Against: controversies, limits and failed assumptions

This file gathers the evidence against showing text folded or nested: null results, disputed studies, and practice that a study contradicts. Terms are defined in frame.

Whether folding helps the reader

On the easiest retrieval task, the fully expanded list and the multi-pane view beat expand/contract (Chimera 1994, p < .01).
On the task of finding the item that covers three topics, the fully expanded list was fastest, not significantly: 46.1 s against 85.9 s for expand/contract and 55.0 s for multi-pane (Chimera 1994).
The authors explain it: the target sat near the start, and its chapter title did not predict its contents, so readers of the folded views opened the wrong chapters.
Chimera 1994 credits part of the expand/contract gain to animation, so folding and animation are confounded.
Shrinking text around the focus cut the time to read and write an essay by about 16%, with grades no better than linear text, and lowered incidental learning (Hornbæk 2003, pp. 140 to 142).
A short first layer made readers faster and less accurate on what it left out; few went further (McDonald 2009, section 4).
Sections open on load drew more reading time than folded ones: 190 s against 146 s at the 90th percentile (Wikimedia 2016).

Whether an overview or a fisheye helps

Readers preferred an overview and were not more accurate with it; on a map built in levels they were faster without it (Hornbæk 2002).
Switching between overview and detail cost time and effort (Hornbæk 2002; Cockburn 2009, section 8).
On large trees, pan and zoom beat a distorting fisheye, and an overview did not change task time (Nekrasovski 2006, through Cockburn 2009).
Distortion-based techniques have largely failed to show performance gains (Cockburn 2009, section 8).
Preference and performance part ways in several of these studies, so a preferred design is not shown to be a better one.

Whether hypertext structure is the lever

Effects vary strongly between studies, and cognitive traits gave effects not clearly different from zero (Chen 1996, secondary).
Method flaws limit every conclusion about hypermedia and learning (Dillon 1998).
Several studies adapting presentation to readers' traits found no significant differences (Brusilovsky 2001).
Pure hierarchy is not always best: hierarchy with cross-links did better in McDonald 1998, while added network links gave nothing in Mohageg 1992.

Whether trees of summaries help a model

Passages kept in their original order matched or beat RAPTOR at matched token budgets; RAPTOR without its summaries beat RAPTOR with them (Laitenberger 2025, sections 3 and 4).
RAPTOR's own ablation used a 3B reader limited to 400 tokens, a setting that favours compression (Laitenberger 2025, section 4).
Within RAPTOR, searching all layers flattened beat descending the tree layer by layer (Sarthi 2024, section 3); the layer ablation in the main text is one story (Table 8).
About 4% of RAPTOR's summaries held minor hallucinations (Sarthi 2024, section 3).
The gain of a summary tree may come from navigation, not from the summaries (Jayraj 2026, secondary).
A second level of disclosure for agents never helped and sometimes broke accuracy (He 2026, Table 1).
Merging summaries up a hierarchy is more coherent and loses detail against a running summary (Chang 2024, secondary).

Weaknesses of the strongest studies

Furnas 1986: 20 subjects, one task of order judgments, one taxonomy.
Chimera 1994: 41 novices, one contents of three levels, one workstation of its time; the title calls it exploratory.
McDonald 2009: an online sample entered for a lottery, one domain (privacy policies), six companies; 7 of 756 were removed for clicking through.
Wikimedia 2016: not a strict A/B test; about 0.05% of users per group, in 2015-12 and 2016-01; it measured time, not understanding.
He 2026 and Jayraj 2026 are preprints; He 2026 uses one benchmark, ∞Bench. Laitenberger 2025 appeared at EMNLP (2025).
The code and tree studies were read only through Cockburn 2009; Hornbæk 2003 was read in full afterwards.

Limits of the practice evidence

Nielsen 2006b gives the two-level limit with no data; its owner sells consulting.
GOV.UK 2026a and 2026b report research findings without numbers or samples.
Wang 2023 (example only, 9 of 20) says folded content is missed; it is one company article.
PageIndex reports 98.7% on FinanceBench itself, with no independent evaluation linked (VectifyAI 2026, example only).
Stars and downloads measure interest, not reading.
Baymard's guideline on collapsed sections is paywalled and was not read.

What the Q/A threads add against

Folding a long block hides it instead of fixing it; regions are called an antipattern (SESE 2011o, accepted answer of 351 votes).
A folded block has no name, scope or interface, so it is not a summary of what it hides (SESE 2014e).
Do not hide answers to save scrolling; readers scroll when they expect something worth finding (UXSE 2012b).
These are opinions; they agree with McDonald 2009 and Wikimedia 2016 on the cost of hiding.

Failed or missing replications

This search found no replication of Furnas 1986's experiment or of Chimera 1994.
RAPTOR's gains did not hold under matched budgets and stronger readers in two later comparisons (Laitenberger 2025; Jayraj 2026, secondary).
No replication of He 2026 was found; it appeared in 2026-07.