How we rate sources

This is a translation; the Greek text is the reference version.

Version 2, 21 October 2026. We welcome comments; every change to the method is recorded with its date at the end of this page.

No outlet is rated on the platform before this method has been published in full. Every rating carries a date, the size of the sample it came from and a history of changes.

What we measure

On each story's coverage bar, the seven lean grades are grouped into three: Left, Centre, Right. Outlets without a rating are shown separately and do not count towards the percentages.

How political lean and stance are determined

We are a small team: one editor, assisted by artificial intelligence. Ratings rest on two sources that we can document, and none is published without the editor's approval.

1. Declared or structural relationship. Outlets owned or controlled by a party, as well as state media, are characterised on the basis of public records. We do not “measure” something that is already known.
2. Content analysis with artificial intelligence. An artificial intelligence model (Claude) scores a sample of each outlet's articles against the criteria in the table below and quotes verbatim the passages on which it bases each score. A second, more capable model combines the results for each outlet and proposes a lean, a stance and a confidence level.
   • The model does not see which outlet an article comes from: the names of the outlet and of its group are removed, so that it judges the text and not the outlet's reputation.
   • The sample gives weight to political and economic articles, because they reveal lean and stance. The target is 200 articles per outlet per quarter.
   • A passage is kept as evidence only if it appears word for word in the article.
   • The government of the day is defined in a dated file. When the government changes, a new round begins.
   • A rating is published only with medium or high confidence and at least 20 political articles. If the label is more than one and a half grades away from the average lean measured in the articles, the editor must review it separately.
3. Blind reader panel, once it is formed. The tools exist: readers from across the political spectrum judge anonymised samples of headlines and lead paragraphs, and each political group counts equally. Until at least 40 readers have been found, the panel is not used and we do not present it as a source.
   • Once it is running: a 7-grade scale for lean and four options for stance; readers in three groups by their self-placement (0–3, 4–6, 7–10) and three by their approval of the government's record, with equal weight; responses excluded if they take under 3 seconds per sample or give the same lean to 20 or more samples; a label only with 10 samples, 30 judgements and readers from all three groups; publication of their agreement (Krippendorff's alpha).

The rating shown is the content analysis, provided the editor approves it. On each outlet's page we show the confidence, the sample size, the short rationale and the articles it rests on, with a link to each. If confidence is low, we publish no rating and the outlet is shown as “Insufficient data”. Artificial intelligence never decides a rating on its own.

This method rests on one model and one person. It is less robust than a reader panel, and that is why we call it by its name on every rating: “Content analysis by artificial intelligence, approved by an editor”.

Criteria (score 1 to 5)

How factuality is determined

• How many times, per 1,000 articles over the last 12 months, a fact-checker certified by the European Fact-Checking Standards Network (EFCSN) cites the outlet as the place where a claim it found inaccurate appeared.
• Whether the outlet publishes corrections and whether it marks the articles it has corrected.
• Whether articles name their sources and link to them.
• How often the headline promises something the text does not contain.
• Repeating the same inaccurate claim weighs more than a one-off mistake.

In the first version we publish indicators, not an overall label, because we have not yet examined whether outlets publish corrections. The indicators come from the political stories in each outlet's sample (at least 20) and are shown next to the median of all outlets:

• headlines that promise something the text does not support (score 4 or 5 on the analysis's 1 to 5 scale),
• anonymous sources as a share of all sources,
• stories without a single named source,
• fact-checks from the last 12 months that found a false or misleading claim in an article by the outlet, with a link to each check.

Like the lean ratings, the indicators are published only with the editor's approval, and the outlet can ask for a review.

Ownership

For each outlet we show the company that publishes it, the ultimate owner (a person, family, party, the state, a cooperative, a non-profit organisation, a foreign group), the owner's other activities (e.g. shipping, energy, construction, banking, a football club) and, where they have been published, the amounts of state advertising it received, with the year. A holding of more than 5% is treated as influence.

