IAB Tech Lab published a technical position paper on April 9, 2026, arguing that shared taxonomies - not natural language prompts - are the foundational layer keeping autonomous advertising agents from making costly, compounding errors. The post, written by Katie Shell, Associate Product Manager at the organization, frames the argument around a structural problem that grows more acute as multi-agent advertising systems scale. When two AI agents negotiate a media buy without a shared vocabulary, every hop in the supply chain introduces a fresh probability of misinterpretation. Taxonomy IDs, Shell argues, eliminate that probability entirely by replacing inference with a deterministic lookup.

The piece is positioned within IAB Tech Lab's broader "Agentifying Standards" series and connects directly to the AAMP initiative - formally named on February 26, 2026 - which consolidates the organization's agentic advertising work under three pillars: Agentic Foundations, Agentic Protocols, and Trust and Transparency. The Agentic Protocols pillar is where Taxonomy Guardrails sit.

The problem with natural language briefs

Shell opens with a scenario that is immediately recognizable to anyone who has worked in programmatic media. A media buyer says, "I want lifestyle content targeting women 21-45 interested in beauty." A human sales rep hears this and applies judgment. Questions get asked. A reasonable outcome is reached. But when a buyer agent sends that same phrase to a seller agent, neither party is exercising judgment. Both are completing probability distributions.

According to the paper, "LLMs generate outputs as probability distributions over language - they do not look up facts, they predict likely completions." The buyer's model might anchor on ages 30-40 as the default wellness demographic. The seller's model might interpret "wellness" as fitness-focused, routing the ad to running content rather than beauty content. Neither agent is technically wrong from its own internal logic. The outcome - an ad on running content shown to an audience tagged as "women 28-42" rather than the intended placement - is a mismatch that no human explicitly approved.

Critically, this error does not occur once. It compounds. According to the paper, a lipstick brand running a campaign with requirements to avoid alcohol and cannabis content would route its brief through a buyer's agent, a seller's agent, an audience agent, and a publisher's ad platform. That is four hops, each with its own LLM re-processing the context. At each hop, the brief is paraphrased slightly. "No adult content" gets interpreted inconsistently across models. "Lifestyle content" creeps in as a synonym for "beauty content." By the fifth hop, the ad appears on a wine review page. Every agent followed the brief as it understood it.

This is the failure mode that structured taxonomies are designed to prevent.

Three taxonomies, one shared contract

The IAB Tech Lab maintains three taxonomies that standardize inventory description between buyers and sellers. The Content Taxonomy describes the "aboutness" of pages, apps, or videos. The Ad Product Taxonomy labels what is being advertised. The Audience Taxonomy provides a consistent naming convention for segments based on demographic, interest, and purchase-intent attributes.

Together these form, in Shell's phrase, "a shared semantic contract." Each taxonomy operates as a controlled vocabulary of integer IDs that any agent can transmit, receive, and match against without interpretation.

The specificity of the IDs is the point. Consider the Content Taxonomy: Automotive > Auto Type > Green Vehicles carries ID 22. Automotive > Car Culture carries ID 25. Both involve cars. In natural language embedding space, both nodes sit close together. An LLM asked to avoid "fuel-related content" for an electric vehicle brand might correctly block content about gasoline prices, but it might also - probabilistically - suppress content about fuel economy tips, which is relevant to EV buyers. Content Taxonomy ID 22 (Green Vehicles) and a targeted blocklist of petroleum-adjacent IDs make that distinction explicit and auditable.

The Ad Product Taxonomy applies the same logic to what is being sold. Alcohol > Wine carries ID 1007. Alcohol > Beer carries ID 1004. A publisher that wants to allow beer advertising but block wine advertising can do so precisely, with no ambiguity, because the IDs are distinct nodes that can be allowed or blocked independently.

The Audience Taxonomy resolves the vagueness in demographic briefs. "Women in their prime spending years" is not a specification. Demographic > Gender > Female [ID: 49] combined with Demographic > Age Range > 30-34 [ID: 6] is an exact specification. The lipstick campaign from Shell's example would transmit Female [49] plus six consecutive age-range IDs (4 through 8, covering ages 21-44), alongside a blocklist that includes Alcohol [1002], Cannabis [1049], and Adult Products and Services [1001]. Those integers pass through every hop in the supply chain unchanged. The publisher's ad server performs boolean matching - not inference, not interpretation. The wine review page carries Ad Product ID 1002 in the blocklist. The impression is rejected automatically, at every node, consistently.

