In a small preprint, adults who use screen readers visited more distinct documents while exploring linked pages than while using an AI question-answering tool. The Document Interface, or DI, averaged 11.44 unique document visits, compared with 9.31 for the Question-Answer Interface, or QAI, a reported 22.9% increase. The comparison was reported with p = 0.03, but confidence intervals were not provided.
That was only one part of a mixed picture. QAI exploration graphs were more tightly connected, and its mental-model graphs were more interconnected after errors were removed. DI elicited more topics and connections, recorded fewer concept-map errors, and showed an advantage in a later decision-based application task.
Two routes through the same information
The study involved 16 participants, all at least 19 and regular users of at least one screen reader. The same participants used both interfaces.
The researchers asked how interface type affected browsing, the integration and application of document knowledge, and interface preference.
Each participant completed two 90-minute sessions on separate days. Interface order and fictional-world order were counterbalanced across participants, so the sequence varied.
One condition offered direct exploration of a document collection through hyperlinked documents. The other used a language model to mediate question answering over the collection.
Each fictional-world set contained 25 documents, including an index. Eight Solana documents and six Dominion documents were diagrams representing spatial or conceptual aspects. Across four phases, the researchers logged exploration, collected concept maps, tested knowledge in a decision-based scenario and conducted semi-structured interviews.
When coverage and connectivity diverged
Participants also made concept maps, diagrams showing the topics they had taken from the documents and the links between them. DI elicited 67% more topics on average than QAI: 9.44 compared with 5.63, with p = 0.02. It also elicited 38% more connections, 20.69 compared with 14.94, with p = 0.01. Subtopic counts were similar, at 9.63 for DI and 8.69 for QAI, with p = 0.55.
DI maps contained an average of 3.31 raw errors, compared with 6.00 after QAI. Their error rates were 0.12 and 0.26, respectively. The paper reports 45% fewer raw errors and a 54% error-rate difference for DI, with both comparisons at p < 0.01.
The QAI exploration network was more compact. Graph density, roughly a measure of how closely items are linked, averaged 0.21 for QAI and 0.13 for DI. QAI's directed diameter, or longest one-way route through the network, was 4.94 compared with 7.75, while its average shortest path was 2.44 compared with 3.77. Counts of transitions between documents were not conclusively different, with all p-values above 0.05.
The same pattern appeared in the concept-map structure. After errors were removed, QAI mental-model graphs were more interconnected than DI graphs, with every reported graph comparison at p < 0.04.
What users thought did not always match
When participants applied what they had learned in a decision-based scenario, the authors observed an advantage for DI. Responses mentioned more topics and subtopics and contained fewer errors than responses after QAI.
Participants' impressions did not fully line up with measured exploration, learning or answer confidence. The narrative account reported an even overall preference split, with eight participants favoring each interface. Interviews identified agency and control, along with effort and cognitive load, as the two main themes shaping those preferences.
Taken together, the findings do not produce a simple winner. DI was associated with broader coverage, broader concept maps, fewer map errors and an advantage in the application task. QAI was associated with more connected exploration and more interconnected mental-model graphs. The reported eight-to-eight preference split reinforces the study's central tension: measures of what users covered and understood did not always match what they preferred.
A bounded result
The evidence stays close to the experiment itself: 16 adult screen-reader users, two fictional 25-document worlds and two 90-minute sessions per participant. Those conditions do not answer how the interfaces would perform during prolonged use or across real-world document collections.
The paper is an arXiv version 1 preprint dated 26 Aug 2026. The research was supported by the University of Victoria, NSERC DG 2020-04401 and the Singapore Ministry of Education Academic Research Fund Tier 1 grant, Project ID 24-SIS-SMU-039.
Paper data and sources
Original title: Q&A or Document-Based? The Effects of Interface Type on How Screen Reader Users Access Interconnected Documents
Authors: Colleen F. Cipriano, Yichun Zhao, Miguel A. Nacenta et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: 10.1145/3797867.3829036
Original paper · Full text