That format shift is exactly the mechanic. Summary-first structure isn't just reader UX — it's how you get your passage into the top-k chunks that actually reach the model's context window during RAG retrieval. The chunk that gets cited is usually the one that answers the query in the first 2-3 sentences, standalone, without needing surrounding context to make sense.
The news sites doing this probably aren't optimizing for AI citations intentionally — but they're converging on the same structure anyway because it works for skimmers and for models. Same underlying reason.
It works for me, because I am a skimmer. The opposite style I still see online, and I really dislike, is when they tease the answer for 5 or 6 paragraphs, only for you to then find out it wasn't all that interesting to begin with. It's a very weak form of click bait to my eyes.
The AI-optimized method lets you know right away if the article is going to be interesting to the human reader as well.
The news sites doing this probably aren't optimizing for AI citations intentionally — but they're converging on the same structure anyway because it works for skimmers and for models. Same underlying reason.