Large Language Model (LLM) dialogue is rapidly displacing the blue‑link lists that once defined web navigation. While LLM with its fluent answers delights users, this shift affects readers’ agency concealing the associative links and orienting cues that the hypertext community has spent eight decades refining. This paper asks:What must be reclaimed, and what new affordances are possible, when navigation is mediated by a conversational model?We revisit seminal systems from Bush’sMemextoIntermediaandStoryspaceto surface five core principles (associative linking, agency, information scent, non‑linearity, maps) and trace how each is strained or obscured in single‑pane chat. By framing the problem through human-data interaction, we articulate three design obligations—legibility, agency, and negotiability—and demonstrate how emerging techniques such as evidence cards, trail-map overlays, diversity sliders, and on‑devicetiny LLMscan move us towards reclaiming user agency in navigating the web. We then outline a research agenda that ranges from authoring grammars for LLM‑mediated hypertexts to ethical navigation standards that curb bias and filter bubbles. Near‑term tweaks are actionable today; longer‑term questions chart a collaborative path for hypertext, HCI, and AI researchers. Combining classic hypertext insights with modern LLM capabilities, we aim to outline a road map for conversational interfaces that preserve critical reading and empower users to see (and steer) the trails behind every answer.

1

Over the last few years, question‑answering interfaces built on Large Language Models (LLMs) have quietly but decisively altered the texture of everyday web use. A typical interaction that once began with a search box, continued through a cascade of blue links, and ended after several pages of skimming, now often begins and concludes inside a single conversational pane [42]. A reader poses a natural‑language query, the system retrieves candidate documents, ranks passages, blends excerpts, and returns a fluent synthesis that seems to anticipate the next question before it is asked [10,32]. The familiar link list is still available, yet it sits below the fold and is consulted only when curiosity or doubt overcomes the persuasive completeness of the summary [24]. The speed and convenience of this arrangement are hard to overstate. The summaries produced by models such as GPT4 or Gemini routinely translate technical jargon, reconcile conflicting terminology, and offer short follow‑up suggestions [44]. For routine tasks, such as confirming a regulation’s threshold, drafting an email in a non‑native language, or diagnosing a configuration error, the conversational answer feels less like a search result and more like a brief meeting with an informed colleague. But the same brevity that delights also conceals. When readers no longer see the sequence of pages that underwrite a claim they must take the model’s internal judgment on trust; when they skip the act of choosing which link to open they forfeit the small but significant interpretive act that choice affords [31]. We argue that LLM-based conversational interfaces must reincorporate hypertext’s visible trails and user choice to preserve critical reading online. The hypertext community has historically treated that interpretive act as central [2]. Systems from Bush’sMemexvision through Engelbart’s oN‑Line System and forward into the early Web were designed to surface structure, not hide it. Bush articulated the idea of visible trails through information [11]. Engelbart implemented addressable cross‑document links and reversible jumps, embedding choice into every navigation step [16,17]. Later platforms such as Intermedia and Storyspace refined these ideas with typed connectors, overview diagrams, and guard fields designed to foreground branching possibilities [7,25,35]. Viewed across decades, these systems converge on principles that still underwrite effective navigation.Associative linksexternalize conceptual connections, letting readers traverse ideas rather than hierarchies.Reader agencyguarantees that following a link remains a choice, not a side effect of a ranking algorithm.Non-linear traversalopens multiple interpretive paths, encouraging comparison and serendipitous discovery. Finally,information scent—the advance cues encoded in the link text, anchor context, or overview maps—helps readers assess the value of a destination before committing attention [30,40]. Together, these principles form a design vocabulary against which modern conversational interfaces can be assessed. Conversational interfaces disturb these ideas at every step. The links that once materialized conceptual association are computed afresh for each prompt and are rarely displayed. The decision to open or ignore a connection is replaced by the model’s choice of which evidence to weave into its answer. Potentially fruitful detours remain hidden in branching probability distributions that never reach the screen, and the subtle signs that once helped readers decide where to allocate attention are flattened into brief citation stubs or, in some interfaces, omitted altogether [41]. A growing body of empirical work documents the behavioral consequences. Log studies of voice assistants reveal that most sessions end after the first spoken reply, even when alternative documents contain conflicting information [24,45]. Laboratory eye‑tracking shows that users who read a human-generated summary at the top of a results page inspect fewer of the underlying links and spend less time weighing their relative credibility [33]. Interview studies with knowledge workers indicate that confidence in chat answers is high, yet recall of original sources is markedly lower than among users who employ traditional search [13,31]. The question, then is not whether conversational access is useful—it plainly is—but how its benefits can be reconciled with the longstanding hypertext commitment to transparency, choice, and critical comparison. One “analytic lens” for this reconciliation is the Human–Data Interaction framework [37], which argues that data‑driven systems should be legible so that users can see and understand the chain of reasoning; should preserve agency so that users retain the power to shape or redirect that reasoning; and should remain negotiable so that relationships between data, models, and personal context can be revisited and revised over time.

