The Panel Is Also a Tab

A whole product category now exists to explain a word to you without making you leave the page: reading assistants, AI answer panels, hover-definition browser extensions, the summary that sits above the search results. All of them are aimed at the same moment, and all of them describe it the same way.

You know the moment. You are reading. You hit a word, a name, an acronym, a clause you do not know. You open a new tab. Twenty seconds or twenty minutes later you come back, and the paragraph you were in has gone cold.

The industry agrees on the problem. What it has converged on is a family of solutions that all make the same move: bring the answer closer. Put it in a panel beside the page. Put it behind a keystroke that floats an overlay above whatever you were doing. Put it directly into the search results so there is no second page to load at all.

Each of these is an improvement on the tab. None of them is what it claims to be, which is an answer that arrives without moving you. The panel is also a tab. It is a narrower tab, docked, with a faster animation — and it still takes your eyes, your cursor and your place in the text.

This paper is about the difference between closer and in place, why that difference is the whole product, and where the in-place answer is the wrong design and the panel is right.

Every answer surface solves the tab problem by getting smaller. The exit path gets longer.

1. A mechanism paper, not a measured one, and what would undo it

This is a mechanism paper, built on observed product behaviour and first-person use rather than a controlled study — stated plainly before the argument rather than left for a footnote to imply something stronger later.

What it argues from. Three things. First, the current generation of mainstream answer surfaces, described as patterns rather than pinned to a product version, because the specific tools will have changed by the time this is read and the shapes will not have. Second, roughly a decade of published cognitive-science research on interrupted tasks and the memory structures that support resuming them — Altmann and Trafton's activation-based model, Trafton, Altmann, Brock and Mintz's study of resumption, Monk, Trafton and Boehm-Davis's work on interruption cost, and Leroy's work on attention residue. None of these studies was run on the specific case this paper is about; §4 is explicit about where the borrowing is sound and where it is an extrapolation. Third, the companion paper, The Second Question, which this paper's §7 relies on without re-arguing.

What it does not argue from. No user study of its own. No measured number for how long re-entry into a reading task actually takes, for the specific case of a self-initiated, in-document lookup — the literature cited in §4 measures externally imposed interruptions, which is a related but not identical situation, and I am not going to borrow an effect size from a different task and present it as if it were measured here. Where a widely circulated number exists for "time to refocus" after an interruption, it is not used anywhere in this paper, because it was not measured under conditions close enough to this one to trust.

What would refute it. Given in full in §12. The short version: if a controlled study finds that resumption cost does not scale with how demanding or engaging the interrupting surface was, the mechanism this paper leans on hardest — that a richer panel conversation makes the return more expensive, not less — is wrong, and the paper's central design claim would need to be rebuilt on a different foundation.

2. What we already argued

Two earlier papers in this series set up the ground. The Web Was Built for Navigation, Not Comprehension made the structural claim: the web's primitives — the link, the page, the tab — are all instruments of going somewhere else, and comprehension is the one task that going somewhere else actively damages. Your Knowledge Gaps Are the Most Intimate Data You Have made the consequence explicit: the record of what you did not understand is the most revealing thing a reading tool could hold, which is a reason to be careful about where that record lives.

This paper is the design argument that follows from the first, and it is narrower than either. It is about a single mechanism: where the answer appears, and what that costs.

3. Three ways the answer got closer

The current generation of answer surfaces comes in three shapes. The labels and the specific interfaces change often enough that pinning this to a version number would date the paper within a quarter, so what follows describes the patterns rather than the products.

The docked panel. A column opens beside what you were reading and holds a conversation. Browser assistants do this; search engines increasingly do it too, inviting you to continue a query with an assistant in a surface that opens alongside the results. The reading area is not replaced — it is squeezed.

The launcher overlay. A keystroke floats a box above everything. This is the Spotlight lineage: a modal surface, summoned from anywhere, that expects a typed command and then disappears. Desktop assistants have widely adopted it because it is the fastest thing to build that is available system-wide.

The rewritten result. The answer is composed directly into the page you would have navigated to anyway — the search results page becomes the answer rather than an index of places the answer might be.

