Editor’s note: Today’s guest post is by Veronica Showers, Heather Kotula, and Marjoirie Hlava from Access Innovations.

For years, scholarly publishers have quietly powered the world’s scientific progress. You published the research that helped design the mRNA vaccine platform. You host the clinical guidelines that determine how physicians treat disease. You steward the discoveries that shape global policy and scientific innovation. Yet for all that influence, publishers were largely spectators during the early wave of generative AI.

In the last two years, a quiet shift has transformed the relationship between scholarly publishers and the world’s most powerful AI companies. Large language models (e.g., OpenAI, Google, Anthropic, etc.), and a growing ecosystem of DSLMs (Domain Specific Language Models), and RAG vendors are coming to publishers asking to license their content. And it reveals something important:

These models can write poems. They can summarize Wikipedia. They can generate marketing material. But they cannot reliably answer expert medical questions, interpret advanced scientific concepts, or synthesize field-specific knowledge. They often struggle with nuance, precision, and terminology. They can create connections to fill gaps when they have incomplete or missing information.

This absence of meaningful and accurate information is becoming a problem for AI companies.

Despite their extraordinary fluency, LLM-related services have only used open-access content and have not had meaningful access to peer-reviewed, paywalled scholarly literature. Studies of LLM training sets confirm that the bulk of model training comes from sources crawled (e.g., Common Crawl), such as public books, free datasets, and open-source text.

This absence is becoming a problem for AI companies. Their systems produce responses that feel magical to the user, but they lack the conceptual rigor, domain specificity, and factual grounding that scientific content provides. To move into professional markets (such as medicine, clinical decision-support, engineering tools, policy systems, environmental modeling, and others), they need curated, authoritative content. And that content resides almost exclusively with scholarly publishers.

Simply holding the content is no longer the differentiator. AI companies don’t need more text. They need meaning and interpretation of results. They need the conceptual scaffolding behind the text, specifically the relationships, hierarchies, parent/child concepts, and synonyms that organize domain knowledge. XML alone cannot express these. XML remains a durable success: a structural backbone that standardized production and ensured preservation. But it describes form, not meaning.

This is where semantic enrichment becomes transformative.

The New Centrality of Meaning

For years, publishers treated semantic enrichment as an add-on, something that improved search, browsing, and recommendations. It was good for SEO. It made curated subject collections possible and easy to create. It allowed users to navigate a semantically structured landscape rather than wander through a maze of full-text search results.

But in the age of generative AI, semantic enrichment does something far more fundamental: it turns unstructured text into organized knowledge.

Semantic enrichment exposes the conceptual architecture within scholarly content. It shows the relationships of content to other content. Ontologies (taxonomies, etc.) tell machines not just what the words are, but what they signify. It resolves ambiguity (“Java” as a language vs. island vs. slang for coffee), links related concepts (melanoma → skin cancer), unifies variant terminology, and preserves hierarchical relationships (Birds → Chickens → Leghorns). It transforms an article from text into a navigable knowledge asset.

This isn’t theoretical. Multiple recent studies now provide clear empirical evidence that semantic or ontology-based layers directly improve RAG (retrieval-augmented generation) performance, including those used along with an MCP (Model Context Protocol) server.

The  OG-RAG architecture of Sharma et al., grounded retrieval in a domain ontology and saw higher factual recall and improved correctness compared with standard RAG across four LLMs. Zhong et al., showed that combining semantic chunking with a knowledge graph “significantly improves retrieval relevance and correctness” versus traditional, text-only RAG. Tiwari’s OntoRAG outperformed both baseline RAG and GraphRAG in comprehensiveness and diversity of answers.  Sun’s environmental QA research states plainly: “using ontology in RAG systems introduces a structured way of organizing information that can significantly enhance the retrieval process” and yield more accurate, context-aware outputs.

Major vendors echo this. Amazon Web Services writes directly that “semantic search enhances RAG results” and improves the quality of context fed to the LLM. Microsoft’s GraphRAG documentation reports “substantial improvements” in question-answering performance versus naïve semantic search over plain-text snippets.

The message is consistent: RAG systems that incorporate structured meaning consistently outperform those that rely solely on unstructured text.

For publishers, this means that enriched content is not only more valuable, but it is also more compatible with how modern AI systems actually work.

Why Enrichment Matters for Licensing and Strategic Value

When AI companies license raw text from publishers, they must still extract entities, deduplicate terminology, resolve ambiguities, build their own ontologies, and maintain conceptual consistency over time. This engineering effort is expensive, ongoing, and error-prone.

When publishers supply content that is preprocessed, specifically, intelligently chunked into coherent, meaningful units (such as overlapping sentences and paragraphs) and semantically enriched at the chunk level (creating a semantic layer), much of that work is already done, and more importantly, it is done by the people who own and understand the knowledge domain. This semantic grounding enables retrieval systems and language models to select appropriate content, rather than merely textually similar passages, and to interpret it within the correct disciplinary framework. Publishers retain control of the organizing logic, the preferred term usage, and the conceptual relationships that define their field. That semantic layer then becomes a defensible intellectual property asset.

