The Web Was Built for Navigation, Not Comprehension

You are reading a paper on watershed governance in sub-Saharan Africa. Forty seconds in, you encounter the phrase "polycentric governance." You know governance. You know poly- means many. But polycentric governance, as a term of art in institutional economics — the specific meaning Elinor Ostrom gave it, the tradition it sits in, the way it differs from federalism — that you don't know. Not quite.

You have two choices. Keep reading and carry the term as an unresolved variable, hoping context will fill it in. Or stop, open a new tab, type the phrase, scan a Wikipedia summary, read half a paragraph, realize you need to understand Ostrom first, open another tab, skim her Nobel lecture, get mildly distracted by something else on the screen, and — some while later — make your way back to the paper you were reading, having lost the thread.

The second option feels like the right one. In practice it is costly to sustained comprehension, for reasons this paper tries to state precisely rather than merely assert.

This is not a self-discipline problem. It is an architecture problem. The web was designed to let you move between documents. It was not designed to help you understand any one of them.

The hyperlink connects documents. It was never built to reach the exact word you stalled on.

1. What this argument rests on, and the statistic it had to drop

This paper argues that the web's core primitives — the link, the page, the tab — were built to solve document connectivity, not in-document comprehension, and that this specific mismatch (not screens generally, not the internet generally) is what makes stopping to look up an unfamiliar term so costly to a reader's progress through a text.

What it argues from. A documented history of two things: the web's own founding design intent, read from Tim Berners-Lee's original 1989 proposal, and the much older marginalia/glossing tradition it structurally replaced, read from a specific, real example rather than an invented generic one. A body of published research on reading comprehension under cognitive load — specifically hypertext reading, dictionary-versus-gloss vocabulary lookup, and task interruption and resumption — checked against primary sources rather than against the version of these findings that circulates informally. A correction made during this research, stated plainly in §4: a widely circulated "23 minutes to refocus" statistic could not be located in either of the two studies most commonly cited for it, and this paper does not use it.

What it does not argue from. It does not argue from a controlled study of in-document term lookup specifically — no published experiment that measures the recovery cost of opening a new tab mid-sentence for a single unfamiliar term, as distinct from a workplace task interruption or a language-learner's dictionary lookup, was found. §4 is explicit about which literature is adjacent to this paper's claim and which would need to exist to settle it directly.

What it does not claim. It does not claim comprehension aids are a fully solved engineering problem — §6 is explicit that interaction design, not model capability, is the unsolved part. It does not claim screens or computers are inherently hostile to deep reading; §3 argues the hostility is a specific, contingent consequence of which affordances the browser built and which it didn't, not a property of digital text as such.

What would refute it. Given in full in §9. The sharpest version: a study that directly measures in-document word-lookup interruptions (not workplace task-switching, not L2 dictionary use) and finds the cost of a tab-based lookup is small — comparable to, say, glancing at a footnote — would mean this paper has misapplied adjacent research to a case it does not actually describe, and the argument would need withdrawing rather than softening.

2. A Link Is an Exit, Not an Explanation

The hyperlink is the defining invention of the web. Tim Berners-Lee's original March 1989 proposal to CERN, "Information Management: A Proposal," framed the problem as one of connecting distributed, evolving technical documentation — he was explicit that generality and portability mattered more than "fancy graphics techniques and complex extra facilities."¹ The link says: this term connects to another document. Click here and you will go there.

What a link cannot do is bring the other document to you. It cannot surface the relevant sentence from a 40-page paper without asking you to navigate to that paper. It cannot synthesize several competing definitions of a term and present them in a single sentence inline with the text you are reading. It cannot distinguish between a reader who needs the definition and one who already has it.

The hyperlink is a navigation primitive. It solves the problem of document connectivity — the problem that mattered most in 1989, when the challenge was getting distributed technical knowledge to link up at all. It does not solve the problem of in-document comprehension, because that problem didn't motivate the design. The architecture optimized for what it was asked to optimize for.

The result is a medium that is extraordinary for discovery and comparatively weak for depth.

This is not a criticism of Berners-Lee or the early web. The early web solved exactly the problem it set out to solve. But architectural choices made for good reasons in one era become invisible constraints in the next — and §4 and §5 below try to state, with real sources rather than received wisdom, how large that constraint actually is.

