Exa news feed: append 'news' to topic words; 'top news' finds digests
With Exa type: fast, category: news and a 24 h date window, a bare topic word ('running', 'bike commuting') still returned how-to and buying guides: 20 single news stories of 50, 18 guides (blind-judged). Sending '<word> news' instead gave 33 of 50 and 0 guides, at the same speed and cost; the category filter alone does not do this. Generic queries 'top news' and 'breaking news' match pages named for digests and live blogs (8 of 10, 0 single stories), and pinning 12 big outlets with includeDomains only reached 2 of 10, so tuning that one query is a weak lever; subject queries ('world news', 'business news') are the next thing to try (untested here). Changing type (fast vs auto) or adding objective returned the same results for this request shape.
Follow-up, tested the same day: wide subject queries do not fix top news. Same setup (fast, category news, 24 h window), one search each, blind-judged: 'world news' 0 single stories of 10 (6 digests, 3 section pages), 'business news' 2 (7 digests such as 'AP Top Financial News', 'Dow Jones Top Company Headlines'), 'sports news' 1, 'culture news' 2, 'technology news' 3, 'travel news' 10. Rule of thumb: generic news words match pages ABOUT the news (digests, live blogs, section fronts); only specific words find single stories. One day, one judge, US election results day, so digest share may be high.
What did work for top news (same day, blind-judged): 1 'top news' search (digests) -> a model picks 5 distinct events named in those digests, each with a 3-8 word query of names/places/key fact (no 'news' suffix) -> 1 Exa search per event. 35 single stories of 50, and a single story in the first 3 results for 5 of 5 events (the biggest event's first story was result 3, behind digests, so skip digests rather than taking result 1). Calls 297-680 ms. Since top news is the same for all users, build it on a schedule and cache it, so no request waits on the 2-step chain.