There's a wave of advice about making content "machine readable" — restructure everything, chunk it, mark it up, write for the model.
The premise deserves a closer look. Language models are good at reading. Making sense of unstructured text is the one thing they demonstrably do; it's why anyone is talking about them at all. Your prose is not the bottleneck.
The bottleneck is earlier. Inference is expensive and context windows are finite, so no model reads your whole site. Something decides which fragments are worth its attention — crawling, parsing, extraction, indexing, retrieval. Plain maths and heuristics, running long before any model sees a word.
That layer has a name. It's search. We've been optimizing it for thirty years.
So the real questions stay unglamorous: can a crawler reach the page, does the markup parse, does rendering hide half the content, do the retrieved fragments hold up out of context, are facts consistent wherever they appear. None of this is new work. It's the same discipline with a new reader at the end of it.
If anything has changed, it's that the reading got better. The finding still works the way it always did.