Determinism versus probability in agentic systems

The distinction between deterministic matching and probabilistic inference is not merely technical. It carries direct implications for campaign performance, brand safety, and measurement accuracy.

Content Taxonomy 3.1, released in December 2024, expanded coverage from approximately 400 categories in version 2.x to more than 1,500 categories. The granularity matters: a finer taxonomy lets AI systems understand content at a contextual level that coarse natural language descriptions cannot reach. In February 2026, IAB Tech Lab received a donated open-source taxonomy mapping tool from Mixpeek that reduces what previously required weeks of manual migration work to seconds, using TF-IDF, BM25, KNN, and LLM re-ranking methods.

Shell's paper also makes a reasoning quality argument that extends beyond pure matching. When an LLM receives a structured, taxonomy-grounded context window, its chain-of-thought reasoning becomes more precise. Unstructured context - "user likes beauty and health content, is a woman in her 30s" - leaves the model free to drift into adjacent categories: health leads to fitness, fitness leads to sports, sports leads to active lifestyle, active lifestyle leads to outdoor gear. Structured context pinning the user to IAB Audience Taxonomy > Demographic > Gender > Female [ID: 49], Demographic > Age Range > 30-34 [ID: 6], and Interest > Style & Fashion > Beauty & Personal Care [ID: 677] constrains what the model is allowed to infer rather than what it finds plausible.

This is why, according to the paper, Retrieval Augmented Generation systems that pre-tag documents with taxonomy labels outperform pure semantic search for precision. The taxonomy acts as retrieval grounding by combining the best of exact-match and semantic reasoning.

Consistent taxonomy labeling also improves measurement. When every participant uses the same content and audience labels, attribution, reach, and frequency calculations are based on the same definitions - a stable shared ground truth that autonomous agents need to optimize campaigns accurately over time.

Where the taxonomy approach fits within AAMP

The paper locates its argument inside the AAMP architecture. AAMP, formally named by IAB Tech Lab on February 26, arranges its architecture in three horizontal bands. Agentic Foundations sit at the base, covering the Agentic Real-Time Framework (ARTF) already released and Agentic Guardrails still in development. Agentic Protocols sit in the middle, covering seven protocol components including Agentic Bid (OpenRTB), Agentic Direct (OpenDirect), Agentic Deals (Deals API), and Agentic Ad Objects (AdCOM). Trust and Transparency sits at the top, incorporating the Agent Registry.

Taxonomy Guardrails occupy a constraint layer within the Agentic Protocols band. As PPC Land documented when AAMP was named, both Buyer Agent and Seller Agent SDKs follow a three-level hierarchy - Level 1 Orchestration (Portfolio Manager), Level 2 Channel Specialists, and Level 3 Functional Agents - all governed by Taxonomy Guardrails. The guardrail layer ensures that agents at any level use standardized vocabularies rather than platform-proprietary definitions.

Shell's paper describes the guardrail function at the transaction level. The buyer agent transmits integer arrays. The seller's ad server does boolean matching. No LLM at any node in the chain is asked to interpret what "alcohol-related content" means. The blocklist ID is there. The page carries or doesn't carry that ID. The impression is served or rejected. The paper calls this "deterministic interoperability."

The stakes are explicitly framed in terms of error compounding. According to the paper, "a misinterpreted brief between two humans is a recoverable mistake. A misinterpreted brief compounding across five autonomous agents, running thousands of times per second, may not be recoverable and cause permanent harm." The scale of agentic systems - transactions executing at machine speed, without humans in the loop - transforms what would be a recoverable error in a manual workflow into a systematic failure mode.

Taxonomy limitations and honest gaps

Shell's paper does not overstate the case. Three categories of limitation are acknowledged directly.

Taxonomies update slowly relative to how fast new content formats and product categories emerge. "AI-generated video content" and "creator economy" content have limited taxonomy coverage today. The pace of content format change is structurally faster than the pace of taxonomy revision, which operates on industry working group timelines.

Self-tagging creates incentive problems the industry has not fully addressed. A publisher labeling its own content has commercial incentives that can produce inaccurate classification. The paper notes that verification standards for tagging accuracy remain an unresolved issue.

