Current advances in web technology and large language models (LLMs) enable us to have a generative web where specific components of websites are dynamically generated on-the-fly by LLMs executed within the user’s browser. While this allows rich personalization, it also raises concerns about balancing customization with privacy and data security. This paper suggests a hybrid client–server architecture for dynamic website generation using LLMs, focusing on personalization, efficiency, and privacy. We use on-device LLM inference via WebAssembly/WebGPU to enable personal data to stay local and examine how this reduces privacy concerns and latency. The architecture is structured in five modular layers, Perception, Reasoning, Layout, Interaction, and Privacy, that decouple system concerns, enabling at runtime adaptive generative web systems, generating personalized user layouts based on local user context, without exposure of personal information to the server-side. We show a proof-of-concept implementation based on Transformers.js and WebGPU for in-browser LLM running and IndexedDB for model and user data caching. Experiments and benchmarking indicate that our method can maintain  80% of native performance for in-browser inference of LLMs with acceptable Time-to-Interactive (TTI) with caching and quantization. We contrast user experience and system metrics (TTI, CPU and network utilization) against current web stacks (e.g. React), demonstrating that generative LLM-driven pages can be competitive in responsiveness while offering adaptive personalized content. Lastly, we propose directions for future work such as on-device fine-tuning for personalization, hybrid cloud–edge inference for scalability, ethical considerations of AI-generated content, and the transformative potential of genuinely adaptive generative web systems.