Possibly because a lot of “AI-company scraping” isn't traditional scraping (e.g., to build a dataset of the state at a particular point in time), its referencing the current content of the page as grounding for the response to a user request.