Reversing the Death of Hypertext: Mechanistic Interpretability for LLM Navigation

The rise of large language models (LLMs) has introduced a new mode of Web interaction called conversational navigation, in which autonomous agents retrieve and synthesize information into a single conversational response. While this offers significant efficiency gains, it conflicts with the architectural principles that originally structured the Web: visible links, discrete documents, user-directed traversal, and inspectable provenance. In doing so, these systems internalize processes that were once visible, obscuring from the user the very epistemic structure the Web was designed to preserve. Existing attempts to address this tension treat transparency as a surface problem, appending citations without reconsidering how LLMs relate to the Web's underlying structure. We propose mechanistic interpretability (MI) as a structural framework for realigning generative systems with the Web's epistemic commitments. Specifically, we analyze three stages at which MI techniques can render generative systems more inspectable: document retrieval, document ranking and integration, and provenance mapping from outputs to sources. Although MI does not restore traditional hypertext traversal, it can reintroduce traceability, visible grounding, and meaningful user agency within the generative paradigm.