EPPO is Not Dead: Every Page Is Page One in the Age of Generative AI

Why AI bots are reading your content instead of humans

While we might be cranky that our sites don’t have the same number of human visitors, we still have plenty of robot visitors. There’s a high chance the person who uses our content never makes it to our website. And yet our content is still being used to supply the answer.

As content professionals, we need to worry about what we have always concerned ourselves with: creating high-quality, correct information that is well suited for our readers.

Now, we need to also worry about LLMs hallucinating. If the answer requires combining information from content on multiple pages, the LLM may not retrieve all of the necessary information, even when links are provided. Public LLMs won’t know your content and its nuances and specialized vocabulary, unless they have been specifically trained on it. Remember, LLMs are looking for statistical similarity, both of the user’s query and their intent and of your content. It’s guessing.

Why AI-synthesized answers aren’t the same as human judgment

When a user asks a question, the sequence is no longer: human searches, finds page, reads page, judges the value of the source content as appropriate and correct. It is: human asks, AI retrieves fragments, AI synthesizes answer, human judges AI’s synthesized response.

EPPO and scent disappears. What the AI retrieves are sections, paragraphs, or blocks. There’s nothing inherently incorrect about this approach. What AI cannot do is judge whether those were the correct pieces to retrieve or what might be missing. AI systems can go through a mechanical process of ranking and re-ranking, which may seem like its ‘judging’ the content, but it’s not exactly the same. It did so solely on statistical similarity, and, maybe, some additional training. That’s why AI summaries provide links to its sources, so that you, a human, can read the source material and judge for yourself whether what the AI pulled supported its result.

But similarity is a ranking exercise, not a completeness check. Retrieval can tell you which fragments look most relevant to each other; it can’t tell you whether the candidate set is missing something that was never included in the first place. It doesn’t know what it didn’t retrieve.

From EPPO’s “topic” to the AI era’s “fragment”

We may not always like it, but a human can tolerate decontextualization. We can notice when something is missing or off and fill in the gaps. We can look around for clues. We might look for a sidebar or click “Up.” An AI works with what its given. A fragment that depends on context not present in the retrieved chunk is apt to produce a confident but wrong answer. Vector-based retrieval is probabilistic; it ranks by similarity without certainty. GraphRAG trades some of that flexibility for symbolic structure, and that structure is what makes complex, multi-hop answers reliable rather than lucky. Authors who write EPPO-compliant topics (complete, well-scoped, explicitly contextualized) produce the high-quality training and retrieval data that AI requires.

How to apply EPPO principles as metadata for AI retrieval

The shift to AI-mediated delivery requires a tactical change in how we handle metadata and structure. When AI synthesizes your content, your content authors shouldn’t abandon EPPO’s principles. Your writers should lean into the principles, making them explicit as metadata where they can be leveraged beyond their semantic qualities.

Before LLMs, metadata helped humans browse. In the AI era, metadata helps LLMs normalize or separate similar terms, and gives retrieval systems the chunk boundaries they’d otherwise have to guess at.

A vague title like Configuration — a common title in many user guides — relies on its place in a hierarchy or the surrounding navigation to provide context. That context disappears in an LLM. The expanded title is “specific and limited purpose” title: (OAuth 2.0, this gateway, this version). It’s explicit about the content of its topic rather than implying it covers configuration more broadly.

Metadata that can be implicit or exposed does something similar for “assume the reader is qualified.” This meant writing for the person who normally performs the task, not explaining every term to every possible reader. Experienced and professional technical writers used their judgment and user research to identify who was ‘qualified.’ Now that needs to be made explicitly: user-persona:security-admin, because the thing consuming the tag now has to be trained on your specific content and user personas to infer qualification level from tone or vocabulary the way a human skimming the page could. The metadata makes it explicit for all user types.

Baker wrote about ‘pages,’ which could roughly be considered a ‘topic.’ (We are using ‘topic’ here in a Wikipedia sense of the word, not an information architecture sense.) With LLMs we need to move into the sub-topic, the fragment, and size that fragment deliberately rather than leave it to chance. Every paragraph should be interpretable in isolation. If an AI retrieves a paragraph about “the reset button,” the metadata or the text itself must identify which device that button belongs to, or the AI may hallucinate context from an adjacent but irrelevant product topic.

The EPPO principle to link by subject affinity is a natural precursor to knowledge-graph structures: rich, deliberate links between related topics are the basis for graphs. Building the graph itself, with typed, machine-traversable nodes an AI can use for multi-hop reasoning, is separate work on top of that linking discipline. That work is exactly what GraphRAG benchmarks measure. But it starts from scratch if the underlying content isn’t already linked with intent. Organizations with deliberate subject-affinity linking have a head start that organizations without it don’t. Unstructured corpora return inconsistent answers regardless of how capable the model is.

Not all seven principles port cleanly for LLMs. “Stay on one level” assumes a reader who chooses when to shift between general principles and concrete examples, following a link when they’re ready to change levels. An LLM doesn’t make that choice the way a person does; it retrieves a topic or it doesn’t. This may be the one place EPPO’s authoring guidance needs to be adjusted: level needs to be stated as fragment-level metadata, not left as something the writer trusts the reader to navigate. (And this is where ‘topic’ could be used in the information architecture sense.)

EPPO content is a head start for AI retrieval

Organizations that already practice Every Page Is Page One — well-scoped, self-contained, richly linked topics — are ahead of the curve. They’re well positioned to take advantage of their content as a strategic asset.

If you want to understand how your current content architecture holds up as an AI retrieval source, schedule a call and we can work through what your corpus looks like from the retrieval side.