I think using vision models for browsing is the wrong approach. It is the same as using OCR for scanning PDFs. The underlying text is already in digital form. So it would make more sense to establish a standard similar to meta-tags that enable the agentic web.