Technical methodology

What we measure and what we don't

Each measure answers a different question:

• Coverage: which outlets reported the same story?
• Attention: how often does an outlet cover a person, topic or event?
• Wording: which aspect of the story does the headline foreground?
• Government stance: does it foreground praise, criticism or a neutral description?
• Policy positions: what do editorials and regular opinion writers support on specific issues?

These observations do not by themselves establish reliability. Frequently naming a politician does not mean supporting them. Criticism of the government does not determine an outlet's political identity. A difference between two headlines does not establish an intention to mislead.

Ownership is presented separately and does not determine an outlet's position on the political map. Our estimates come with their supporting evidence and limitations.

From articles to a story

1. We collect headlines, available short summaries and links from outlets' feeds.
2. We group articles about the same event. A shared topic is not enough: two articles about health may describe different stories.
3. We retain the source and available publication and collection times. When we observe a changed headline, we record its new version.
4. Readers can compare the headlines and open the original article on its publisher's website.

Collection has gaps. A feed may be unavailable or omit some articles. Story groups can also need correction. This is why an article we did not find is not necessarily an article that was never published.

Which outlets cover the same story

Each outlet counts once per story. Five related articles still count as one outlet covering the event. The number of articles answers a separate question: how often did the outlet return to it?

We compare the stories two outlets share and those found in only one. Before treating a missing observation as absent coverage, we need adequate monitoring and time for the other outlet to publish.

An outlet can cover a story without naming the prime minister in its headline. Filters select the stories to examine. Coverage counts every outlet with an eligible article in those stories within the same time window. Name mentions are counted separately. Before treating a missing article as missing coverage, we check whether collection was sufficient.

What an outlet pays attention to

We distinguish a headline naming a person, referring to them through their office, and an article substantively covering a topic. A keyword alone does not establish the third.

The same approach can examine a specific event, such as the UN General Assembly, or a government activity. A story solely about a foreign government does not count as one about the Greek government.

We show article counts and their share of the same outlet's assessed articles in the same period. For example, 10 relevant articles out of 100 assessed articles give 10%. Pending and uncertain articles are reported separately. Opinion articles have a separate population.

Overall totals and outlet comparisons answer different questions. Changes over time use the same suitable set of outlets in both periods. We withhold the comparison when there is insufficient observation history.

What a headline emphasises

We start by comparing headlines about the same event. We also check that they concern the same development: an initial report of a problem and an announcement of a solution the next day are different developments.

We are testing seven separate measures:

• Government foregrounding: is the government the main actor or placed in the opening phrase?
• Response substitution: is a government response foregrounded over an independently evidenced negative event?
• Responsibility: is the responsible actor named when their role is supported by evidence?
• Attribution of claims: are different expressions of scepticism used for comparable claims?
• Similarity to official communications: are distinctive phrases or their meaning repeated?
• Missing context: are facts known at the time, such as cost, cause, delay or affected people, omitted?
• Conflict emphasis: is the event presented mainly as party or personal conflict?

Several of these judgements need evidence beyond the headlines, such as the specific official statement or independent evidence of the event. We use only material available at the relevant time. Another headline is not by itself proof.

When required evidence is missing, the case remains uncertain. It is not counted as an assessed absence. The seven measures are not combined into an overall bias score. Each needs its own evaluation before publication.

The map

Instead of a single bar from left to right, each outlet is a dot on a map with two axes. We measure five axes. The map always opens on the two core ones, the economy and society. The three additional ones appear only once they pass their checks (see “When a topic becomes an axis”), and then the reader can put any two of the five in place of the core ones. We do not show all of them together in one chart, which would be hard to read; on each outlet's page its position is shown separately on each axis.

The map is published every night from the opinion texts read so far, and it changes as more are read. Opinion texts come from the outlets' feeds. For the months before we began reading the feeds, in July 2026, we take them from each outlet's own archive, from the same opinion sections, so that the map rests on the whole of the last 12 months. From the same archives we also take the texts the feeds missed. Columns published in opinion sections that are not opinion, such as the daily reprint of an old front page, or columns on television, shopping or local news, do not count as opinion texts. An outlet appears on an axis as soon as it has enough positions (see “When we show no position”), so the map fills in gradually. The checks of step 8 are not a condition for publication: the map page shows each of them with its current value, whether it passed, has not passed yet or is in progress. Anyone can dispute a position, and the decision is taken by a person on the editorial team, not by a model.

The two core axes

Economy: "State intervention ↔ Free market". It concerns the role of the state in the economy. Towards "State intervention" lie redistribution (moving income, through taxes and benefits, from those who have more to those who have less), regulation and public ownership. Towards the "Free market", competition, private enterprise and lower taxes.

Society: "Progressive ↔ Conservative". It concerns rights, equality, migration, religion, the nation, law and order. Nationalism belongs here, not on the economic axis: among the Greek parties of the CHES 2024 survey the two scales almost coincide (correlation coefficient 0.93).

