Headlines and links only, from ~50 public RSS feeds worldwide; all content belongs to the publishers. Outlets marked [state media] are government-owned or -controlled (TASS, RT, Xinhua, CGTN, Global Times, PressTV).
Refresh: feeds are fetched in 3 shards, merged, grouped and scored, and cached for about 15 minutes.
Grouping: headlines are stemmed, stopwords removed, demonyms mapped to countries (French→France), synonyms normalized (riots/unrest→protest), proper nouns treated as entities, and compared with TF-IDF cosine similarity; groups sharing their top entities with similar wording are merged.
Score = (8 × distinct outlets [state media counts 4] + 10 × (countries − 1) + 6 × (regions − 1) + Wikipedia points + Google Trends points) × recency.
Wikipedia points = min(15, 3 × log10(avg daily English Wikipedia views of the story's main name)), top 25 stories only; Wikimedia data lags several days, so this measures background interest.
Google Trends points = 12 per matching daily-trending list (US, UK, India, Australia), max 24.
Recency = 1 / (1 + hours since newest headline / 24).