2.3 release, Tech Lab has been busy defining innovative new standards as new use cases are suggested and discovered by the industry. One of the most exciting new use cases is the exchange of audience data in a privacy-preserving manner using the concept of Vector Embeddings via the Agentic Audiences standard, donated to IAB Tech Lab by LiveRamp.
But the standard by itself does not add value until you can operationalize it. This is now possible, thanks to an OpenRTB community extension and a Prebid module update that enable passing the necessary data to use the standard in a bid request. This not only makes the data exchange interoperable, but also transaction-ready.
For those of you whose eyes are glazing over and still stuck on “What the heck are Vector Embeddings?” we will start with a little explanation. (Feel free to skip over the next section if cosine similarity and Euclidean distance are already in your day-to-day vocabulary!)
What are Vector Embeddings?
Vector embeddings are numerical representations of data like words, images, or user IDs, mapped into continuous, high-dimensional space so that similar items lie close together. Created by machine learning models, they capture semantic meaning, context, and relationships beyond simple labels. For example, words with related meanings have vectors with small distances between them. Embeddings enable tasks like search, recommendation, clustering, and classification by allowing algorithms to compare items using distance or similarity metrics. They are foundational in modern AI systems, powering language models, semantic search engines, and personalization systems by transforming complex, unstructured data into structured, computable forms for many real-world applications.
A simple analogy: imagine rating every movie on scales like “how much action,” “how romantic,” “how funny,” “how dark.” Each movie gets a set of scores. Movies with similar scores are probably similar movies. An embedding does this automatically, but instead of four human-picked categories, it uses hundreds of dimensions that a model has learned on its own, capturing nuances no person would think to label.
Why this matters for advertising:
- Ad targeting: a user who’s been browsing “trail running shoes” and “lightweight hiking gear” gets an embedding reflecting that interest. Ads with embeddings nearby (a new trail running shoe launch, a hydration pack) get served to them, even if the ad copy never uses the words “trail” or “hiking.”
- Creative matching: if you have hundreds of ad variants, embeddings let you find which ones are conceptually similar so you’re not accidentally running near-duplicate creative in the same auction, or so you can group variants by theme (humor-driven vs. price-driven vs. aspirational) without manually tagging each one.
- Lookalike audiences: your best customers get turned into embeddings, then you find other users whose embeddings land nearby: people who behave similarly even if they don’t share obvious demographics.
- Brand safety: an ad’s embedding can be compared against a webpage’s embedding to catch mismatches in tone or topic before the ad runs there, not just relying on keyword blocklists.
- Performance clustering: thousands of past ad campaigns can be clustered by embedding similarity to see which creative themes actually drove conversions, surfacing patterns like “aspirational lifestyle imagery outperforms product-shot ads for this segment.”
Semantic Map Vector Embedding Example
The interactive example below is an honest projection of the same vectors calculated as numbers, so two dimensions can’t capture everything. The neighbor lines use the full-dimensional vectors, which is why some “close” relationships connect points that look far apart on the flattened map. You can play with this Vector Embeddings demo in more detail here.
In advertising, vector embeddings can be used to find the closest match between the right ad and the right user and/or content more intelligently than traditional keyword or rule-based systems based on exact match.
Users, ads, and content can be converted into embeddings that capture meaning, interests, and context. For example, a user who reads about “electric cars” and “renewable energy” will have an embedding close to ads about EVs or solar products, even if the exact keywords don’t match. This enables semantic targeting rather than just keyword targeting. When in place, this also makes the user targeting more flexible.
They’re especially valuable in a post-cookie world, where they enable platforms to build audience segments from behavior patterns and target these audiences in a privacy-preserving manner, since the raw identity data is encoded. This is where Agentic Audiences comes in. It standardizes interoperability between systems.
Agentic Audiences is an Envelope
At a high level, Agentic Audiences, which is open-sourced on IAB Tech Lab’s GitHub repository, defines how Buyer and Seller agents communicate information about how embeddings that represent audience intelligence are created.
The specification defines how autonomous agents (such as Buyer and Seller agents developed from the AAMP SDKs) can exchange audience intelligence that they can use in their models, as long as the Buyer and Seller are using the same model. Rather than relying on static audience segments or batch-processed data, each agent continuously generates signals about users based on real-time intent and ad response. These signals are structured into three layers:
- identity (persistent or probabilistic identifiers)
- context (current environment or activity)
- reinforcement (feedback such as engagement, conversions, or outcomes).
