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.

While accessibility research often emphasizes the experiences of blind users as consumers of online content, far less attention has been given to their role as content authors in Content Management Systems (CMS). For blind users, the accessibility of CMS platforms is fundamental to ensure efficient navigation, content creation, and overall usability. This study investigates the accessibility challenges faced by blind authors when working with WordPress as a representative CMS, identifies barriers related to navigation and content management, and offers insights to improve inclusive design. A mixed evaluation strategy was employed, combining automated testing tools with a manual assessment based on the barrier walkthrough method. Three blind participants and one sighted participant completed representative CMS tasks—such as navigation, content editing, and menu creation—using NVDA and JAWS screen readers. In addition, an expert evaluation analyzed the severity of the identified barriers. Findings indicate that block-based and drag-and-drop interfaces remain highly inaccessible for screen reader technologies, while AI-powered image descriptions help reduce dependence on sighted assistance. Results also highlight a clear mismatch between pages considered accessible by automated tools and the barriers observed in manual testing, underscoring the limitations of automated evaluations. Overall, the study stresses the need for CMS platforms to improve navigation consistency, reduce dependence on dropdown submenus, and integrate AI support to better address the needs of blind authors.

Recent research on web accessibility has explored the use of artificial intelligence (AI), particularly large language models (LLMs), to support accessibility remediation. However, the field lacks a theoretical perspective explaining how LLMs can be integrated to systematically support this process. This study proposes a theory of LLM-assisted web accessibility remediation. It is built through the integration of qualitative evidence, prior literature, accessibility standards, and empirical studies on LLM-based remediation. The resulting theory provides an explanatory framework describing how LLMs can assist web accessibility remediation through iterative cycles of analysis, transformation, and validation, and identifies key factors including prompting strategies, input representations, and validation mechanisms. This work provides a conceptual foundation for understanding and systematically studying LLM-assisted accessibility remediation, and supports both research and practice by guiding future studies and informing the design of models, methods, tools, and accessibility engineering practices.