How Russian News Monitor works

We read what 13 Russian news outlets publish and measure, article by article, how each one treats 196 subjects: 80 countries, 43 organisations and groups, and 73 politicians. Everything on this site is counted from those measurements.

What we read

New articles are collected from Komsomolskaya Pravda, Lenta.ru, Moskovsky Komsomolets, Gazeta.ru, Kommersant, Nezavisimaya Gazeta, RBC, Argumenty i Fakty, Vzglyad, Life.ru, Fontanka, EADaily, Vedomosti. We store the full text as published and never revise it afterwards.

Collection started on 18 September 2026. Articles from 14 to 17 September were added afterwards from the archives of the outlets that keep one (Lenta.ru, Moskovsky Komsomolets, Life.ru, EADaily, Vedomosti, part of Kommersant, and from 17 September Komsomolskaya Pravda); the other outlets are missing on those days.

The list is not the whole of Russian media. Outlets under EU sanctions, among them RIA Novosti, RT, Izvestia and Rossiyskaya Gazeta, are not collected, and neither are TASS and Regnum. The most overtly state-run voices are therefore missing from the numbers.

What we ask of every article

Each article goes to a classification model that answers fixed questions with probabilities. It does not write text, so it cannot invent a claim; it can only answer wrongly. We ask:

  • Mention. Does the article mention or discuss the subject? Counted when the model is at least 50% sure.
  • Friendly, neutral or hostile to Russia. Which side does the article place the subject on: friendly to Russia (cooperates with it, supports it, opposes sanctions or military aid to Ukraine), hostile to it (fights, sanctions, threatens or accuses Russia, arms or funds Ukraine), or neither? This is the rating shown across the site. A country is judged as a state: one citizen, company or opposition party does not decide its side. Friendly and hostile count only when the model is at least 80% sure; anything less reads as neutral.
  • Salience. Is the subject what the article is about, or a passing mention?
  • Frame. Is the subject portrayed as a threat to Russia, a victim, declining, ridiculed, a partner of Russia, successful, or in no particular way? Threat and partner are about Russia only: a missile test aimed at someone else is not a threat here.
  • Topic and genre of the article as a whole.

The Russian state and its officials are not rated friendly or hostile to Russia: the question has no meaningful answer for them. Their pages show how they are framed instead.

Until 21 September 2026 the site showed a five-level rating from very negative to very positive. We stopped showing it because it mostly measured whether the news was bad for the subject, not how the outlet wrote about it: a plain report of a politician’s poor poll numbers came out “negative”. It is still collected.

The model is typesafe/jev-1.13, and the version is pinned: a silent model change would put a step into every trend line. When we change the wording of a question, the answers record which wording produced them.

The English headlines are machine-written summaries of the Russian originals. Every article links to its source.

The text headed “How Russian media writes about…” on a subject page is also written by a language model, once a week, from nothing but those summaries and the ratings described above. It describes what the outlets wrote and how; it does not check whether what they wrote is true, and like any machine summary it can get details wrong.

How accurate it is

We test the classifier against a labelled set of articles. Those labels were drafted by a larger language model and have not yet been checked by human reviewers, so the figures below are agreement between two machines, not accuracy. As of 21 September 2026:

Is the subject mentioned?98% precision, 96% recall
Friendly, neutral or hostile to Russia?87%
Main subject or passing mention?93%
Portrayal (seven frames)80%
Article topic82%

Finding mentions is reliable. Side and frame are judgement calls: the two models disagree on the side in about one case in eight, most often where our classifier reads a country arming Ukraine as hostile to Russia and the other model calls it neutral. Read the shares as a measure across many articles, not as a verdict on any single one.

How to read the numbers

  • Percentages on a subject page cover the last 7 days unless the page says otherwise.
  • A share built on a handful of articles is noise. The front page ranks shares only for subjects with at least 10 mentions.
  • “Hostile to Russia” describes how the articles present a subject, not what the subject is. It is a statement about Russian coverage.
  • Neutral is the large remainder, and mostly means Russia did not come up: an article about a Chinese panda says nothing about where China stands. Read the hostile and friendly shares, not the neutral one.
  • If a rating looks wrong, the “Incorrect?” button on the article sends us a correction.