3. What We Lost When We Gave Up Marginalia

Before the web, serious readers had a technology for comprehension that worked remarkably well: marginalia. Annotations, glosses, cross-references, interlinear notes — the scholarly tradition of writing on and between the text rather than separate from it.

The clearest documented example is the Glossa Ordinaria, the standard medieval Latin commentary on the Bible, compiled in the first half of the twelfth century and associated above all with Anselm of Laon. Its distinctive layout placed a central column of biblical text surrounded by marginal glosses — longer exegetical notes — and interwoven with short interlinear glosses between the lines of the text itself: two parallel commentaries running alongside the primary document at once.² It survives in thousands of manuscripts and was, for several centuries, the standard reference text for biblical study in European schools and universities. The reader moved through base text, interlinear gloss and marginal gloss together. The explanation was co-located with the confusion, not filed in a separate volume.

The print revolution disrupted this, not deliberately but structurally. Printed books are harder to annotate than manuscripts prepared for a glossing tradition. The economics of the printing press favored clean, replicable pages. Footnotes and endnotes emerged as compromises — explanations that preserved the visual cleanliness of the page while exiling the commentary to the margins or the back of the book. The further the note sits from the text, the more context is lost by the time a reader arrives there.

The digital era promised to reverse this. Hypertext theorists in the 1980s imagined linked, annotated documents as a return to the richness of manuscript culture. What arrived instead was documents even harder to annotate than printed pages — smooth, rendered HTML with no natural place for reader-generated commentary, surrounded by interfaces designed to move a reader through content rather than let them linger in it.

Reading on a screen is navigation. The scroll gesture, the back button, the address bar, the tab — every affordance of the browser is a navigation affordance. Comprehension affordances — pause, annotate, ask, expand — have to be bolted on after the fact, against the grain of the medium.

4. What the Research on Reading and Interruption Actually Shows

A widely circulated figure claims it takes 23 minutes to refocus after an interruption, usually attributed to Gloria Mark's research and dated to 2005. Checked directly against Mark's two most relevant published studies, the figure appears in neither of them. Mark, Gonzalez and Harris's 2005 CHI paper, "No Task Left Behind? Examining the Nature of Fragmented Work," observed 24 information workers at a financial-services outsourcing firm over several hundred hours and found they spent an average of 11 minutes 4 seconds in a given "working sphere" before switching or being interrupted, with 57.1% of working-sphere segments interrupted — a real and striking finding about workplace fragmentation, but not a 23-minute recovery figure, and not about reading.³ Mark, Gudith and Klocke's 2008 CHI paper, "The Cost of Interrupted Work: More Speed and Stress," is a controlled lab experiment with 48 subjects performing an email task under three interruption conditions, and its headline finding is arguably the opposite of what the circulated figure implies: interrupted work did not take longer to complete, because people compensated by working faster — at a measurable cost in stress, frustration, time pressure and effort, not in elapsed time.⁴ Neither paper supports the 23-minute claim, and this paper does not use it. Readers of this corpus will recognise this as the same untraceable statistic flagged and avoided in a companion paper on AnswerTab's design;⁵ it is evidently in wide informal circulation and worth retiring rather than repeating.

What the genuinely relevant literature shows is narrower and, I think, more useful. DeStefano and LeFevre's 2007 review of cognitive load in hypertext reading, synthesising a body of experimental studies, found that the decision-making and visual-processing demands hypertext adds — deciding whether to follow a link, tracking where you are, holding multiple paths in mind — measurably impair reading performance relative to linear text, with the effect falling hardest on readers with lower working-memory capacity and less prior knowledge of the subject.⁶ That is a direct, verified finding about exactly the kind of in-document decision a reader faces when they hit an unfamiliar term: not "will I eventually resume my original task," which is what Mark's research measures, but "does stopping to navigate, mid-sentence, cost comprehension even before any tab is opened."

The second body of literature concerns what happens once a reader does stop. Research on second-language vocabulary acquisition has compared marginal glosses — a brief explanation placed next to the difficult word — against dictionary lookup, which requires leaving the text. Knight's 1994 study of university Spanish learners found dictionary access improved comprehension scores for lower-verbal-ability readers substantially (a 29.0% gain) and higher-verbal-ability readers only marginally (7.4%) — evidence that the cost of interrupting a reader to resolve a gap is not uniform, and falls hardest on exactly the readers who can least afford it.⁷ Hulstijn, Hollander and Greidanus's 1996 study of advanced French learners found that marginal glosses and dictionary use both supported vocabulary learning better than unassisted reading, with roughly a quarter of looked-up words retained after a single exposure — a real benefit to stopping and checking, which this paper is not arguing against.⁸ What neither study measures directly is reading speed or time cost, which is the gap this paper's own claim — about the comprehension layer's effect on throughput, not just outcome — cannot yet close with existing literature; §10 registers this directly.

