Ratings Methodology

1. How stories are grouped

Reports about the same real-world event are grouped into one story automatically, twice a day. The grouping first matches articles that share specific named entities — people, places, organizations; generic terms are ignored — and then a multilingual semantic model merges groups that describe the same event in different languages. Articles that turn out not to fit the group they landed in are removed and regrouped. A story therefore shows one event as covered by many outlets, in several languages, over a window of a few days.

Grouping is fully automatic and imperfect: occasionally two related events merge, or one event splits in two. Because regrouping runs with every update, such cases usually resolve within a day.

2. Ownership type

We classify each outlet as Independent (privately owned, with no direct state or party control), State (a public broadcaster or a state-owned or state-funded outlet), or Party-aligned (owned by, or strongly tied to, a political party or interest group). This is based on public ownership records, regulatory filings, and reputable reporting on who owns and funds the outlet. It is curated by hand and periodically re-audited against current sources, and each outlet's page shows the specific basis for its classification. If you believe an outlet's ownership is recorded incorrectly, please tell us using the form below.

3. Editorial lean

The lean estimate reflects an outlet's overall editorial tendency across the spectrum from progressive to conservative. It is an estimate of editorial tendency, not a claim about any individual article or journalist. We combine three signals:

• Our estimate: an informed editorial prior based on the outlet's reputation and history.
• Measured framing: an automated analysis of the framing of the outlet's coverage over time, judging word choice, emphasis, and sourcing rather than the topic an article covers.
• Community input: aggregated bias votes from readers.

4. Per-report lean and political charge

Independently of the outlet-level estimate, every analyzed article receives its own framing rating on the spectrum from Progressive to Conservative. The rating judges how the article is written — word choice, emphasis, sourcing — not the topic it covers, and it is a statement about that single article, not about its author or outlet.

Each article is also scored for political charge: how politically contested its subject is. Only politically charged articles count toward a story's bias bar and an outlet's measured lean, so sports results and weather reports don't pull every outlet toward Center.

5. Factuality and objectivity

For each story, an AI model compares every article against the primary source it cites and the consensus across the other coverage, and scores factual accuracy (factuality) and neutral, non-editorial framing (objectivity) from 0 to 100. We aggregate these into an outlet-level indication. These are AI-assisted estimates and can be wrong on any given article.

The two scores are deliberately kept separate: an article can be accurate yet slanted, or neutral in tone yet wrong. Each score is published with a short written rationale, translated into every interface language, so you can see why a report was scored the way it was.

6. The neutral summary and its sources

For every story covered by at least two outlets, we generate our own neutral summary from the full text of the underlying reports. The summary must use original wording — it never copies sentences from the source articles — and stories with too little source material are skipped rather than padded out, so the summary cannot drift into invention. For larger stories we also extract the key factual claims, counting how many distinct outlets assert or dispute each one, and add a short note per side of the spectrum on what that side emphasized or omitted.

The "Primary sources" panel under a story follows three hard rules:

• Real: a link is shown only if it actually appears in the text of one of the underlying articles — a link the AI merely "remembers" is discarded.
• Specific: bare homepages are rejected; a link must point at the actual document or page being cited.
• Working: every link is checked and removed once it is definitively dead, and re-checked periodically.

7. Overlooked stories

A story is flagged as overlooked when at least two outlets cover it but all of that coverage sits on one side of the spectrum with no Center coverage — or when it has received almost no coverage at all. The Overlooked feed and the badge on story pages surface these blind spots.

8. Translations

The interface and all editorial content exist in seven languages (English, Slovenian, German, Croatian, Italian, Spanish and French). Article titles and summaries, our neutral story summaries, score rationales, claims and science summaries are machine-translated in advance by translation models we run ourselves, with a commercial engine as fallback. Links always lead to the untranslated original report. Machine translation is imperfect; where a nuance matters, consult the original.

9. Social pulse and science studies

The social pulse panel on a story shows public posts about that story from X and Bluesky. Posts are filtered before display: posts that are irrelevant to the story, likely automated, or from organizational or brand accounts are removed, and the remainder is summarized per language with an overall sentiment. Filtering is precision-first and errs on the side of keeping genuine human posts.

The Science tab is separate from news coverage: it lists new research from a fixed list of leading journals (Nature, Science, The Lancet and others), combined from the journals' own feeds and the OpenAlex database. Editorials and corrections are filtered out; each study gets a plain-language summary, a "why it matters" note and a link to the original paper. We never generate or alter studies — only select and summarize them.

10. AI use and human curation

All analysis on this site — article analysis, story grouping, summaries, scores, translations — is produced in batches twice a day on our own hardware, by open-weight AI models we run ourselves. Nothing you read is generated live while you browse, no reader data is sent to any AI provider, and every consequential AI output (ownership rationale, source links, claims) is constrained by verifiable inputs as described above.

Humans stay in the loop where judgement matters most: the ownership classification and the editorial-lean prior are curated by hand with cited sources, and every correction request is reviewed by a person.

11. Limitations

• Our ratings are AI-assisted and human-curated estimates, not statements of absolute fact (ownership aside).
• Measured signals depend on how much coverage we have analysed and on which stories happened in a given period, so short-term figures can be noisy.
• Story grouping and machine translation are automated and can occasionally be wrong; both are continuously refreshed.
• We refine our methods over time, so ratings can change.
• Ratings are an aid to your own judgement, not a substitute for it.

12. Corrections

We want our ratings to be fair and accurate, and we welcome corrections, especially from the outlets themselves. If you think an ownership classification, a lean estimate, or a score is wrong, use the form below or email corrections@objective-news.org with the outlet and what should change. We review every request and update our ratings when warranted.