The third of these is genuinely different and genuinely good, and it is worth saying so plainly. It removes a navigation. But it only helps when the question was a search. It does nothing for the case this paper is about, which is a question that arose inside a document you are already committed to reading.

4. What the first two have in common

Both the panel and the overlay solve the tab and leave the cost.

Consider what actually happens when you use them. Your eyes leave the sentence. Your attention re-anchors on a new region of the screen. You read an answer composed without reference to the specific sentence that confused you, because you had to restate the question in a box. Then you go back — and going back is the expensive part. You have to find the sentence again, reload the argument up to that point, and recover the reason you stopped.

The tab was never the cost. Re-entry was the cost. The tab was just the most visible carrier of it.

This is not only a claim from use; it is the shape of a specific, published mechanism. Altmann and Trafton's activation-based model of goal memory treats resuming an interrupted task as a retrieval problem: a goal's activation decays once attention moves away from it, and what reactivates the goal is an associative cue in the environment — for a reading task, the cue is the sentence you were on, visible and in the state you left it.¹ A panel that squeezes the column, and an overlay that covers it, both reduce the availability of exactly that cue, at exactly the moment the reader most needs it back.

Monk, Trafton and Boehm-Davis's study of interruption cost adds a sharper and somewhat counter-intuitive point: the cost of resuming a suspended goal scales with both the duration of the interruption and its demand — a longer, more cognitively engaging interruption is more expensive to return from, not less.² Applied here, this cuts against the intuition that a good, helpful panel conversation is a strict improvement over a bad one. A richer, multi-turn exchange in a docked panel is, by this finding, a more demanding interruption than a single quick glance at a definition — which means the panel's own quality works against the reader's ability to return, precisely when the panel is doing its job well.

Trafton, Altmann, Brock and Mintz's study of task resumption found that perceptual availability of the primary display — not just memory of it — measurably aids re-encoding the interrupted goal when the interruption ends.³ An overlay fails this specifically and structurally: it does not merely add a demanding side conversation, it removes the primary display from view while that conversation happens, at the exact moment the reader would otherwise be using it to hold the question in mind.

A panel reduces the distance you travel and leaves the re-entry intact, now with an added cost proportional to how good the conversation was. An overlay is worse on a second axis: it covers the thing you were reading, so the context you needed to formulate the question is occluded at the exact moment you are formulating it. This is why the launcher pattern, which is excellent for launching, is a poor fit for understanding. Spotlight is optimised for "go do a thing." It has no notion of the thing in front of you. Asking it to help you read is asking a doorway to be a desk.

None of this is a claim about how many seconds or minutes anything costs for the specific case of a self-initiated, in-document lookup. The three studies above measured externally imposed interruptions — a phone ringing, a task switch commanded by an experimenter — not a lookup the reader chooses to make mid-sentence. The directional claim survives that gap: resumption has a cost, the cost is separate from travel time, and the cost scales with how demanding the interruption was. The specific number for this case remains unmeasured, and CITATION-NEEDED: a study measuring resumption lag for a lookup initiated from within a reading task, rather than imposed from outside it is carried into Appendix B rather than estimated here.

Leroy's work on attention residue supplies the last piece: switching away from an unfinished task leaves a measurable drag on performance of the task you return to, driven by incomplete attentional disengagement from the one you left — which means even a "successful" panel exchange that fully answers the question can leave the reader measurably worse at the reading task for a period afterward, simply because part of their attention is still finishing the conversation.⁴

5. What "in place" would actually require

If the mechanism is re-entry, then the design target is not proximity. It is not having left. That is a stricter specification than it sounds, and most things that call themselves inline do not meet it. Four conditions, each traceable to a specific piece of the mechanism above:

1. The context is taken, not retyped. The question is raised by a selection. The tool already knows the sentence, the surrounding paragraph, and the document they sit in. If the user has to restate what they are asking about, the tool has already made them leave — restating requires holding the context in your head instead of on the screen, which is the condition Altmann and Trafton's model says is exactly when a goal's activation is most vulnerable to decay.¹

2. The answer is anchored to what raised it. It appears at the selection, not in a region reserved for answers. The distinction matters because an answer in a fixed region has to be found; an answer at the selection is already where your eyes are — the perceptual-availability condition Trafton et al. found necessary for cheap re-encoding.³

