LLMs attempt that, but their success is mixed - sometimes the summary is very bad.

Imagine how long it would take to load a page of HN search results, or the table of contents of an online book, or a page with a lot of dead links, or a page with links to a whole bunch of non-textual content that it has to figure out is non-textual.

You don't need to load where the link is going. The base line is the context a sighted person would have when they see the "click here" link.

While that would work in many situations, I think the number of edge cases would be too great to make it reliable as a primary mode for communicating meaning in an accessible interface. As an alternative assistive feature? Sure. Something to rely on? I'll believe it when I hear it.