Sources, in order of weight: the Register of Print and Online Press, the national media ownership database under the European Media Freedom Act, the shareholder records of the National Council for Radio and Television (ESR), the General Commercial Registry (GEMI), each website's imprint page («Ταυτότητα»), and well-documented reports of acquisitions. Every item carries its source and the date it was checked. Anything not documented by a public source is shown as “Not available”.

Blindspots

A story is a “blindspot” for one side when:

• at least 8 rated outlets cover it,
• that side has at most one third of the share it holds among all rated outlets (when left-leaning outlets make up 24% of the rated outlets, at most 8% of the story's outlets may be left-leaning),
• the opposite side has at least the share it holds among all rated outlets,
• so few outlets from the first side would turn up less than once in twenty times if the outlets that covered the story had been drawn at random from the rated ones,
• and outlets with low factuality make up at most 35% of the coverage (this condition will apply once factuality ratings exist).

The same rule applies to the axis of stance towards the government. We compare each story with the make-up of the rated outlets, so that rating more outlets on one side does not by itself create blindspots. On 21 October 2026, with 29 rated outlets (16 on the right, 6 in the centre, 7 on the left), the rule marked 6.0% of stories as blindspots. The target is 5 to 8%. On the “Blindspots” page we show how many of the rated outlets are on each side.

What artificial intelligence does

• It classifies articles by topic and region and groups them by story.
• It writes short summaries only from what the articles themselves report. Every sentence is accompanied by the outlets that report it. A sentence without a citation is not published.
• It translates foreign outlets' headlines into Greek and writes a summary of up to two sentences only from what the outlet itself publishes in its feed, with a link to the original.
• In each fact-check, it separates the links to the place where the claim appeared from plain references.
• It gathers evidence for the ratings, which people then read.
• It picks the stages of a developing story from among the outlets' headlines.
• It compares how each side's outlets present a story, from the headlines of rated outlets only and without knowing which outlet wrote each headline. Every sentence of the comparison points to one headline from the same side, which we show with its link. It does not judge whether a headline is accurate.
• It proposes the daily edition from the day's summaries. Only an editor publishes it, after reading it.

Every summary carries the notice: “Summary by artificial intelligence (Claude) based on N articles. It may contain errors; check the sources.” 2% of summaries are checked again every day against the full articles. If a sentence is not supported by the articles or the summary is not neutral, the summary is withdrawn. Any reader can report an error. The results of the quality checks are published on the “Quality checks” page and are updated continuously.

Fact-checks

We do not carry out our own checks. We show checks published by fact-checkers certified by the EFCSN (Greece Fact Check, FactReview, Ellinika Hoaxes, Fact Check Cyprus). A check is linked to an outlet only when the fact-checker itself cites an article by the outlet as the place where the claim appeared. The verdict we show is the fact-checker's.

Appeals

Any outlet can ask for a review of its rating, of a fact-check we linked to one of its articles, of a summary or of its ownership details, from the “Appeals by outlets” page. The editor replies within 30 days; a rating is re-examined with a new sample of articles and may be withdrawn. The decision (accepted or not, with reasons) is recorded publicly on the outlet's page, without the text of the appeal. The email address given for the reply is deleted as soon as the decision is made.

What we do not do

• We do not accept state advertising, nor advertising from owners of outlets we rate.
• We do not republish outlets' texts or photos without permission. We show the headline, outlet, time and link.
• We do not get around the limits that outlets set on automated access (robots.txt), nor their protection systems.
• We do not rate individual journalists.

Our own ownership

[To be completed before publication: company, shareholders, every funder above €5,000. The platform is rated with the same method.]

Changes to the method

• Version 2, 21 October 2026. Blindspots are judged against the make-up of the rated outlets. Until then, fixed thresholds applied: one side at most 10% of the coverage and the other at least 40%. With 29 rated outlets, 16 of them on the right, the fixed thresholds marked 8.7% of stories as blindspots, all on the same side, mainly because we rate more right-leaning outlets. The new rule marks 6.0%.
• Version 1, 21 October 2026. First publication.