Finally, on why resumption specifically carries a cost once a reader has left the page: Altmann and Trafton's model of task resumption as a retrieval problem, cued by what is perceptually available in the environment, and Trafton, Altmann, Brock and Mintz's finding that perceptual availability of the primary display is what drives successful re-encoding of a suspended goal — both already established elsewhere in this corpus for a closely related product — explain mechanistically why a reader returning from a new tab, with the original document no longer on screen, has a harder time re-entering the sentence than a reader whose eyes never left the page.⁹ This is the most defensible version of the "cost of leaving" claim available: not a specific number of minutes, but a mechanism, independently documented, for why leaving costs more than staying.

5. The Scrollbar Became the Progress Metric

There is a second, subtler problem with the web as a reading medium, separate from the interruption cost in §4. The dominant metaphor for reading progress online is physical position in the document — the scrollbar.

The scrollbar is a navigation metaphor. It tells you where you are in the document, not whether you understood what you just read. It measures exposure to words, not comprehension of them. And because the scrollbar is visible and moving feels like progress, the medium implicitly encourages a reading behavior that privileges completion over understanding: skim to the bottom, see the scrollbar reach 100%, close the tab.

This isn't the reader's fault. It's a design affordance. The medium rewards coverage. The marginalia tradition rewarded depth.

Mortimer Adler and Charles Van Doren's How to Read a Book calls books "absent teachers" — present to instruct, but not present to answer a question or notice confusion the way a live teacher can.¹⁰ Elsewhere in the same book, discussing a reader who skates past unfamiliar words as if they were already understood: "you might just as well be reading a newspaper, for the book cannot enlighten you if you do not try to understand it."¹⁰ Adler and Van Doren were writing about printed books, decades before the web. The same gap exists online — but the web adds a wrinkle their book could not have anticipated: the interface actively discourages the interrogation they were recommending. A book sits still while you think about it. A web page is surrounded by notifications, related links, autoplay videos, and a toolbar full of escape routes. The architecture of the browser is the architecture of distraction.

6. The Comprehension Layer That Doesn't Exist

If you sketch what the web is missing, it looks like this:

A layer that sits between the reader and the document. That watches what you're reading, not to surveil you but to be ready. That activates when you select a phrase that you don't fully understand, or when you reach for a new tab. That surfaces the explanation in context — not as a link to another document, but as a few sentences, generated for this reader, in relation to this passage, right now.

Not a summary (summaries are still navigation — they move you above the document). Not a link (links are exits). An annotation. The gloss that the web never had — the Glossa Ordinaria's interlinear note, rebuilt for a reader whose text is a PDF rather than a manuscript.

The technical components to build this have existed in crude form for several years and in considerably more capable form since large language models became broadly accessible. Browser extension APIs can observe selected text. The marginal cost of a well-targeted explanation is close to zero.

What has been harder to get right is the interaction design, and this is the part a companion paper in this series, which describes building exactly such a tool and living with it, treats as the actual finding rather than a footnote: the explanation has to appear at the point of confusion without a mode switch, has to be brief by default and expandable by choice, and has to work with the grain of the reading experience rather than adding friction to it — a sidebar that requires context-switching into it, or a chat box that asks the reader to type a question, defeats the purpose it was built to serve.¹¹ This paper argues why that layer should exist; that companion paper and its sequel report what happened when Flux actually built one.

7. BYOK as the Architecturally Correct Response

There is a separate problem with building a comprehension layer for serious readers: the intimacy of the data.

What you look up, what you don't know, what confuses you, what questions you ask about the things you're reading — this is not generic behavioral data. It is a map of the edges of your understanding. It reveals your domain, your level of expertise, your research agenda, your intellectual gaps. A companion paper in this corpus develops the privacy argument for this category of data — epistemic data — in full; this paper states the narrower design conclusion it implies.¹²

A comprehension tool that routes your queries through a proprietary platform to be logged, analyzed, and potentially used to train models or target advertising is not simply a comprehension tool with a privacy cost attached. It risks becoming a surveillance system dressed as a reading aid. The users who would benefit most from serious reading assistance — researchers, journalists, policymakers, analysts — are precisely the users who can least afford to expose their knowledge boundaries to a third party.

