AI has become one of the heaviest consumers of your content, yet most organizations are entirely unaware of their inability to deliver it in a way that the machines can consume it effectively. Search bots, AI crawlers, and autonomous agents are reading your content every day, and, despite what you may think, they don’t care about your beautiful layouts and clever headlines. They care about the raw text and the limited context or structure that you provide alongside it. If there is no structure to guide them or, worse still, it is misleading, then AI has to rely on guesses to get to the right answer about your business.
Machines read differently than humans
when confronted with visual information such as pages or even paragraphs. While a human eye can recognize elements on the page and make assumptions based on their relative location and appearance, an LLM is not capable of performing these tasks. To recognize that the bold text at the top of the page is the price of the product, it has to have the word “Price” somewhere attached to it along with the word “Product.”
This is why most companies are failing when it comes to web crawlers, RAG pipelines, and agents. These are the new readers of your content, and they are fundamentally different from human readers your web page was designed for. They do not scroll down the page looking for answers; they copy-paste blocks of text somewhere else, run similarity checks, and tune the response of a language model using the retrieved content. Source, retrieval, response – each of these steps penalizes ambiguity and rewards clarity.
While companies that optimized their content for people have succeeded, they failed to realize that the same page would also be read by somebody else. This is not a hypothetical scenario, either; we have seen incorrect pricing, expirations, and even made-up product specs appearing in the responses of LLMs representing real-world companies.
Hallucinations are usually a data problem
There is a common belief among people trying to battle hallucinations that it is primarily a model problem, one that is expected to be resolved as models get better. While it is not untrue – there certainly is a model component – it is significantly more useful to recognize the data component.
According to the hallucination leaderboard maintained by Vectara and benchmarked against major LLMs, the fabrications made up by models span from roughly 3% to 27%, depending on the model. This is a significant range, and it demonstrates that even the best models are still fairly likely to make up information when they cannot confidently retrieve it. When source text is ambiguous, conflicting, or unavailable, models are much more likely to guess at it, and guess incorrectly, often producing answers that sound impressively certain.
The reason this matters so much for data quality is that most production-grade systems today do not rely on training models to know everything. Instead, they employ retrieval augmented generation (RAG) that uses separate tools to retrieve relevant documents from a database before generation. Grounded generation, as this practice is known, has far fewer hallucinations than zero-shot generation relying on parametric knowledge because models can pull factual information from the retrieved documents. However, this only works when there is factual information to pull, and the more structured the better. By feeding a RAG pipeline unstructured or conflicting information, companies regularly produce grounded but utterly wrong answers.
Why RAG makes your content the real bottleneck
RAG has become the architecture of choice for most enterprise-grade applications, from chatbots and AI copilots to the AI summaries appearing above search results. It is difficult to overestimate the importance of this technology to the future of the internet. While the nuances of the generator part of RAG are often discussed, the mechanics of retrieval are far more interesting. Quality of retrieval depends on the quality of embeddings, essentially mathematical representations of your text chunks, and those, in turn, mostly depend on the quality of the source text.
Well-structured text produces embeddings with desirable properties, such as clustering. This allows similar text chunks to be grouped together and facilitates the process of retrieving relevant documents. On the other hand, text with unnecessary repetition or contradictions will produce faulty embeddings that trick the model into finding irrelevant passages.
It is worth mentioning that long texts are also detrimental to the performance of LLMs. Structured retrieval systems solve this problem by allowing the model to focus on the most relevant and well-structured text chunks while avoiding the attention of generic, untargeted, and often misleading information. The more organized your knowledgebase is, the more relevant information it can provide, and the less irrelevant information it can avoid providing, making retrieval a core part of knowledge management.
None of this has much to do with improving the LLMs themselves. It has everything to do with structuring the text that they process.
Structured data is the label set machines actually need
This is where Schema.org markup fits in as the ideal solution to the problem. The specification, which is implemented as JSON-LD these days, defines a set of tags and labels that can be added to HTML to help search engines understand what the page is about. JSON-LD itself serves as infrastructure-level markup that enables most of the rich results and featured snippets that search engines display. It is not surprising that most people who have heard of structured data associate it with SEO, as it has long been used to communicate with search engines.
What few people realize is that it has enormous potential to facilitate interaction with other machine readers of your content. Search engines are just one use case among many, and while they may be the easiest to reason about, they are not necessarily the most important. The same structured labeling that powers rich results can be consumed by a RAG pipeline to extract relevant information or used by an autonomous agent to perform specific tasks. For example, if an agent needs to locate a return policy, the ability to recognize a page as a Pledge and extract its relevant information will make it exponentially easier to find compared to searching through unstructured text.
