From Cluttered to Clear: Improving the Web Accessibility Design for Screen Reader Users in E-commerce With Generative AI
Yaman Yu, Bektur Ryskeldiev, Ayaka Tsutsui, Matthew Gillingham, Yang Wang · 2025 · ASSETS '25: Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663547.3746353
Summary
This paper explores how Generative AI (GenAI) can be used to automatically restructure the HTML of shopping websites to improve accessibility for screen reader users. Blind and low vision users face persistent barriers on e-commerce sites — complex layouts, inconsistent heading hierarchies, unclear labels, and visually-driven designs that prioritize aesthetics over navigability. While prior tools address isolated elements like alt text or product descriptions, they fail to tackle the structural and navigational challenges across entire webpages. The authors conducted a three-phase study: formative interviews with 6 screen reader users to identify barriers and coping strategies, development of a GPT-4o-powered Chrome browser extension, and evaluation through both automated accessibility audits and user testing with 15 blind and low vision participants. The extension offers two modes: Option 1 (Regenerated HTML) completely rewrites the page's HTML for optimal screen reader navigation, while Option 2 (Reorganized HTML Tags) modifies only the tags without changing the visual layout. The formative study revealed key challenges including inappropriate use of HTML tags, too many or too few headings, disorganized heading hierarchies, unclear labels on buttons and images, insufficient product descriptions, and difficulty comparing products across tabs. Participants already used GenAI tools like ChatGPT and Be My Eyes to describe product images and compare items, indicating readiness for AI-assisted accessibility solutions.
Key findings
Automated accessibility audits across Amazon, Nordstrom, and Mercari showed both GenAI versions consistently reduced Level A WCAG violations. For example, Amazon Product Pages dropped from 16 violations (original) to 1 (Option 1) and 3 (Option 2) in Google Lighthouse. Content integrity was preserved, with aggregated semantic similarity scores between original and regenerated pages frequently exceeding 95%. In the user evaluation with 15 screen reader users, Option 1 (Regenerated HTML) significantly outperformed the original website across all five evaluation metrics: overall user experience, heading clarity, content hierarchy understanding, efficiency in locating sections, and ease of accessing key features. Task completion times on product pages were significantly faster with Option 1 (p < 0.001). Participants rated the regenerated version 5.0/5.0 for screen reader browsing experience versus 3.14 for the original. Key improvements participants valued included logical reordering of sections to match linear screen reader navigation, addition of summary headings for page sections, enhanced navigation flexibility through multiple labeling (headings, links, and list items for the same element), and removal of unnecessary category headings that cluttered navigation. Option 2 showed moderate improvements but was constrained by preserving the original visual layout.
Relevance
This research demonstrates a promising application of GenAI for web accessibility that goes beyond fixing individual elements to restructuring entire page layouts for screen reader users. The approach is significant because it addresses structural accessibility issues that automated testing tools often miss — the mismatch between visually-oriented page designs and the sequential, heading-driven navigation patterns of screen reader users. The finding that complete HTML regeneration outperformed tag-only modifications suggests that meaningful accessibility improvements often require rethinking page structure, not just patching individual violations. Important caveats noted by the authors include the risk of content loss or hallucination during regeneration, the tool's limitation to static content (challenges with JavaScript-heavy dynamic sites), and the principle that such tools should support developers rather than shift accessibility responsibility onto end users. The work opens a significant research direction for using LLMs as accessibility remediation tools while highlighting the need for careful content integrity safeguards.
Tags: screen readers · web accessibility · generative AI · e-commerce · blind and low vision · HTML structure · automated remediation
Standards referenced: WCAG · ARIA