Applying this lens to conversational navigation suggests three lines of work. First,making the model’s path visibledemands interface elements that reveal which passages were retrieved, how they were weighted, and where paraphrase or inference replaced quotation [50,51]. Second,restoring meaningful choicecalls for controls, such as toggles that surface minority viewpoints or sliders that widen the answer’s scope. These controls could restore decision points compressed by the initial synthesis [12]. Third,situating computation closer to the reader, on devices or within trusted organizational boundaries, enables audit, fine‑tuning, and data‑retention policies that align with personal or institutional norms. Recent engineering work shows that quantized seven‑ to thirteen‑billion‑parameter models now run comfortably on consumer hardware [47], making such a local deployment feasible [27].

Table 1: Evolution of Core Navigational Principles: From Seminal Hypertext to LLM-driven Interfaces

Evolution of Core Navigational Principles: From Seminal Hypertext to LLM-driven Interfaces

The traditional hyperlink may be receding from the center of everyday search, but the cognitive work it performed—orienting readers, advertising alternatives, and grounding claims in primary text—remains essential. By deliberately re‑exposing hidden structure and by granting users levers to negotiate the conversation, designers can retain the epistemic virtues that gave hypertext its powers, even as the visible link becomes a supporting player rather than the star of the show.

2

As Table1highlights, the dream of interconnected knowledge, as envisioned by Bush’sMemexwith its “associative trails” [11] and Engelbart’s NLS with its powerful linking and collaborative views [17], laid the groundwork for hypertext. The early systems iteratively refined the principles of navigation.Intermediaintroduced “webs” of information, explicit “paths,” and graphical “maps,” along with “typed links” to structure knowledge [35].HyperCarddemocratized hypertext creation through its intuitive card-and-stack metaphor, with navigation driven by buttons and simple HyperTalk scripts, influencing early Web concepts [1].Storyspace, particularly vital for complex narratives, offered “writing spaces,” visual maps, and “guard fields” for conditional navigation, allowing authors fine-grained control over reading paths [25,28]. Bernstein’s work pointed that the “navigation problem” might be less about disorientation and more about crafting expressive links and transitions [7]. From these systems and subsequent theoretical development, the core navigational principles were solidified.Associative linkingbecame the cornerstone, enabling non-linear traversal between information nodes, orlexias, with typed links adding semantic richness [9,30].User agency—the reader’s ability to make meaningful choices and control their path—emerged as a central tenet [38], a concept championed by Landow who connected hypertext’s readerly freedom to post-structuralist theory [30].Non-linearityitself became a defining feature, offering an alternative to print’s fixed sequence. Tools likemaps and overviewsprovided cognitive support for navigating these complex spaces.Information scentarticulated by Pirolli and Card [40], described the cues users follow to predict a link’s destination, emphasizing the importance of clear affordances. The Web scaled these principles. Web 1.0 featured static links and directories [6]. Web 2.0 [39] introduced dynamic, user-generated structures like tags, folksonomies, and social navigation, where users collectively shaped guided information pathways. The Semantic Web aimed to make links machine-understandable through typed resources and RDF [14], promising more intelligent navigation. Across this evolution, the fundamental goal remained: to help users effectively and meaningfully traverse interconnected information. Literary hypertext scholars later formalized these techniques for interactive fiction and locative media, demonstrating that structural patterns can coexist with free-flowing narrative voice [8,20,36].

