Designing Document Links for Humans and Machines: Getting It Right

So what? The way regulatory and medical writers design document hyperlinks now influences not only how efficiently human reviewers read but also how effectively AI tools, like FDA’s Elsa, parse, cross- reference, and retrieve content. Link design is no longer a simple way to manage information economy as defined within lean writing precepts. Link design directly shapes comprehension for humans and interpretability for machines.
​
The Human Reader Challenge
Many biopharma document users must read under time pressure. They work within and across long documents with dense appendices, moving between summary sections, data, methods, and supporting detail. Poorly structured links force them to jump back and forth with little context. That jump is not harmless.

Cognitive psychology highlights two distinct challenges:

• Split attention effect. Readers must hold partial, spatially separated information in working memory while navigating elsewhere for the missing pieces. The extra load strains memory capacity, leaving less mental energy for comprehension.
• Misdirected attention effect. In-text hyperlinks require constant micro-decisions about whether to click (what I refer to as the “should I stay or go now” phenomenon), increasing cognitive load and diverting attention from the main text. Poorly conceived links send readers into irrelevant or low-value detail. When they return, context is lost and momentum broken. Attention has been wasted on material that does not advance the argument.

For example:

• A phrase such as “see appendix for details” without previewing what is in that appendix increases cognitive load.
• A chain of multiple links (from main text → appendix → sub-appendix → external document) fragments attention further.
• A sentence that presents three or more hyperlinks in sequence—such as “refer to 5.3.2, 6.4.3.1, 8.4.5.3”—forces readers to scatter attention across several targets at once, with no clear priority. The reader must also decide: Do I need to check all three? In what order? After reviewing one section, do I return to the original sentence before moving to the next? Each of these decisions adds unnecessary cognitive load and fragments comprehension.

Link formatting is not neutral—it influences what stands out and what gets lost. The Von Restorff effect shows that distinctive items are more likely to be noticed and remembered. Purposeful formatting of critical links can therefore guide reviewers’ attention to essential evidence. But overuse or inconsistent styling dilutes the effect, turning the page into visual noise. Instead of helping reviewers prioritize what matters, the links compete for attention and distract from the argument.

The result: reviewers not only lose the argument thread but also misinterpret the evidence when links overshoot or underserve their purpose. Poor link design magnifies two distinct risks—split attention and misdirected attention—each undermining comprehension in different ways.

The AI Reader Challenge
Hyperlinks can both improve and challenge Natural Language Processing (NLP) parsing in large technical documents by providing additional context and structure while also introducing noise and formatting complexities. For modern Large Language Models (LLMs), hyperlinks can be invaluable resources to enhance the quality of text-based applications like Retrieval-Augmented Generation (RAG).

AI tools such as FDA’s Elsa approach links differently. Machines do not skim, infer, or guess. They parse hierarchies and rely on structure. A vague cross-reference like “see above” or “refer to Appendix 1” leaves a machine with no anchor point.
For AI:

• Consistency matters. A link must follow a stable format—such as “Appendix 3.2.S.4.1 Dissolution Data”—not shorthand like “see Table in Appendix.”
• Persistence matters. Anchors must point to stable IDs or tags, not text that may shift during drafting.
• Context matters. Machines need metadata around the link to understand relationships. For example, is the link pointing to supporting evidence, regulatory precedent, or comparative data?

The anchor text of a hyperlink—the clickable words—often provides a concise, semantically meaningful summary of the linked content. An NLP model can use this information to better understand the linked document’s topic and relevance. Internal links that connect different sections of the same document can act as a roadmap, informing NLP models of the document’s structure, similar to how a table of contents functions.

But not all anchor text is descriptive. Hyperlinks with generic text like “refer to Section 6.4.2” or “see Appendix 1” introduce noise for NLP systems, as they provide no semantic information. Another factor is technical documents in various formats, such as PDFs or legacy file types, may lack consistently marked- up hyperlinks—posing a major challenge for accurate hyperlink extraction.

Without consistent structure and metadata, AI parsing has constrained value. Tools may mis-index or
mis-categorize evidence, leading to gaps in automated review or flawed analytics.

The Dual Design Challenge
The challenge for regulatory writers is designing links that serve two audiences at once.
Human-centered design:

• Preview what the reader will find on the other side of the link.
• Integrate the link into the sentence logically, so readers don’t lose context.
• Minimize unnecessary toggling by including summaries or excerpts before the link.

Machine-centered design:

• Use structured patterns—consistent numbering, explicit section identifiers.
• Anchor to stable, persistent IDs rather than vague references.
• Provide metadata or descriptive labels that help AI categorize the relationship.

These principles are complementary, not competing. Links designed well reduce cognitive load for
people and improve interpretability for machines.

Why This Matters Now
Two shifts make link design urgent:

• Human workload is rising. Regulatory reviewers face expanding data volumes. Poor navigation multiplies their effort and increases the risk of oversight.
• AI oversight is accelerating. Tools like Elsa are entering mainstream regulatory review. Documents that lack structured, machine-readable links may slow automated checks or create mistrust in sponsor submissions.

In short, link design is a risk management decision. If links distract the human or confuse the machine,
the regulatory argument weakens.

Design Principles Going Forward

• Preview before you link. Give readers context so they know why they are leaving the page. This reduces split attention by keeping meaning in view and prevents misdirected attention by signaling whether the link is relevant.
  
• Reduce toggling. Summarize supporting evidence inline before directing to an appendix. This keeps working memory free (limiting split attention) and ensures only essential links are followed (limiting misdirected attention).
  
• Think metadata. Add descriptive tags or labels clarifying the link’s purpose—whether it supports evidence, methods, or regulatory precedent. This helps AI parse relationships and signals to human readers whether the detail is worth following.
  
• Audit your links. Review for vague references, irrelevant detail, and inconsistent styles. Each unchecked problem adds either memory strain (split attention) or wasted effort (misdirected attention).

Bottom Line
Future-ready regulatory documents must support two modes of reading: fast, context-seeking human
review and precise, structure-dependent AI parsing. Writers who treat links as part of information
design—not just formatting—reduce cognitive load for reviewers today and build trust with AI systems
tomorrow.

The real question is not whether your documents contain links. The real question is: Are your links
designed for humans and machine?

https://www.linkedin.com/pulse/designing-document-links-humans-machines-getting-right-gregory-
cuppan-veeec