I see where you’re coming from—AI tools have become much better at extracting information and recovering structure from PDFs. But, in my opinion, that does not make HTML versions any less necessary: they are still much easier to make fully accessible, mobile-friendly, searchable, and machine-readable without requiring tools to infer the document’s structure from its rendered layout. Many accessibility guidelines currently recommend publishing in HTML wherever possible, using PDF when a fixed layout is genuinely needed.
I’d argue that the long-term solution is to preserve the Typst source and its dependencies as the reproducible record, then generate a stable PDF and a structured HTML version from the same source.