To scale such an approach you could have the LLM generate JS to walk the DOM and extract content, caching the JS for each page.