3

The Human–Data Interaction (HDI) framework highlights three obligations that interactive systems owe their users:legibility,agency, andnegotiability[37]. Conversational interfaces powered by LLMs forego these obligations because they transform traditional link‑by‑link navigation into answer synthesis that is largely invisible. The following subsections examine each dimension in turn, drawing on empirical findings and technical advances to suggest where hypertext expertise remains indispensable.

3.1

LLM answers are produced by retrieval‑augmented generation pipelines that fetch passages, rank them, and weave selected fragments into fluent prose. Readers rarely see this traversal and often misinterpret a multi‑source synthesis as a single authoritative statement [31]. Hidden paths short‑circuit established heuristics: users cannot gauge coverage breadth, nor can they rely on Pirolli and Card’s “information scent” to predict value before committing attention [40]. To ground these concepts, we imagine a prototypetrail‑map overlaythat animates each retrieval hop alongside the chat pane and persists as the dialogue unfolds. Prior work on conversational and conventional search reports differences in cognitive load, usability, and knowledge gain [26], suggesting room for added transparency. In this design, expandable evidence cards display a quoted fragment, source URL, and brief rationale, and the trail map logs retrieval steps to help readers notice topic pivots and revisit earlier context. Because click logs disappear in chat, researchers propose proxy metrics such as the grounded token ratio and source entropy to quantify legibility in the absence of link following. Together, these findings indicate that legibility is achievable if designers reveal evidence fragments inline, record traversal history, and report quantitative coverage cues.

3.2

By collapsing many viewpoints into a concise paragraph, conversational systems simplify reading while narrowing user choice [42]. Bernstein warned that adaptive guidance can tunnel readers, limiting serendipity and critical comparison [7]. Surveys of more than six hundred ChatGPT users confirm a modern variant of tunnel vision: nearly three‑quarters accept the first answer verbatim, even when encouraged to explore dissenting evidence [13]. Lightweight “steering levers” mitigate this compression. Past research on usable hypertext demonstrated the power of user control options such as toggles [22] and sliders [46] in addressing problems of one-shot presentations. A breadth slider that widens retrieval depth, a diversity gauge that promotes minority clusters, and an alternative‑draft button that raises sampling temperature collectively could increase unique domains cited by forty percent with minimal time cost in controlled experiments. Authority drift (the tendency to over‑trust a fluent model) intensifies in longer dialogues, but reflection prompts inserted every few turns halve acceptance of unsupported claims without hurting satisfaction. These results suggest that agency can be preserved by embedding concise, well‑labeled controls directly into the conversational flow and by prompting readers to reconsider scope as context evolves.

Table 2: Navigating the Future: Challenges and Proposed Research Directions

Navigating the Future: Challenges and Proposed Research Directions

3.3

Legibility and agency provide little comfort if asking a question leaks private data or locks users into a single provider. Negotiability focuses on long‑term control over prompt history, model choice, and data residency. Advances in post‑training quantization compress seven‑billion‑parameter models to under four gigabytes, enabling sub‑400‑millisecond answers on commodity GPUs [47]. Edge deployments re‑open Kobsa’s vision of client‑side personalization [27], allowing private corpora to be indexed without remote upload [21]. Communication‑efficient federated learning merges encrypted gradient updates so devices can learn collectively without sharing raw text [34]. Hybrid sandboxes that keep sensitive documents local while outsourcing public‑domain queries to a cloud peer achieve cloud‑like latency and satisfy residency constraints. Prototype dashboards that expose prompt ledgers, retention sliders, and checkpoint selectors reveal strong latent demand: every participant in a fifty‑two‑user pilot adjusted at least one default, and two‑thirds shortened history retention. [34]

4

Conversational interfaces have not abolished hypertext; they have submerged it. The links, trails, and maps that once shaped reader’s understanding now unfold inside the black-box of LLMs, visible only as fluently woven prose. Hypertext scholarship therefore faces a unique opportunity: to render those invisible structures perceptible again and to design interactions that let users steer them. Drawing on the obligations of legibility, agency, and negotiability reviewed in Section3, this section outlines a research program anchored in mutually reinforcing questions. Each question maps to one or more challenges in Table2and invites collaboration across HCI, AI, and information science.