3. Nothing is displaced. The paragraph stays put. A layout that reflows to make room has moved you, because the sentence you were on is now somewhere else. This rules out both the docked panel that squeezes the column and the overlay that covers it, and it is the condition that keeps Monk, Trafton and Boehm-Davis's demand-scaling cost from applying: a conversation that does not displace anything is not competing with the primary display for the reader's perceptual attention in the way a docked panel does.²

4. Dismissal leaves no residue. When the answer goes away, the page is exactly as it was, and your eye is on the word you stopped at. This does not eliminate the attention-residue effect Leroy describes — that is a cognitive cost, not a layout one — but it prevents the tool from adding a second, purely mechanical residue on top of it.⁴

Meet all four and the interaction stops being a lookup and becomes something more like an annotation you happened to summon. Miss any one of them and you have built a faster tab.

6. The second question is the tell

Here is the part that changed my view of what this product is, and it came from using the thing rather than designing it.

An explainer that answers once is a dictionary. The first answer is almost never the end of the matter — it resolves the word and raises the next thing. What you actually want to say is "yes, but why does that apply here", and v1 could not take that question. The companion paper develops this finding properly, and grounds it in information-science research that anticipated it by four decades;⁵ it is enough here to note the consequence for this paper's own argument.

Now notice what the surfaces above do to follow-ups. In a panel or an overlay, asking again is cheap within the panel and expensive to relate back to the text. So people compensate by asking one large, well-formed question — because the switch has already been paid for, you may as well make it count. In place, the incentive inverts. Asking is so cheap that you ask three small questions instead of one big one, and the small ones are better: they are specific to the sentence, they build on each other, and each one is answerable.

The cost of asking determines the shape of what gets asked. That is the real consequence of where the answer appears, and it is not a matter of seconds saved. A tool that makes follow-up free produces a different kind of understanding than one that makes it merely possible.

7. Where in place is the wrong answer

The claim above is bounded, and the boundary is not a footnote — it is most of the map.

In-place answering is for comprehension in flow: you are committed to a document, and something in it is blocking you. That is a real and frequent situation, and it is badly served today.

It is the wrong design for everything else. Comparing four sources, holding a table open while you read against it, building up a synthesis over an hour, anything where the answer is the artifact you are working on rather than an unblocking — for all of these the panel is correct, because you want the answer to persist, to have room, and to be somewhere you can return to deliberately. A popover that politely vanishes is exactly wrong for work you intend to keep.

So this is not an argument that the panel is a mistake. It is an argument that the panel has been asked to cover a case it structurally cannot, and that the case it cannot cover is the common one.

8. Objections I take seriously

"This is introspection with footnotes." Fair, in part. The cognitive-science literature in §4 was not run on this task, and borrowing its mechanisms to explain a different phenomenon is an inference, not a replication. I think the inference is a reasonable one — the mechanisms (activation decay, perceptual availability, demand-scaled cost) are general properties of interrupted-task memory, not properties specific to the experimental tasks those studies used — but a reader should treat §4's application of them as argument, not as established fact for this case. §12 states the study that would settle it.

"The demand-scaling finding argues for making panels less engaging, not for eliminating them." A serious objection, and one the paper has not fully answered. If Monk, Trafton and Boehm-Davis's result is right, a sparse, low-demand panel might resume as cheaply as an in-place answer, which would undercut the case for in-place specifically rather than for "less demanding panels generally." I have not tested a low-demand panel against an in-place surface, and the honest position is that this objection identifies a real, untested alternative design this paper does not rule out.

"If the panel is wrong for comprehension-in-flow, why do search engines' assistant panels feel useful?" Because a search results page was already a tab in the sense this paper cares about — the reader had not yet committed to a specific document, so there is no primary reading context for a panel to compete with. §3's third category, the rewritten result, is this paper's own account of why that case is different and already well served.

