I built an AI website builder for Webround. Then I killed it.

An AI website builder was created for the Webround platform, only to be abandoned later. The idea behind it was to integrate an AI assistant that could construct an entire site from a single user prompt. Webround supports both a visual no-code editor and a React environment, so the builder aimed to combine these capabilities using AI. However, the integration faced several issues that led to its discontinuation.

The AI builder utilized two mainstream large language models (LLMs): Gemini 3.5 Flash and Gemini 3.1 Pro Preview. The planning phase collected user requests, while the execution phase generated multi-step flows to implement those requests using typed and clearly defined operations. Despite this structured approach, the AI-generated website components were often mediocre in visual quality, with layouts, multi-device configurations, and styles frequently failing to meet expectations.

The real challenge lay in the AI's inability to produce visually effective layouts, as well as its tendency to miss crucial configurations. The model's output required extensive cleanup due to its inability to respect the component type systems. When the AI was tasked with generating React code directly, the results were vastly superior, producing beautiful landing pages with balanced colors and coherent copy.

React's extensive training enabled the AI to generate correctly configured components with visually appealing designs.

Despite these successes, the high costs associated with the AI-generated output proved to be a significant drawback. Generating entire sites resulted in expenses that could quickly accumulate, making the integration inefficient. As a result, the no-code generation strategy, though inconsistent and unsatisfying at times, became the more reliable and cost-effective approach. Ultimately, the decision was made to put the AI builder on hold, opting for Webround's React environment instead.