Methodology
Where the data comes from, how often it updates and how each number on the site is calculated.
Sources
- News — the GDELT Project. An open database that indexes the world's online media every 15 minutes. We ingest two streams: English and translated (about 65 languages, machine-translated by GDELT). Roughly 300,000 articles a day.
- Telegram. Public channels by topic, added in stages, starting with Russian-language channels.
- We store each article's headline, link, outlet, language, date and labels. We do not store or republish article text — we link to the original.
Updates
New articles reach the database 15–20 minutes after publication. Topic and trend pages are recalculated every few minutes, the topic directory hourly. Topic pages cover the last 30 days.
How we measure
- Mentions per 10,000 articles. The total number of articles varies by day, so we show dynamics as a share: topic articles per 10,000 articles that day. That way a topic's growth is not confused with growth of the whole stream.
- Tone. GDELT's dictionary-based score from −10 (very negative) to +10. News normally runs slightly negative; on the site below −1.5 is “negative”, above +1 is “positive”.
- People and organisations. Extracted from the text automatically. A page appears once at least 25 different outlets covered the person or organisation in a week, which filters out one-off mentions and sites repeating one story hundreds of times.
- Today's trends. We compare the last 24 hours with the previous week's daily average. Only topics with at least 80 articles and 25 outlets in a day count.
- Topic search. If a query does not match a person or organisation, we search headlines. For now search runs on original-language headlines, so English queries find more.
Contested framing
For each story we state a checkable claim (“Sanctions are not working”) and label every source: supports the claim, rejects it, reports it neutrally, or mixed. A language model labels from the headline and lead; a sample is checked by hand. The bar on a card shows the share of sources in each position. We do not judge who is right: we show how framing diverges and link to the sources so readers can check.
Limits
- GDELT does not see every outlet: regional and paywalled media are covered less than large ones.
- Machine translation and automatic name extraction make mistakes, especially with rare languages and namesakes.
- Dictionary tone does not understand irony or context.
- Blocks marked “demo data” show how a feature will look; their numbers are illustrative.
Found a mistake or want to suggest a topic? Write to us.