For standard responses you're not wrong, but increasingly there are a lot of LLMS that are essentially doing RAG against search results. For example this is I believe how Kagi works, and Google AI overviews. I have increasingly seen Claude and chat GPT also doing the same thing where instead of answering a question from the knowledge Bank they will do a web search and cite the responses that were used.