WebGenie: An LLM-Based Method for Automated Website Design Recommendations

Abstract

In today’s fast-paced digital environment, creating an effective website has a profound impact on brand presence and business operations, but it also brings increasingly complex demands for web design, which may consume significant time and resources. At the same time, emerging AI tools like ChatGPT demonstrate impressive generative capabilities but still require users to master prompt-crafting and often fall short in conveying a brand’s emotional identity when creating a web design.

The study aims to explore how Large Language Models (LLMs) and Prompt Engineering techniques can be leveraged to assist users in generating personalized web design recommendations based on their requirements, thereby improving the efficiency of the design process and providing additional references for subsequent development.

To achieve this goal, we propose a guided LLM-based framework, WebGenie, that divides the web design process into four phases which are Requirement Analysis, User Interface Design, Web Development, and Quality Assurance, and we design corresponding prompts and task structures to guide the LLM effectively. Additionally, we integrate the relationship between visual design elements and website personality into the prompt design to enhance the emotional expression and style alignment of the generated outputs.

To ensure the best possible performance, we adopt an automated LLM evaluation process to decide the most suitable model for each phase. In terms of method’s performance evaluation, we compare the proposed method with the existing LLM tool, ChatGPT, to evaluate its effectiveness in generating web design recommendations. The experimental results show that our method improves design efficiency by approximately 48% compared to using ChatGPT alone, and has demonstrated practical potential in supporting early-stage ideation and team communication.

Overall, the study not only highlights the potential application of LLMs in web design tasks but also presents an extendable and practical framework that offers a technical foundation for future related research and system development.