Two records are kept apart: most adopted (what people use) and best evidence (what a study measured). Terms are defined in frame; sources are in review.
Sub-question
Most adopted
Best evidence
Measured how
Q1 folded and expandable text
accordion and details (WHATWG 2026; GOV.UK 2026a, 2026b)
Chimera 1994; McDonald 2009; Hornbæk 2003
timed retrieval on a 1,296-item contents, 41 novices; accuracy and time on privacy policies, 749 adults; essays and questions on scientific papers
task time, accuracy, preference; no technique wins across tasks
Q4 levels in one view
three heading levels (reader-ergonomics shallow outline); at most two disclosure levels (Nielsen 2006b, example of practice); no accordion inside another (GOV.UK 2026a)
for menus, two levels beat three (file-structure, linked); for agents, one level of disclosure beat two (He 2026)
no study of reading an outline in an outliner was found
Q6 trees of summaries for models
GraphRAG and PageIndex repositories (counts below)
Laitenberger 2025; He 2026; Sarthi 2024
matched-token comparisons on QuALITY, NarrativeQA, ∞Bench
Q7 hidden content left unread
sections folded by default on mobile Wikipedia (Wikimedia 2016)
McDonald 2009; Wikimedia 2016
accuracy on topics outside the first layer; reading time with sections open or folded
Adoption counts, descriptive, taken 2026-10-29
Count
Value
Caveat
stars, Logseq (outliner, created 2020)
45,407
interest, not use (GitHub 2026c)
stars, SiYuan (block editor, created 2020)
46,984
same
repositories tagged outliner
137
self-tagged
repositories tagged org-mode
1,271
self-tagged
downloads, Obsidian Outliner plugin
1,471,149
of 9,606 community plugins; the largest has 8,763,578 (Obsidian 2026)
downloads, Obsidian Zoom plugin (a bullet as the root)
113,488
same
stars, RAPTOR reference code (created 2024)
1,775
interest, not use
stars, GraphRAG (created 2024)
36,477
a graph of summaries, not a tree
stars, PageIndex (created 2025)
39,880
its benchmark figure is the vendor's own (VectifyAI 2026, example only)
Counts come from the public GitHub API and Obsidian's published plugin statistics, read without a key.
No count of web pages using details was found: the Web Almanac markup chapters of 2021 and 2024 do not report it.
Upper bounds (achieved)
A fisheye of a regular tree is a logarithmically compressed view of it, computable in time proportional to the view (Furnas 1986, section 3).
On order judgments in an unfamiliar hierarchy, correct answers rose from 52% with two flat windows to 64% with one fisheye and 75% with two (Furnas 1986, 20 subjects, pages 20 to 21).
On a three-level contents of 1,296 items, expand/contract and multi-pane views were both significantly faster than the fully expanded list on four of nine tasks, and expand/contract on a fifth (Chimera 1994).
A hypertext book with an expandable contents beat print on search and essays (Egan 1989, secondary).
A tree of summaries raised QuALITY accuracy to 82.6% with GPT-4, against a previous best of 62.3% (Sarthi 2024, Table 7).
Retrieval in document order reached 93.1% on ∞Bench at 30K tokens (Laitenberger 2025).
One level of disclosure kept agents at 0.46 on English open questions over 20 books, where raw reading fell to 0.26 (He 2026).
Lower bounds and limits (not shown, or shown not to hold)
No study found varies depth in storage while holding depth in view fixed; whether storage depth matters to a reader is not measured.
No study found shows a document more than three levels deep in one view to readers.
A second level of disclosure never helped agents and cut accuracy in some cells, from 0.9126 to 0.6398 (He 2026, Table 1).
With strong long-context readers, the summaries in a tree did not help: the same pipeline without them did better (Laitenberger 2025, section 4).
Distortion-based fisheyes have rarely shown performance gains over the usual alternatives (Cockburn 2009, section 8).
A view of at most k lines cannot hold the whole ancestral line of a node deeper than k, so a capped view of an unbounded tree must elide ancestors. This follows from the definitions; it is not a study.
Folded or layered text was read less. Readers answered worse on what the first layer left out, and the authors infer few went past it (McDonald 2009). Folded sections drew less reading time (Wikimedia 2016).