Accessible Web Content Generation Using LLMs: An Empirical Study on Prompting Strategies and Template-Guided Remediation
Abstract
Web accessibility remains a persistent challenge, particularly for visually impaired users who rely on screen readers. This study investigates the potential of large language models (LLMs) to remediate accessibility issues only through structured prompt engineering without accessibility specialization. We evaluate GPT-4o and Gemini 2.0 Flash across 20 variants from two websites using different input formats (HTML, Markdown) and template-guided strategies. Outputs were assessed with automated tools, token efficiency metrics, and manual evaluations with experts and blind users. Results show an average Lighthouse score of 93.25, with WAVE errors reduced by 92.85%. Usability evaluation yielded an average success rate of 95.83% on completed tasks, with accuracy values reaching up to 0.86. GPT-4o demonstrated greater token efficiency, while Gemini produced more visually dynamic outputs. Certain violations persisted, confirming the need for human-in-the-loop validation. Overall, findings suggest that effectively guided LLMs can streamline remediation and foster more inclusive web experiences.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
These tools can support concrete accessibility tasks when properly configured.
“The complementary finding is that these tools can support concrete accessibility tasks when properly configured”
From A Protocol for Evaluating the Accessibility of AI-Generated Educational Materials: Prompt Configuration, WCAG-Derived Criteria, and Content Overload (2026) · cites this paper as [15]
AI-generated educational content can become substantially more accessible when the generating tool is deliberately configured.
“when generative AI tools are given explicit, operationalized accessibility instructions, through carefully engineered prompts [15,16], structured remediation pipelines [17,18], or context-aware generation conditioned on the surrounding material [19], measured compliance improves substantially compared to unconfigured use of the same tool.”
From A Protocol for Evaluating the Accessibility of AI-Generated Educational Materials: Prompt Configuration, WCAG-Derived Criteria, and Content Overload (2026) · cites this paper as [15]
The improvement observed under Condition B extends to educational content specifically findings already reported for web code, PDF tagging, and alt-text generation individually.
“The improvement observed under Condition B (Section 4.1) would extend, to educational content specifically, findings already reported for web code [15,16], PDF tagging [18], and alt-text generation [17,19] individually.”
From A Protocol for Evaluating the Accessibility of AI-Generated Educational Materials: Prompt Configuration, WCAG-Derived Criteria, and Content Overload (2026) · cites this paper as [15]
Cites
Still being checked.
Cited by
Sorted by citations
Large Language Models for Web Accessibility: A Systematic Literature Review
Aljedaani, W., & Mollik, R. H. (2026) · International Web for All Conference (W4A) · doi:10.1145/3800424.3800452 · Cited by 3 · see citing works
Vera-Amaro, G., Vera-Amaro, R., Mata-Rivera, M.-F., & Rojano-Cáceres, J. R. (2026) · Technologies · doi:10.3390/technologies14050287 · Cited by 2 · see citing works
Self-citationA Conceptual Model for Born-Accessible Web Accessibility Regeneration Using Large Language Models
Vera-Amaro, G., & Rojano-Cáceres, J. R. (2026) · Computers · doi:10.3390/computers15060343 · Cited by 2 · see citing works
Self-citationAccessibility Challenges for Blind Users Authoring in Content Management Systems: An Empirical Study
Vera-Amaro, G., & Rojano-Cáceres, J. R. (2026) · IEEE Revista Iberoamericana de Tecnologias del Aprendizaje · doi:10.1109/rita.2026.3681871 · Cited by 1 · see citing works
Self-citationAmado-Salvatierra, H. R. (2026) · arXiv · doi:10.48550/arxiv.2608.00749
“The improvement observed under Condition B (Section 4.1) would extend, to educational content specifically, findings already reported for web code [15,16], PDF tagging [18], and alt-text generation [17,19] individually.”
InfluentialInteraction with LLM-Based Systems: A Structured Review and Taxonomy of Mechanisms and Autonomy
Nejašmić, D., Mladenović, S., & Granić, A. (2026) · Applied Sciences · doi:10.3390/app16105001
Can LLMs Generate Accessible EPUBs?
Charaf, A., & Ribera, M. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-29653-5_11
Evaluating Large Language Models for Web Accessibility Remediation in Real-World Websites
Aljedaani, W., Gautier, K., Parthasarathy, P. D., & Clemmer, C. (2026) · International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS) · doi:10.1145/3797867.3841243
Silva, G. A. A. E., Silveira, P. H. D. A., Colombo, L. G., de Souza, V. F., Lima, J. C. S., & Freire, A. P. (2026) · Brazilian Symposium on Multimedia and the Web · doi:10.5753/webmedia.2026.30915
- Authors
- Vera-Amaro, G., & Rojano-Cáceres, J. R.
- Affiliations
- Universidad Veracruzana (MX)
- Venue
- IEEE Latin America Transactions
- Published
- 2025-12
- Volume
- 23
- Issue
- 12
- Pages
- 1230-1239
- ISSN
- 1548-0992 (ISSN-L)
- OpenAlex
- W4416010301
- Cited by
- 7
- References
- 0
- Citations by year
- 2026: 7
- FWCI
- 42.25
- Type
- journal article
- Language
- en
- Open access
- closed
- Locations
- IEEE Latin America Transactions (publishedVersion)
- 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 Applications and Data Management
- Keywords
- web accessibility, large language models, prompt engineering, screen readers, visually impaired users, GPT-4o, Gemini, waves, lighthouses, token efficiency, human-in-the-loop validation, inclusive web design
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.
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