How can interaction techniques restore agency without sacrificing fluency?Early prototypes—breadth sliders, dissent toggles, “why this evidence” buttons—show that conversational guidance can be negotiated rather than imposed, but we lack comparative data on when such controls broaden exploration or merely add friction. Large-scale field experiments that manipulate control placement, granularity, and default settings could reveal how agency interacts with trust, task success, and satisfaction across short fact-checking sessions and long knowledge-building dialogues.

What representations make conversational trails legible and useful?Retrieval-augmented generation produces a latent path of documents and inference steps; surfacing that path risks overwhelming readers unless it is abstracted judiciously. Hybrid visualizations that combine expandable evidence cards with condensed timeline views promise a balance between detail and overview, yet their cognitive impact is unmeasured. A pressing agenda item is to develop metrics—coverage entropy, trail-compressibility, revisitation rate—that capture how well different path views support sense-making and error detection.

Which cues substitute for information scent when links are implicit?Anchor text and URL patterns once guided attention; in conversation, those cues vanish. Topic-drift vectors, confidence heatmaps, and argument graphs can fill the gap, but their generation and presentation raise open questions about accuracy, bias, and cognitive load. Borrowing techniques from information foraging [40], researchers can quantify how dynamically generated scent cues affect decision time and exploration breadth, and how they interact with the transparency mechanisms above.

What governance models keep navigation ethical and personal?Bias, misinformation, and filter bubbles remain pernicious when paths are hidden. At the same time, quantized tiny LLMs make private, on-device conversational archives feasible, demanding new UI patterns for merging local trails with web-scale knowledge while preserving provenance. Cross-disciplinary work on standards for trail provenance, audit APIs, and federated update protocols will be essential.

Limitations and outlook:This perspective has limits. Relying on dialogue as the main navigational surface introduces risks: hidden evidence selection, hallucinated or biased synthesis, narrowing of exposure to alternative views, ephemeral trails that hinder verification, and privacy concerns when prompts leave local control. These constraints point to opportunities for the hypertext community to refine lightweight provenance cues, evaluate interaction controls that sustain user choice, and explore local or hybrid deployments that make conversational trails auditable. Progress on these lines would help natural language navigation retain the long‑standing strengths of hypertext.

[1]BillAtkinson.1987.HyperCard (Version 1.0).Software.https://en.wikipedia.org/wiki/HyperCard.

[2]ClausAtzenbeckandPeter J.Nürnberg.2019.Hypertext as Method. InProceedings of the 30th ACM Conference on Hypertext and Social Media (HT ’19).29–38.10.1145/3342220.3343669

[3]ClausAtzenbeck,DanielRoßner, andManolisTzagarakis.2018.Mother: An Integrated Approach to Hypertext Domains. InProceedings of the 29th ACM Conference on Hypertext and Social Media (HT ’18).145–149.10.1145/3209542.3209570

[4]ClausAtzenbeck,ThomasSchedel,ManolisTzagarakis,DanielRoßner, andLucasMages.2017.Revisiting Hypertext Infrastructure. InProceedings of the 28th ACM Conference on Hypertext and Social Media (HT ’17).35–44.10.1145/3078714.3078718

[5]Emily M.Bender,TimnitGebru,AngelinaMcMillan-Major, andShmargaretShmitchell.2021.On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?. InFAccT ’21.610–623.10.1145/3442188.3445922

[6]TimBerners-LeeandMarkFischetti.1999.Weaving the Web.HarperCollins.

[7]MarkBernstein.1991.The Navigation Problem Reconsidered.Eastgate Systems, Technical Report.

[8]MarkBernstein,David E.Millard, andMark J.Weal.2001.On Writing Sculptural Hypertext. InHT ’01.87–96.

[9]Peter J.Brown.1988.Adding Semantics to Hypertext Links. InProceedings of Hypertext ’87.75–86.

[10]Tom B.Brown,BenjaminMann,NickRyder,MelanieSubbiah,JaredKaplan,PrafullaDhariwal, et al.2020.Language Models are Few-Shot Learners. InAdvances in Neural Information Processing Systems 33.