Entity extraction is another critical area where structured data provides massive benefits. By assisting in the identification of relevant concepts within a text, structured markup helps extract entities and populate a knowledge graph. A knowledge graph, in turn, can help organize information in relation to other entities on the page, facilitating tasks such as finding similar items. If a specific item has a return policy that differs from the standard practice, then a properly structured page can allow the knowledge graph to recognize those exceptions and facilitate their identification.
Unstructured archives hit a ceiling as they grow
Any knowledge repository hits a point where it stops scaling efficiently. This slowdown manifests in several ways – pages with redundant information start appearing, old articles contradict newer ones, orphaned pages with no connections to the rest of the archive appear, or, perhaps, nobody maintains the pages at all. People encounter these issues first when they attempt to use the knowledge base to research a particular topic. They come across old information and either try searching for newer alternatives or give up frustrated.
This is not as severe of a problem as it may seem – a human researcher is capable of identifying differences and determining whether the information that they found was correct. However, the same cannot be said for a machine. A system is unlikely to realize that certain information is out-of-date unless it is explicitly told to look for a publication date and, if instructed, identify conflicting pieces of information. Instead, it is likely to utilize the page with the most similar text and no apparent contradictions as the most relevant search result and attempt to use it as a source of truth, potentially presenting it as such.
While an unorganized archive may be annoying for a human to navigate, for a machine it becomes an overwhelming deluge of unmanageable information. What constituted a useful knowledge base for a human user suddenly becomes a source of tremendous amounts of misinformation for an AI consumer.
Turning documentation into infrastructure
Unfortunately, there is no simple solution to most of these problems. Mostly, it boils down to curating all of the information scattered across your manuals, product documentation, knowledge repositories, and other resources to a singular source accessible to both your customers and your AI. The only way to do that is to build an infrastructure around knowledge management that supports the efforts of both. The solution will center around the creation of a central knowledge repository that structures the information that it contains while educating the company’s AI and systems about that structure.
This is the essence of what the knowledge management field is trying to achieve – turning the disparate efforts of various departments into a single coherent system that utilizes the best practices of knowledge organization while providing support to the widest variety of tools. The work is extremely multifaceted, spanning taxonomy and ontology design, content governance and curation, as well as the technical aspects of the knowledge repository itself. These aspects reinforce each other, however, creating a system around standardized information that is accessible and consumable by many different parties. Taxonomy and ontology work are far more important than most people realize – a simple unordered list of articles is not much better than having no system at all, while controlled vocabularies enable consistent tagging that is critical for large-scale knowledge management. Similarly, content governance is what ultimately prevents a free-for-all situation in a knowledge repository when it grows beyond a small size.
Agents are the next machine customer
Another development that is quickly approaching maturity and should be considered a high priority is agents. These are essentially autonomous or semi-autonomous programs that interact directly with your business systems to perform specific tasks, such as checking a return status or completing a multi-step transaction. They are not significantly different from a human interacting with your website, except for the fact that they cannot read web pages like a human would. An agent checking a return confirmation generally does not care about the pretty prose that is meant to reassure the customer; it needs a machine-readable, structured block of text, or, even better, an API.
A properly structured knowledge repository with a well-defined set of taxonomic labels essentially acts as an API endpoint for agents; one that is far more reliable and consistent than most endpoints that companies expose to the public. An unstructured KB, on the other hand, will fail to provide the agent with meaningful information, and the agent will fail to complete its task. If the task the agent was attempting to complete relied on the information that was provided, then your company has just trained its machine learning models on erroneous data.
This is, in many ways, the reality of B2B and Customer Service spaces for the foreseeable future. The companies that embrace the opportunity to provide structured information for these agents will be the ones to shape the future of their respective fields.
The shift nobody’s naming yet
For years, the competitive environment has been dominated by content; quantity and quality, broader coverage, better optimization – all of these have been crucial to determine the superior player in the space. This does not change entirely, although it certainly shifts, as the next frontier is about fueling the next wave of technologies. Semantic search, RAG, and various graph and agent-based innovations have one critical weakness that is unlikely to be addressed directly – the accuracy and structure of the information that is fed to these systems.
Semantic search, RAG, knowledge graphs, agentic AI – they all share one critical link in the chain that determines their ultimate success or failure. Companies that want to succeed in this environment need to be aware of it and optimize their knowledge management practices to take advantage of the situation. Get it right, and your company will appear as the reliable source in every AI summary, recommendation, or search result. Get it wrong, and your brand will inevitably be associated with hallucinated information with little potential to correct it.
Comments