Bring Your Own Key (BYOK) — the reader's own API key, their data and queries going directly to the model provider, no intermediary platform capturing the signal — is the architecturally correct response to this specific risk, not a response to every risk a comprehension tool might carry. It does not, by itself, guarantee the model provider does not log or train on the traffic; it removes one intermediary, not all of them, and §9 names this as a real limit on the claim.

This is not primarily a privacy-as-virtue argument, and it is worth being honest that it is also a convenient one: the users most likely to pay for a serious comprehension tool are also the users with the highest reading volume and the highest stakes attached to what they read, who have typically already thought about where their data goes and will find a BYOK design reassuring rather than inconvenient. A design choice that is architecturally right and also commercially convenient for exactly the users worth building for is not thereby suspect, but it is worth naming rather than presenting as disinterested.

8. What the Comprehension Layer Changes

The case for a comprehension layer is not primarily about convenience. It is about what kind of knowledge work becomes possible when the friction of deep reading drops.

The effective reading speed of a serious knowledge worker on complex material is throttled not by how fast they read but by how often they stop, and how expensive each stop is. A long policy paper with many unfamiliar terms or contested claims takes disproportionately longer than its word count would suggest, because each unfamiliar term can spawn a lookup that spawns a wider distraction. This paper does not have a measured, general figure for how much time a comprehension layer reclaims — §4 is explicit that no study measures exactly this case — but the mechanism described in §§2–4 (navigation cost plus resumption cost, compounding across many stops in a single document) predicts the effect should be real and should scale with a document's term density, which is a testable claim rather than an assumption, and §9 states it as one.

The harder case is what happens to comprehension quality, not just speed. The tab-opening pattern doesn't just cost time. It plausibly punishes depth: a reader who has learned, through repeated bad experience, that an unfamiliar-term-heavy text means a string of costly interruptions has a real incentive to avoid such texts — which selects against exactly the material most worth reading closely. This paper does not have direct evidence this avoidance effect operates at scale; it is named here as a hypothesis the mechanism in §4 makes plausible, not a measured outcome.

Vannevar Bush, writing in 1945 about a hypothetical memory-augmenting device he called the memex, put the underlying premise plainly: "The human mind does not work that way. It operates by association. With one item in its grasp, it snaps instantly to the next that is suggested by the association of thoughts," he wrote, describing thought as moving along "some intricate web of trails carried by the cells of the brain."¹³ Bush was writing eight decades before anything resembling the tool he described was technically buildable at the quality this paper is discussing. The attention economy that grew up around hyperlinked media in the decades since is built on navigation — on keeping a reader moving through content rather than deepening in it — which is a direct structural opposite of what Bush was describing and what §6's comprehension layer tries to restore. A platform whose business model depends on time-on-platform has little incentive to build a tool whose explicit purpose is to help a reader finish one document well and leave.

9. Objections I take seriously

"Flux has a commercial interest in this conclusion." AnswerTab, the tool named directly in §6, is a Flux product, and this paper's argument — that the web has a structural comprehension gap only an inline explanation layer can close — is also, uncharitably read, a justification for exactly the product Flux sells. I am not going to pretend otherwise. The mitigation available is narrow: §§2–4's account of the web's navigation-first design and the relevant cognitive-load and vocabulary-acquisition literature would be true whether or not Flux had ever built AnswerTab; what is specific to Flux is the interaction-design claim in §6, which is reported there as the companion papers' own finding, not asserted here as proven.

"This paper leans on research about workplace interruptions and L2 vocabulary learning, neither of which is really about in-document word lookup by a fluent adult reader — the correction in §4 doesn't go far enough." This is a fair and serious objection, and I think it is largely right. §4's table is explicit about the gap between what the cited studies measure and what this paper actually claims; the honest position is that no study directly measures this paper's central case, and the argument rests on adjacent literature plus a stated mechanism (§4's last paragraph, on perceptual availability) rather than a direct measurement. §10 states this as a limitation rather than treating §4's correction as having closed the gap.