"Flux has a commercial interest in this conclusion." Flux sells AnswerTab, and this paper argues directly against the design most of its competitors have converged on. That conflict is real and worth stating without softening. The mitigation available is the same one the companion paper offers: the paper's own conclusion — that the panel is correct for sustained comparison, synthesis, and any case where the answer is the artifact — concedes most of the reading-tool market to the design this paper is not selling. A paper written purely to sell an in-place product would not spend §7 arguing that the competing design is right most of the time.

9. What would settle it

The honest position is that this paper argues from mechanism and from use, not from measurement. The study that would settle it is not difficult to specify: same readers, same documents, same set of genuinely unfamiliar terms, and a comparison of comprehension and completion between an in-place answer and a docked panel — with follow-up question count recorded, since the prediction here is specifically that in-place raises the number of questions asked and lowers the size of each, and with interrupting-surface demand varied, since §4's borrowed mechanism predicts that a more demanding panel should cost more to resume than a sparse one.

The prediction is falsifiable and I have not run it. Stating that plainly is worth more than a number that would not survive being checked.

10. The gap between mechanism and measurement

No measured number exists for resumption cost in the specific case of a self-initiated, in-document lookup. Every cost claim in §4 is borrowed from studies of externally imposed interruptions and applied by inference, not by replication.

It does not claim the four conditions in §5 are jointly sufficient for a good reading tool, only that violating any one of them reproduces a specific, named failure. A tool could meet all four and still be bad for other reasons.

It does not claim panels are a worse design in general. §7 is explicit that for sustained comparison and synthesis work, the panel is the correct surface, and nothing in this paper argues otherwise.

And it does not resolve the objection in §8 about low-demand panels, which remains a real, untested alternative this paper's evidence does not rule out.

11. What would falsify this

Stated as commitments, in decreasing order of how badly each would damage the paper's claim.

A controlled study finds that resumption cost after a panel interruption does not scale with the panel's duration or engagement, contrary to Monk, Trafton and Boehm-Davis's finding for externally imposed interruptions. That would remove the mechanism §4 leans on hardest — that a better panel conversation is a more expensive one to return from — and the paper's sharpest claim would need to be withdrawn, not merely qualified.

A low-demand, minimal panel is shown to resume as cheaply as an in-place answer, for the same readers and tasks. That would not contradict the mechanism in §4, but it would falsify the design conclusion drawn from it — that the fix has to be in-place rather than merely lighter-weight — and §8's second objection would stand confirmed.

A measured study of self-initiated, in-document lookups finds the resumption cost is negligible compared to the cost measured for externally imposed interruptions. That would mean the borrowing in §4 was invalid — the two situations are not similar enough for the mechanism to transfer — and the paper's whole evidentiary base would need to be rebuilt from a study of the actual case rather than an adjacent one.

Readers shown both surfaces do not, in fact, ask more and smaller follow-up questions in the in-place condition. That would falsify §6's specific claim about the shape of what gets asked, while leaving the re-entry-cost argument in §4 intact; the two claims are separable and this would only take down one.

Appendix A — Evidence table

Sourced rows are cited in the footnotes. The unverified rows are measurements this argument needs and does not have; they are left empty rather than estimated.

Row 6 is the measurement every other row in this table is downstream of. It is the same open row carried in The Second Question's own Appendix A — neither paper has it, and I have not found a published study that does.

Appendix B — Citation register

CITATION-NEEDED — resumption lag for a self-initiated, in-document lookup. Appendix A row 6. The task-resumption literature cited in §4 measures externally imposed interruptions; a lookup the reader chooses to make mid-sentence is a related but distinct situation, and the directional claim in §4 does not depend on the two being numerically identical, only on the mechanism (activation decay, demand-scaled cost, perceptual availability) transferring. The widely circulated "minutes to refocus" figure sometimes quoted for interruptions generally is deliberately not used anywhere in this paper, because it was not measured under these conditions.

CITATION-NEEDED — comprehension/completion study, in-place vs. docked panel, controlled for demand. Appendix A rows 7–8. Specified in full in §9 and §11; not run by the author or, as far as I have established, by anyone else.

CITATION-NEEDED — low-demand panel vs. in-place resumption cost. Appendix A row 9. Raised as a serious open objection in §8 and restated as a falsification condition in §11; this paper's evidence does not rule it out.