The Agentic Audiences specification, critically, is NOT an embedding model. The specification knows nothing about how the embeddings were generated or what is in them. Those models are largely proprietary and are the secret sauce of the companies utilizing the specification. Agentic Audiences is the transport layer between counterparties using the same model.
The Agentic Audiences standard is a messaging envelope that describes information about an embedding and the model that was used to generate it. A sender inserts an embedding, and a receiver can look at the envelope and determine what model it needs to interpret the embedding within it. Of course, for an embedding to be useful, the receiver of an embedding must use the same model that was used to generate the embeddings.
Think of it as audience segments, but improved, since models can do a much better job at clustering and identifying relationships than humans at scale. Instead of a set of targeting segments using lists of cookie-based user IDs, a campaign will have a set of targeting embeddings. An incoming bid request would have a set of embeddings the user has been tagged with, and a bidder would then just look at the incoming embeddings, look at the model that was used, and compare the embeddings against a campaign’s targeting embeddings to evaluate the bid. If any related embeddings are retrieved, it means that the incoming bid request embedding matched against a campaign’s targeting.
Embeddings Solve Targeting. IDs Still Own Attribution.
Vector embeddings are great at answering “who’s likely interested in this?” They compress behavior into a similarity space, so targeting becomes a nearest-neighbor problem: find users whose vectors sit close to your best customers, and serve them the ad. The fuzziness is a feature: two users can have near-identical embeddings while being totally different people, and that’s fine, because targeting only needs “close enough.”
Attribution needs the opposite. You need to know, with certainty, that user 4471829 saw this exact impression at this exact timestamp, then converted on order #12345 twelve days later. That’s a deterministic join, not a similarity match. A vector can tell you two users look alike; it can’t tell you which specific user clicked which specific ad and bought which specific product. That takes stable identifiers such as device IDs, hashed emails, click IDs, and order IDs stitched into a real causal chain.
So the two live side by side: embeddings decide who sees an ad, IDs decide whether that ad gets credit for the outcome. It’s also why the death of third-party cookies has hit attribution harder than targeting. Contextual signals and embeddings can patch targeting reasonably well, but there’s no similarity-based substitute for “this is definitely the same person.”
One caveat: probabilistic attribution does lean on similarity techniques (device fingerprinting, pattern matching) when a deterministic ID isn’t available. But that’s a lower-confidence fallback, used precisely because it lacks the certainty a real ID provides.
Agentic Audiences in OpenRTB and Prebid
Outside of the usage in agentic advertising, which is still in its early stages, Agentic Audiences can add value in the existing programmatic ecosystem, given its privacy-preserving nature. To support this, we have added a community Agentic Audiences extension to OpenRTB. This extension simplifies the original specification by using only the base set of fields needed for bidders to process the embeddings. Those fields are: ID, name, version, vector, dimension, model, and type.
The vector field is the embedding itself; the other fields are used to describe the type of embedding and the model used to generate it. To significantly reduce the size of the embedding as it is sent through the bidstream, the task force decided to encode the vector using base64. The result is a simple yet efficient extension that contains just enough information to process the vector embeddings by upstream bidders.
To further enable adoption and experimentation with Agentic Audiences, we developed a Prebid module that conforms to the OpenRTB extension. Vendors can use this example module to build their own modules or use the module that is already integrated into Prebid and is available here.
Open Source Scorer
One additional artifact that was donated by LiveRamp and developed by the Agentic Audiences Task Force to simplify adoption is an open-source scorer that demonstrates the power of embeddings and how they can be used in programmatic advertising using the new OpenRTB extension. With this tool, implementers can simulate how to:
- process an OpenRTB Agentic Audiences request
- extract the embedding
- compare it against campaign targeting embeddings
- match and score how closely related the incoming embedding is compared to the targeting embeddings
The work to date has been driven by the Tech Lab Agentic Task Force. As adoption increases and the use cases expand, we have decided to launch a separate task force to accelerate the work: the Agentic Audiences Task Force. You can’t afford to be left behind by state-of-the-art agentic workflows, so email membership@iabtechlab.com to sign up.
Help develop the future of agentic audiences in the Agentic Audiences Task Force and start testing Agentic Audiences by reviewing and implementing the specification in GitHub.
Anthony Katsur
CEO
IAB Tech Lab