Soft targeting preferences - "premium editorial feel" or "brand-safe but culturally relevant" - remain more expressible in natural language and LLM semantic reasoning than in any fixed taxonomy. The paper is explicit that these limitations argue for better taxonomies, better tagging practices, and hybrid architectures that use taxonomy IDs for hard constraints while allowing LLM reasoning for discovery, enrichment, and novel categories.

The paper does not present the limitations as an argument for abandoning structured classification in favor of natural language. It presents them as an argument for improving the taxonomies themselves.

Context for the marketing industry

The taxonomy argument lands in a specific industry moment. IAB Tech Lab CEO Anthony Katsur warned on December 29, 2025, that the industry was racing toward agentic AI without resolving foundational infrastructure questions, including supply chain transparency, privacy compliance, and measurement frameworks. The taxonomy guardrails paper is a direct response to one dimension of that concern: semantic ambiguity at the agent-to-agent negotiation layer.

Kochava opened its StationOne platform to public beta on March 25, 2026, providing the first accessible sandbox where buyers, sellers, agencies, and ad tech specialists can experiment with AAMP-compliant agentic workflows. The platform exposes 19 specialized skills organized across 8 functional areas, all running through the official IAB Tech Lab reference implementation MCP Server. This is the environment in which taxonomy-based guardrails would function in practice.

The IAB Tech Lab Agent Registry, which reached 10 participants in early March 2026, uses taxonomy category selection as part of its registration form - a structural reinforcement of the paper's argument that taxonomy classification is load-bearing infrastructure rather than an optional annotation layer.

The open-source AI mapper donated to IAB Tech Lab by Mixpeek in February 2026 addressed one practical barrier to taxonomy adoption: migration cost. Any SSP, DSP, publisher, or brand safety vendor can now map their IAB Content Taxonomy 2.x labels to version 3.1, gaining deterministic, confidence-scored mappings without external API access or cloud processing.

IAB Tech Lab's Privacy Taxonomy, launched for public comment in September 2024, extends the taxonomy approach into data governance, giving models a common language for understanding what data is sensitive, what uses are restricted, and what should be off-limits.

Shell closes the paper with an argument about the direction of model capability. As LLMs become more capable, the paper does not argue they will outgrow the need for shared taxonomies. It argues the opposite: according to the paper, "as models grow more capable, they do not outgrow the need for shared structure. They become more dependent on it." The reasoning is that more capable models executing more autonomous and consequential actions at greater scale require a more reliable semantic contract, not a less constrained one.

For the advertising industry, this means the investment case for taxonomy adoption - and for improving taxonomy quality, coverage, and verification standards - grows as agentic systems expand into media planning, buying, optimization, and reporting functions that currently require human oversight at each step.

Timeline

Summary

Who: Katie Shell, Associate Product Manager at IAB Tech Lab, authored the paper. IAB Tech Lab is a global non-profit technical standards consortium established in 2014.

What: Shell published a technical position paper arguing that IAB Tech Lab's three taxonomies - the Content Taxonomy, Ad Product Taxonomy, and Audience Taxonomy - serve as the essential semantic contract enabling deterministic interoperability between autonomous advertising agents. The paper explains how integer taxonomy IDs replace probabilistic language interpretation in agent-to-agent transactions, preventing brief misinterpretation from compounding across multi-hop supply chains at machine speed. The paper also acknowledges taxonomy limitations including slow update cycles, self-tagging incentive problems, and the inability to express soft targeting preferences.

When: The paper was published on April 9, 2026, as part of IAB Tech Lab's ongoing "Agentifying Standards" content series, following the formal naming of AAMP on February 26, 2026.

Where: Published on the IAB Tech Lab website. The taxonomies and AAMP protocols described are intended to function globally across all participants in the programmatic advertising ecosystem - buyers, sellers, agencies, publishers, SSPs, DSPs, and brand safety vendors.

Why: As advertising systems shift from human-supervised execution to autonomous agent-to-agent transactions running at machine speed, semantic ambiguity in campaign briefs transforms from a recoverable human error into a systematic failure mode. Taxonomy IDs provide a shared, machine-readable vocabulary that makes interoperability possible between agents trained by different companies on different data, preventing compounding misinterpretation errors that no single participant in the supply chain intended or approved.

Share this article
The link has been copied!