Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index; 5 of the 20 sources its claims cite were checked for support.

From Cluttered to Clear: Improving the Web Accessibility Design for Screen Reader Users in E-commerce With Generative AI

Open access
Yu, Yaman, Ryskeldiev, Bektur, Tsutsui, Ayaka, Gillingham, Matthew, Wang, Yang · 2025 · Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility

Abstract

Figure 1: From left to right: (a) the original website, (b) Option 1 (Regenerated HTML), and (c) Option 2 (Reorganized HTML Tags).Option 1 is generated using a GenAI-powered extension that rewrites the HTML to enhance accessibility, while Option 2 reorganizes the original HTML tags to address accessibility issues without altering the visual design.This example highlights the changes made in both versions compared to the original website.Both versions simplify navigation by removing single-category headings, reducing clutter and minimizing fatigue for screen reader users.

Overview

CoChat Summary

TL;DR

The paper presents a GenAI-powered Chrome extension that uses GPT-4o to either regenerate or reorganize e-commerce HTML to improve screen reader navigation. In formative interviews and a user evaluation, the regenerated HTML version improved accessibility audit scores, content similarity, and user ratings over the original site.

Key findings

  • Automated audits across Amazon, Nordstrom, and Mercari showed both LLM versions reduced WCAG Level A violations; for example, Amazon Product Page Lighthouse violations dropped from 16 in the original to 1 in the regenerated version and 3 in the reorganized version, and Nordstrom Home SortSite violations dropped from 5 to 1 in both versions.
  • Regenerated pages frequently exceeded 95% Aggregated Semantic Similarity; the lowest reported score was Nordstrom Home Page at 91.60%.
  • In the user evaluation with 15 screen reader users, the regenerated HTML version received an average overall experience rating of 5.00, the reorganized tags version 4.57, and the original website 3.14; 13 participants reported the regenerated HTML version provided the best experience.
  • Statistical tests showed the HTML version outperformed the Original on all five Likert questions, while the Tag version showed significant improvement only in overall user experience (Q1) compared to the Original; HTML vs. Tag differed significantly only in Q1. Task completion times differed significantly across versions (Friedman χ²(2)=30.00, p<0.001), with Option 1 significantly faster than both Original and Option 2 (W=0, p<0.001), and Option 2 significantly faster than Original (W=0, p<0.001).

Methods

  • Three-phase study: formative semi-structured interviews with 6 blind or low vision screen reader users (ages 25–54) about shopping barriers and think-aloud browsing on Mercari; development of a Chrome extension using GPT-4o (temperature=0.2, max_tokens=16,384, top_p=1, frequency_penalty=0, presence_penalty=0) with two modes—regenerated HTML (Option 1) and reorganized HTML tags (Option 2); and a within-subjects user evaluation with 15 screen reader users (ages 18–44) on Mercari, counterbalanced across original, Option 1, and Option 2, using think-aloud and a 5-question Likert survey.
  • Automatic accessibility audit on Amazon, Nordstrom, and Mercari home, search results, and product pages using Google Lighthouse, SortSite, and AChecker; content integrity measured via Aggregated Semantic Similarity (cosine similarity of screen-reader-accessible elements, with a 90% acceptance threshold).
  • Analysis included deductive/inductive thematic analysis for interview data, and Wilcoxon signed-rank tests with Bonferroni correction (α adjusted=0.0167) plus Friedman tests for Likert and task-completion comparisons.

Limitations

  • Stated in the paper: the formative sample was fairly homogeneous (all from similar demographics); the user evaluation focused on desktop use and did not account for mobile or tablet navigation; generalizability to dynamic, JavaScript-heavy platforms remains an open challenge; LLM-driven restructuring carries risks of content loss, subtle omissions, and hallucinations; and Option 2 was limited by its inability to alter the original layout or content order.
  • Evident from the text: samples were small (6 formative participants, 15 user-evaluation participants) and the user evaluation was limited to one website (Mercari); per-page processing required 1–5 minutes and around 220,000 tokens per session, costing $0.50–$2.20 per page, which may constrain deployment.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3663547.3746353

Cites

62 of 62 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY · From the PDF
Figure 2 Methodology OverviewNote that participants of the user evaluation (step ❹ ) do not overlap with those in steps ❶ .
Figure 2 Methodology OverviewNote that participants of the user evaluation (step ❹ ) do not overlap with those in steps ❶ .CC-BY · From the PDF
Image 3
CC-BY · From the PDF
Figure 3 Option 1 pipeline overview, prompting GPT-4o model to completely rewrite HTML code
Figure 3 Option 1 pipeline overview, prompting GPT-4o model to completely rewrite HTML codeCC-BY · From the PDF
Image 5
CC-BY · From the PDF
Image 6
CC-BY · From the PDF

Verification

Claims the paper makes

3 Valid · 3 Not valid · 10 Not checked · 125 Own results and interpretations

53 more claims

Create a free account to see all 141 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 43
  • Probably not a substantial claim. 53
  • Borderline: one of the tests for a substantial claim is unclear. 1
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 54 references

Research works printed but not linked: 1

  1. Smirity Kaushik, Natã M Barbosa, Yaman Yu, Tanusree Sharma, Zachary Kilhoffer, JooYoung Seo, Sauvik Das, and Yang Wang. 2023. { GuardLens} : Supporting Safer Online Browsing for People with Visual Impairments. In Nineteenth Symp. on Usable Privacy and Security (SOUPS 2023). 361–380.

Cited by

See all 8

Metadata

Authors
Yu, Y., Ryskeldiev, B., Tsutsui, A., Gillingham, M., & Wang, Y.
Affiliations
  • University of Illinois Urbana-Champaign (US)
  • University of Tsukuba (JP)
  • Mercari R4D, Tokyo, Japan
Venue
Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility
Publisher
ACM
Published
2025-10-22
Pages
1-19
OpenAlex
W4415445080
Cited by
8
References
54
Citations by year
2026: 8
FWCI
9.80
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS) (publishedVersion, cc-by)
Topics
  • Social Sciences › Human Factors and Ergonomics › Digital Accessibility for Disabilities
  • Physical Sciences › Computer Science › Human-Computer Interaction › Usability and User Interface Design
  • Physical Sciences › Computer Science › Information Systems › Web Data Mining and Analysis
Keywords
web accessibility, screen readers, e-commerce platforms, generative AI, user fatigue

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: From Cluttered to Clear: Improving the Web Accessibility Design for Screen Reader Users in E-commerce With Generative AI. https://app.cochat.ai/papers/10.1145/3663547.3746353
The paper
Yu, Y., Ryskeldiev, B., Tsutsui, A., Gillingham, M., & Wang, Y. (2025). From Cluttered to Clear: Improving the Web Accessibility Design for Screen Reader Users in E-commerce With Generative AI. Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility. https://doi.org/10.1145/3663547.3746353

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.