"The comprehension layer described in §6 could just as easily increase distraction as reduce it — an always-available explanation might encourage readers to stop at every unfamiliar term rather than push through productive difficulty, the way a GPS can erode a traveller's own sense of direction." I take this seriously and do not have a study that settles it either way. It is a real risk the interaction-design problem in §6 does not solve by construction; a comprehension layer built badly could plausibly make the scrollbar-skimming behaviour in §5 worse, not better, by making every stop feel costless.

"BYOK is a smaller privacy improvement than §7 implies, and convenient for Flux besides." Correct on both counts, and §7 says so directly rather than treating BYOK as a complete solution. BYOK removes one intermediary's access to a reader's query history; it says nothing about the model provider's own logging or retention practices, which §9's falsification conditions name as a real exposure this paper has not investigated.

10. What would falsify this

Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.

A study directly measuring in-document, mid-sentence word-lookup interruptions — as distinct from workplace task-switching or L2 dictionary use — finds the cost is small, comparable to a glance at a footnote. That would mean this paper has misapplied adjacent research (§4) to a case the cited literature does not actually describe, and the central claim would need withdrawing, not softening.

A comprehension-layer tool, deployed and measured, is found to increase total reading-session abandonment or reduce comprehension-test scores relative to a plain hyperlink baseline — the "always-available explanation erodes productive struggle" objection in §9, confirmed rather than merely raised. That would mean §6's proposed fix makes the problem described in §5 worse, not better.

The model provider behind a BYOK integration is shown to log, retain, or train on query traffic regardless of the BYOK architecture, undermining §7's claim that removing the platform intermediary meaningfully protects a reader's epistemic data. That would not refute §§2–5's account of the web's design, but it would substantially weaken §7's specific architectural recommendation.

A large-scale analysis of real reading sessions finds no relationship between a document's density of unfamiliar terms and either reading time or abandonment rate, contrary to §8's prediction that the mechanism in §§2–4 should scale with term density. That would mean the mechanism this paper describes, even if real in the lab studies cited in §4, does not actually operate at the scale or in the way §8 predicts for ordinary reading.

11. Limitations, and what this paper does not claim

No study exists that directly measures the cost of an in-document, mid-sentence tab-switch for a single word lookup; §4 states plainly that none was found, and the argument instead combines adjacent literature (hypertext cognitive load, L2 glossing-versus-dictionary research, task-resumption mechanism research) with a stated inferential bridge.

It does not claim, and no longer uses, the "23 minutes to refocus" figure that circulates widely in discussions of this topic; §4 documents why that figure could not be verified against either of the two studies most commonly cited for it.

It does not claim the comprehension layer described in §6 is risk-free; §9's second objection — that an always-available explanation could erode productive reading struggle rather than support it — is a real, unresolved risk this paper does not have evidence to rule out.

It does not claim BYOK solves the privacy problem named in §7 completely; it addresses one intermediary's access to query data, not a model provider's own logging practices, which this paper has not investigated.

And it does not claim this is a general theory of digital reading or attention; it is a claim about one specific architectural mismatch — navigation primitives standing in for comprehension primitives — in one specific medium, the hyperlinked web page.

Appendix A — Evidence table

Appendix B — Citation register

Fabrication found and removed, not merely flagged. The previous draft of this paper attributed a direct quotation to Mortimer Adler ("Reading is learning from an absent teacher. The author cannot see your confusion...") that does not appear verbatim in How to Read a Book. It has been replaced in §5 with a verified short quotation from the actual text and an accurate paraphrase of the book's "absent teachers" framing, sourced to the specific page content located during this revision.

Figure corrected, not merely softened. The previous draft attributed "23 minutes" to "Mark 2005" as a recovery-time finding. Checking both of Gloria Mark's most relevant published studies directly found this figure in neither; §4 documents the correction and the paper does not use the figure anywhere.

CITATION-NEEDED — in-document word-lookup cost, specifically. Appendix A row 11. No study directly measuring this paper's central case (a mid-sentence tab-switch to look up one term, by a fluent adult reader, outside a workplace-interruption or L2-vocabulary-acquisition setting) was found.

CITATION-NEEDED — comprehension-layer effect on productive struggle, at scale. Appendix A row 12. §9's second objection names a real, plausible risk this paper cannot rule out with available evidence.

CITATION-NEEDED — term density and real-world reading behaviour. Appendix A row 13. §8's prediction that the mechanism should scale with a document's density of unfamiliar terms is stated as a testable claim, not a measured one.