Equally as important, semantic enrichment reduces misinterpretation. It creates a conceptual guardrail, a guide to meaning.

For AI companies, that distinction is critical. Raw text is expensive to process. It requires massive engineering efforts to clean, normalize, classify, and disambiguate. When content arrives already enriched, the cost of integration drops, accuracy rises, time-to-market shrinks, and the system’s overall reliability improves.

This is why enriched content is more valuable in practice, even if the license price isn’t explicitly “per tag” or “per concept.” It is not that AI companies will always pay more simply because tags exist. Rather, enrichment strengthens your negotiating position. It turns your content from “just more text” into a knowledge product. It gives you the ability to offer not just content dumps but APIs, knowledge services, RAG-ready feeds, and domain-specific assistants built on top of your own structure.

Additionally, LM developers do not think like publishers. Publishers think in terms of articles and prefer to license years’ worth of information via subscription. LMs think in terms of tokens and small data chunks (sentences to paragraphs). Licensing of intellectual property needs to be based on both points of view.  As such, adding semantic structures increases the licensing options available to publishers.

This is more than an operational advantage. It creates new monetizable components: text, metadata, ontologies, and semantic tags. These can be priced and licensed separately. Analysts studying AI training ecosystems have noted that structured, rights-cleared data and metadata are emerging as distinct economic assets. Semantic enrichment turns that possibility into a publisher-owned reality.

Publishers as Active Participants in AI Integration

Publishers are no longer engaging with AI solely as content licensors. Increasingly, they are adopting emerging integration frameworks, e.g., Model Context Protocols (MCPs), to connect their content directly to AI systems, retrieval layers, and agentic workflows in controlled and auditable ways. An MCP is a standardized way for AI models to securely access external information sources, such as publisher content, databases, APIs, or retrieval systems, at the moment a question is asked, rather than relying solely on what the model learned during training.

As publishers explore MCP-enabled integrations, a critical requirement becomes clear: AI systems can perform only as well as the structure and grounding of the content to which they are given access. MCPs do not interpret raw documents; they broker access to information that must already be prepared in a form language models can reliably consume. This shifts responsibility upstream to the publisher to ensure that content is genuinely AI-ready before it is exposed through these integration layers.

In this context, preprocessing is a requirement. Raw full-text XML or PDFs, even when structurally well-formed, are insufficient for AI use without preprocessing. To function effectively within MCP- and RAG-based workflows, content must be intelligently chunked. Equally important is semantic tagging at the chunk level. Applying ontologies ensures that each content unit is explicitly anchored to authoritative concepts, domains, and relationships.  Creating this workflow: Content to Preprocessing to Vector Database to RAG to MCP to LM, shown in the figure below.

When Semantic Enrichment Isn’t Used

Despite its importance, semantic enrichment is not universally used. Vector-only RAG pipelines are fast and convenient for vendors, which is why many still rely on them. They retrieve passages based on surface similarity rather than conceptual correctness, often misinterpret domain-specific terminology, and are prone to hallucinating. When content is licensed solely for training or fine-tuning, structure is ignored entirely: the model consumes raw text rather than meaning. And because AI companies must merge information across medicine, engineering, finance, and law, they often avoid adopting any single publisher’s ontology for fear of constraining their flexibility. The result is that vendors rebuild their own internal view of your discipline, one that may be incomplete or inaccurate, yet it becomes the version millions of users experience. In these systems, it is their model, not your editorial expertise, that defines the meaning of your field.

This is precisely why semantic enrichment remains foundational for publishers, even if vendors do not directly ingest it. Enriched content is cleaner, more consistent, and far harder for AI systems to misinterpret. It preserves the logic and relationships your discipline depends on, stabilizes terminology, and ensures that, as RAG evolves toward knowledge-aware architectures, the conceptual map the AI relies on is anchored in the authoritative version you provide rather than the one a model guesses at. In a world where AI increasingly mediates how knowledge is discovered, interpreted, and trusted, safeguarding the meaning of the discipline is critical. It is the publisher’s most important responsibility.

A Future Built on Meaning

We are entering a period where AI will mediate more and more research discovery, clinical practice, regulatory analysis, and public understanding. In this future, the difference between content that machines can read and content that machines can understand will determine which publishers thrive and which fade into the background.

Semantic enrichment elevates content into a knowledge model, and that knowledge becomes the foundation of search, discovery, licensing, and AI applications. Publishers are no longer licensing documents; they are licensing the conceptual structure of their field. In today’s world, semantic enrichment unlocks the future.

We have entered a moment when publishers must stop thinking of themselves as providers of articles and start recognizing themselves as stewards of knowledge. You are not licensing text. You are licensing meaning. The structure you apply today determines how your content will be understood, retrieved, reasoned over, and valued in every AI system that comes next.

Your content was never trained into the models, and that is your advantage. Because now, as the industry is finally getting ready for it, you have the opportunity to shape how your knowledge enters the next generation of technology. With semantic enrichment, you don’t just prepare your content for AI. You make it indispensable to AI.

That is the kind of value no model can manufacture on its own.