Ad Tech's New Standard Is a Vector With No Ground Under It

Buying decisions are becoming multi-modal. The modalities have embeddings. The join that would fuse them into one decision is unwritten.

This is the first post of a thread, and the thread has one subject: the AI models that
are not LLMs.

Terms first, because the thread depends on them. AI, as I will use the word, is machinery
that learns a function from data instead of taking its rules from a programmer. Three
jobs cover most of it. Predictive models turn inputs into a judgment, such as the
probability that this impression converts. Embedding models turn things into vectors, so that similar places, people, and products sit near each other in a space a machine can search, and most of them earn those vectors generatively: the model learns by predicting what is missing from its input, and the vector is the understanding it keeps. Generative models produce new instances of what they were trained on outright, and the famous family generates one step at a time, each step conditioned on everything produced so far; the technical term is autoregressive. An LLM is one species in that family, an autoregressive generator whose steps are words. Run the same machinery over a sales history and you get a forecaster. Run it over a user's event sequence and you get what Meta's researchers call a generative recommender. All of it is generative AI. Almost none of it is an LLM.

That distinction is the subject, because when an advertising campaign gets bought and
measured, the deciding is done by the other kinds. A production ad ranker composes 20 to 40 embedding tables per impression; Meta's published architecture carries roughly 30. [9] Spatial models now embed every few square kilometers
of ground. Temporal models forecast by the hour, and population models carry the
denominator under every reach number. LLM agents are arriving to coordinate that
machinery, which sets up the sentence this thread will keep returning to: the agent is a
language model, and the decision is not.

The question I will keep asking is what these models need in order to compose into one decision, and what the industry has to standardize before the open version can compete with the walled gardens. One development at a time, one chart, sources included.

Here is the first one.

In November, the IAB Tech Lab shipped a standard whose roadmap ends with software agents exchanging embeddings instead of text. Agentic Audiences, proposed by LiveRamp, defines the audience as a dense vector of 256 to 1024 dimensions, served to a bidder in under 100 milliseconds. [1] The audience is now officially a vector.

The ground already had its vector. In April, Google moved its Population Dynamics
research model into commercial preview on Maps Platform, delivered straight into
BigQuery: a 330-dimensional embedding for every cell of roughly three to six square
kilometers, refreshed monthly, across 17 countries. [2] DeepMind's AlphaEarth goes finer: a 64-dimensional embedding for every 10-meter square of the planet, 1.4 trillion
embedding footprints a year, and as of this month the archive costs nothing to pull. [3]

The audience has a vector. The ground has a vector. The join is unwritten.

Lay the modalities next to their shipped vectors and one column is empty. The audience standard specifies its dimensions and its latency. It specifies no spatial grain and no time bin. The same gap runs through the IAB's other new workstream: Project Eidos, the campaign-data taxonomy that closed public comment this month, standardizes pod position and ad format and says
nothing about where or when. [4]

The reason the empty column matters is what a buying decision has become. An agent
optimizing a campaign is fusing modalities: the audience, the creative, the page it runs
on, the place, the hour, and increasingly the measurement itself, because the newest
rankers bid on predicted incrementality, a number trained on experiments rather than
clicks. [10] Modalities fuse in a model only when they share a representation, which embeddings now provide, and a ground reference, which nobody has specified: a cell at a declared resolution, and an hour. That key already exists as an open proposal, a hexagonal cell grid crossed with the week and the hour, and it is sitting in the Ad Context Protocol's review queue right now. The queue moved this weekend: the cell proposal was revised with its design questions settled on Sunday, and its reporting companion, filed Thursday, holds the temporal question open. [5] With the key in the standards, the modalities compose into one decision inside the warehouse where
the datasets already live. Without it, every "AI-powered" buy is a mashup: an audience
score stapled to a location report, with no model able to see both at once.

The economics of embedding retrieval are not hypothetical. Meta rebuilt ads retrieval on embeddings and reported an 8 percent ads-quality gain. [6] The Trade Desk credits its AI-scored buying with $5.40 of return per advertising dollar across 665 measured
campaigns in 2025. [7] The platforms internalized the vector and kept it proprietary. The open standard is where the rest of the industry gets the same physics, and ground-truth embeddings are the input that makes the open version competitive. The research side has already run the test: a May study found the population embeddings cut unexplained variance in subnational population estimation by a median of 20 percent, which is the denominator every reach and frequency calculation stands on. [8]

The IAB's own repository states the destination:

"Enabling interoperability from prompt-based LLM coordination to embedding-based optimization." [1]

The rest of this thread is about getting the join written.

How the fused decision works today, inside the walled gardens, is in
How a Modern Ad Ranking System Works, the explainer this thread builds on.

The longer argument for the join key is in Marketers of Catan, and the proposal itself is the 168 Grid, from my work with Ether Data.

[1] IAB Tech Lab, Agentic Audiences v1.0 (released November 2025, renamed from User Context Protocol January 2026): github.com/IABTechLab/agentic-audiences and
iabtechlab.com/standards/agentic-audiences/.

[2] Google Maps Platform, Population Dynamics Insights, April 22, 2026: S2 level-12
cells, 330-dimensional embeddings, monthly updates, 17 countries.

[3] DeepMind, AlphaEarth Foundations: 10 m pixels, more than 1.4 trillion embedding
footprints per year; the 2025 annual layer shipped March 2026 and the Cloud Storage
archive moved to provider-pays (free egress) in July 2026.

[4] IAB, Project Eidos Campaign Data Standards: public comment closed July 6, 2026.

[5] Ad Context Protocol, RFC 5969 (geo_cells targeting; revised July 26, 2026 with
cell-system enum and resolution semantics settled, awaiting working-group review) and RFC 6014 (cell-grain reporting, filed July 24, 2026; whether temporal dimensions are addressed now or deferred is open question 4).

[6] Meta Engineering, Andromeda retrieval engine: +6% recall, +8% ads quality on
targeted segments.

[7] The Trade Desk, Kokai results page; TTD's figure across 665 campaigns, published
with its own disclaimer that results are not guaranteed.

[8] "Geospatial foundation-model embeddings improve population estimation unevenly across space and scale," arXiv:2605.01650 (May 2026): PDFM embeddings, median 20.1% reduction in unexplained variance across Brazil, Nigeria, and the United States.

[9] Naumov et al., Deep Learning Recommendation Model (DLRM), arXiv:1906.00091 and github.com/facebookresearch/dlrm: roughly 30 sparse embedding tables in the published architecture; production rankers compose 20 to 40.

[10] Gordon, Moakler, and Zettelmeyer, "Predicted Incrementality by Experimentation
(PIE) for Ad Measurement," NBER Working Paper 35044, April 2026:
nber.org/papers/w35044.