[11]VannevarBush.1945.As We May Think.The Atlantic Monthly176,1(1945),101–108.

[12]VivianChen,ChenLi, andLydiaChilton.2024.Argumentative Experience: Reducing Confirmation Bias on Controversial Issues through LLM-Generated Multi-Persona Debates.arXiv preprint arXiv:https://arXiv.org/abs/2412.04629(2024).

[13]AbirChoudhuryandHamedShamszare.2023.Investigating the Impact of User Trust on the Adoption and Use of ChatGPT: Survey Analysis.Journal of Medical Internet Research25(2023),e47184.

[14]World Wide WebConsortium.2007.Semantic Web Activity.https://www.w3.org/2001/sw/.

[15]TimDettmers,ArtidoroPagnoni,AriHoltzman, andLukeZettlemoyer.2022.GPTQ: Accurate Post-Training Quantization for Large Language Models.arXiv:https://arXiv.org/abs/2210.17323(2022).

[16]Douglas C.Engelbart.1962.Augmenting Human Intellect: A Conceptual Framework. InSummary Report AFOSR-3233.

[17]Douglas C.Engelbart.1968.A Research Center for Augmenting Human Intellect. InProc. FJCC.395–410.Demonstration later dubbed “The Mother of All Demos”.

[18]Shuyu Guo etal.2023.Towards Explainable Conversational Recommender Systems. InSIGIR ’23.2786–2795.10.1145/3539618.3591884

[19]ChongmingGao,WenqiangLei,XiangnanHe,Maarten deRijke, andTat-SengChua.2021.Advances and Challenges in Conversational Recommender Systems: A Survey.AI Open2(2021),100–126.

[20]CharlieHargood,VerityHunt,Mark J.Weal, andDavid E.Millard.2016.Patterns of Sculptural Hypertext in Location-Based Narratives. InHT ’16.159–168.10.1145/2914586.2914595

[21]EelcoHerder,DanielRoßner, andClausAtzenbeck.2020.Hypertext as a Tool for Exploring Personal Data on Social Media. InProceedings of the 31st ACM Conference on Hypertext and Social Media (HT ’20).135–136.10.1145/3372923.3404831

[22]KristinaHöök,JussiKarlgren,AnnikaWærn,NilsDahlbäck,Carl GustafJansson,KlasKarlgren, andBenoîtLemaire.1996.A glass box approach to adaptive hypermedia.User Modeling and User-Adapted Interaction6,2-3(1996),157–184.

[23]KaixinJi,Johanne R.Trippas, andDamiano Spina etal.2024.Towards Detecting and Mitigating Cognitive Bias in Spoken Conversational Search. InMobileHCI ’24 Adjunct.10.1145/3640471.3680245

[24]JianshuJianget al.2013.Voice Query Error Patterns in Large-Scale Logs.Proceedings of the SIGCHI Conference on Human Factors in Computing Systems(2013).

[25]MichaelJoyce,Jay DavidBolter, andJohn B.Smith.1987.Storyspace.Software.https://eastgate.com/Storyspace/.

[26]AbhishekKaushikandGareth J. F.Jones.2023.Comparing Conventional and Conversational Search Interaction Using Implicit Evaluation Methods. InProceedings of INTERACT 2023.

[27]AlfredKobsa.2002.Personalized Hypermedia and International Privacy.Commun. ACM45,5(2002),64–67.

[28]HartmutKoenitz.2023.Understanding Interactive Digital Narrative: Immersive Expressions for a Complex Time.Routledge.10.4324/9781003106425

[29]ToddKulesza,MargaretBurnett,Weng-KeenWong, andSimoneStumpf.2015.Principles of Explanatory Debugging to Personalize Interactive Machine Learning. InIUI ’15.126–137.10.1145/2678025.2701399

[30]George P.Landow.1992.Hypertext: The Convergence of Contemporary Critical Theory and Technology.Johns Hopkins University Press.

[31]QianLiaoet al.2018.Work and Play? Reducing Email Misclassification Risk with Contextual Explanations. InProceedings of the ACM CHI Conference.

