Humans interpret text as rendered by the browser. That includes a lot of visual information and even text that only becomes visible in response to user interact

Humans interpret text as rendered by the browser. That includes a lot of visual information and even text that only becomes visible in response to user interaction. For an LLM (or an AI in general) to do what humans do, it would have to process the rendered output and interact with the site like we do (such as clicking on a dropdown to make the available options visible). The same information represented as semantic HTML (i.e text) is far cheaper to process for an AI. I think cost is a key consideration in everything AI. If it's too expensive then it won't happen, even if it could theoretically be done.