"Every Website Is a Puzzle!": Facilitating Access to Common Website Features for People with Visual Impairments
Natã M. Barbosa, Jordan Hayes, Smirity Kaushik, Yang Wang · 2022 · ACM Transactions on Accessible Computing · doi:10.1145/3519032
Summary
This paper addresses a fundamental challenge for screen reader users: while sighted users can rely on visual conventions to find common website features (log in at top right, contact at bottom), these shortcuts are inaccessible to users with visual impairments who must linearly parse page content. The researchers developed "CrowdIntent," a browser extension that enables command-based access to common website features like "log in," "sign up," "find store," and "contact." The system uses three mechanisms to map user commands to target pages: (1) machine learning models (SVM classifiers) trained to classify hyperlinks leading to common features based on link text, URL structure, and page position; (2) crowdsourced human input where users can mark pages as correct and request help from other users; and (3) a "feeling lucky" background search as a fallback. The ML models achieved F1 scores ranging from 0.74 to 0.95 across different features. The key paradigm shift is from "passive listening" (users linearly consuming page content to find what they need) to "active solicitation" (users directly requesting desired features via commands). One participant captured the problem succinctly: "Every website is a puzzle" that users must figure out individually, despite common features serving identical functions across sites.
Key findings
The user study with 15 participants (9 with visual impairments, 6 without) demonstrated significant improvements. The System Usability Scale score was 81.2 for the command-based system versus 52 for conventional browsing. For blind participants specifically, SUS scores were 88.6 for the system versus 43.6 for conventional methods. Task completion times were dramatically reduced for users with visual impairments: blind participants averaged 42 seconds per task with the system versus 112 seconds with conventional browsing. The difference was even more pronounced when using search engines—users with visual impairments spent 55.6% of their time navigating search results in the baseline, reduced to 36.3% with the system. 80% of participants preferred the command-based system. Key benefits cited were consistency ("since no two websites are the same... you just type in login and it puts you right where you need to be"), fewer steps, and reduced trial-and-error. The system was deemed most useful for unfamiliar, cluttered, or infrequently visited websites—and notably, users without visual impairments also found it helpful in these scenarios, supporting universal design principles. Participants raised concerns about privacy when requesting human help, handling of unsupported commands, and reduced discoverability of website features when bypassing normal navigation.
Relevance
This research has significant implications for how we think about web accessibility. Rather than focusing solely on making page content accessible, the authors argue for task-level assistance—helping users accomplish goals rather than just navigate content. This "active solicitation" paradigm could fundamentally change how assistive technologies work. For practitioners, the findings suggest that websites could benefit from a "consistency infrastructure"—metadata or standards that map common features to standardized commands, enabling personal assistants and browsers to interact with sites on users' behalf. The machine learning approach demonstrates that this doesn't require websites to change; patterns can be learned from existing hyperlink structures. The universal design aspect is noteworthy: a system designed for users with visual impairments also benefited sighted users on unfamiliar or cluttered websites. This challenges the notion that accessibility features are only for disabled users. Limitations include that the ML models support a limited set of intents, crowdsourced help raises privacy concerns, and the system was only tested with typed commands (not voice). Future directions include integration with personal assistants like Siri and Alexa, support for voice commands, and website-specific features beyond the common set.
Tags: visual impairment · screen readers · web accessibility · machine learning · crowdsourcing · browser extension · universal design · intelligent personal assistants