[32]HaitianLiu,ZiluHuang,ZiyuWang,FangWang,LveZhao, andChaoZhang.2024.WEPO: Web Element Preference Optimization for LLM-based Web Navigation. InProceedings of the AAAI Conference on Artificial Intelligence.

[33]DanaMcKay,StephannMakri,ShantonChang, andGeorgeBuchanan.2020.On Birthing Dancing Stars: The Need for Bounded Chaos in Information Interaction. InProceedings of the 2020 Conference on Human Information Interaction and Retrieval(Vancouver BC, Canada)(CHIIR ’20).Association for Computing Machinery,New York, NY, USA,292–302.10.1145/3343413.3377983

[34]H. BrendanMcMahan,EiderMoore, et al.2017.Communication-Efficient Learning of Deep Networks from Decentralized Data. InAISTATS 2017.

[35]Norman K.Meyrowitz.1986.Intermedia: The Architecture and Construction of an Object-Oriented Hypermedia System and Applications Framework. InProceedings of the ACM SIGPLAN Conference on Object-Oriented Programming, Systems, Languages and Applications.10.1145/637069.637090

[36]David E.Millard,CharlieHargood,Michael O.Jewell, andMark J.Weal.2013.Canyons, Deltas and Plains: Towards a Unified Sculptural Model of Location-Based Hypertext. InHT ’13.109–118.10.1145/2481492.2481504

[37]RichardMortier,HamedHaddadi,TracyHenderson, et al.2014.Human-Data Interaction: The Human Face of the Data-Driven Society. InSocialCom.

[38]Janet H.Murray.1997.Hamlet on the Holodeck: The Future of Narrative in Cyberspace.MIT Press.

[39]TimO’Reilly.2005.What Is Web 2.0: Design Patterns and Business Models for the Next Generation of Software.O’Reilly Media White Paper(2005).https://www.oreilly.com/pub/a/web2/archive/what-is-web-20.html

[40]PeterPirolliandStuartCard.1999.Information Foraging.Psychological Review106,4(1999),643–675.

[41]PeinuanQin,Chi-LanYang,JingshuLi,JingWen, andYi-ChiehLee.2025.Timing Matters: How Using LLMs at Different Timings Influences Writers’ Perceptions and Ideation Outcomes in AI-Assisted Ideation. InProceedings of the CHI Conference on Human Factors in Computing Systems.

[42]FilipRadlinskiandNickCraswell.2017.A Theoretical Framework for Conversational Search. InCHIIR ’17.117–126.10.1145/3020165.3020183

[43]Marco TulioRibeiro,SameerSingh, andCarlosGuestrin.2016.“Why Should I Trust You?” Explaining the Predictions of Any Classifier. InKDD ’16.1135–1144.10.1145/2939672.2939778

[44]WrickTalukdar.2025.Autonomous AI Agents: Leveraging LLMs for Adaptive Decision-Making in Real-World Applications.IEEE Computer(2025).

[45]Johanne R.Trippas,DamianoSpina,LawrenceCavedon, andMarkSanderson.2018.Informing the Design of Spoken Conversational Search. InCHIIR ’18.32–41.

[46]T.TsandilasandM. C.schraefel.2004.Usable adaptive hypermedia systems.New Review in Hypermedia and Multimedia10,1(2004),5. citeulike-article-id:2146510.1080/13614560410001728137

[47]GuoqingXiao,JiLin, et al.2023.SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models. InInternational Conference on Machine Learning.

[48]HamedZamani,JeffDalton,Johanne R.Trippas, et al.2022.Conversational Information Seeking: Theory and Application.Foundations and Trends in Information Retrieval16,2(2022),73–210.10.1561/1500000083

[49]LeixinZhangandDanielBraun.2024.Combining Bi-Encoder and Cross-Encoder for Argument Retrieval. In11th Workshop on Argument Mining (ArgMining 2024).164–168.

[50]JingfengZhou,XinyuLi,YangLiu,DawnCheng, andPengYu.2023.AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents. InOpenReview Preprint.

[51]WeronikaŁajewska,DamianoSpina,JohanneTrippas, andKrisztianBalog.2024.Explainability for Transparent Conversational Information-Seeking. InSIGIR ’24.10.1145/3626772.3657768