The three additional axes

Europe & the West: "Europe & the West ↔ Sovereigntism". It concerns European integration, NATO and the United States, Russia and Ukraine. "Sovereigntism" is the view that the crucial decisions should stay with the nation state and not with European institutions or Western alliances. It is not simply the conservative end of the social axis: among the European parties of CHES 2024 the two core axes predict only 52% of this position. The most sovereigntist Greek party is the KKE, and Russia divides the national right: on the CHES scale from 0 ("more defence against Russia") to 10 ("more trade and diplomacy"), Greek Solution and Niki stand at 9.5 and Voice of Reason at 2.5.

National issues: "Dialogue ↔ Hard line" (trial). It concerns Greek-Turkish relations, the Cyprus question and relations with the neighbours: dialogue and compromise, or an unyielding stance and deterrence. It is the most frequent opinion topic in Greek media, with 2.6% of opinion headlines over six months, more even than the economy. We are testing it before we publish it, for two reasons. It may repeat the social axis, to which nationalism belongs. And it may repeat the stance towards the government, since dialogue with Turkey is also government policy. CHES has no question for this axis, so the parties are placed only from their programmes and their order is not checked against another source.

Institutions & system: "Institutional ↔ Anti-system" (trial). It concerns trust in institutions, or the rejection of the political system as a whole: referendums, trust in the courts, in elections and in experts, the idea that "they are all corrupt". Criticism of a particular government does not count here; it counts in the stance towards the government. In CHES 2024, the question whether "the people" or elected representatives should decide is the one most independent of the core axes, which predict it by only 15%. When it is combined with anti-elite rhetoric, however, it moves with Europe: among the Greek parties the correlation coefficient is 0.77. So the axis has to show that it divides the outlets beyond the Europe & the West axis as well.

An axis is shown as trial when it has passed every check in the next section except the sixth, the parties' order, which can be checked only where CHES has a question for the axis and enough parties have a position from their programmes. The label stays beside its name, and the note on trial axes, on the map and in the quiz, says whether experts have judged outlets' places on it; their check does not change the label.

When a topic becomes an axis

Every new axis passes seven checks, and passes them again in every edition of the map:

1. It measures something new. The existing axes predict less than 70% of the parties' positions on it, among the European parties of CHES 2024. Once there are data, the same check is made on the outlets.
2. It divides the parties. The Greek parties are far enough apart: a standard deviation of at least 0.30 on the scale from −1 to +1.
3. It divides the outlets. A standard deviation of the outlets' positions of at least 0.20.
4. There are texts. At least 12 outlets have taken a position on half the issues of the axis.
5. The checking model agrees with the readings. κ ≥ 0.70 on 150 positions of the axis itself.
6. The parties hold. The order of the parties from their official texts agrees with their order in CHES (ρ ≥ 0.70). Where CHES has no question, the check is not made.
7. It is not the stance towards the government under another name. Among the outlets, its correlation with the index of the stance towards the government is below 0.70 in absolute value.

And when the reader picks two axes that move together among the outlets in the current edition (a correlation of 0.60 or more in absolute value), the map says so: the outlets spread along a diagonal and the second axis adds little information.

Topics that did not become axes

We also examined other topics as candidate axes. Most turned out to be parts of the core ones:

One more thing holds for the environment: the Greek parties differ on it less than on almost any other CHES topic (a standard deviation of 1.72 on the scale from 0 to 10, against 3.14 on migration).

The shares of opinion headlines are our own measurement: 6,136 headlines of opinion texts from 46 outlets, in the 180 days to 24 September 2026, by keywords. Headlines understate how often a topic is written about, so the shares are only comparable with each other.

Topic comparison. These topics are not lost. In "Compare by topic" the reader picks a topic, for example "Environment & climate", and sees each outlet's stance on each of its issues separately, with the quotes that support it. We do not sum it up in a dot. The topics of the comparison are: Environment & climate, Migration, Church & religion, Rights, Labour & wages, Housing, Russia & Ukraine, Greek-Turkish relations. Some issues we measure only for this comparison ("off the axes"); they move no dot.

Why these axes

The two core axes match the way political science describes Greek politics. In the Chapel Hill Expert Survey (CHES), political scientists place the parties of European countries, the Greek ones among them, on common scales: among others on the economic axis from left to right, on the GAL-TAN axis (from green, alternative and libertarian values to traditional, authoritarian and nationalist ones) and on their stance towards the European Union.

Katsanidou and Otjes (2016) showed that the debt crisis reshaped the Greek political space: because accepting the bailout memoranda was tied to staying in the eurozone, economic and European issues merged into one axis, while cultural issues formed a second. That is why we keep cultural issues apart from economic ones. And we measure Europe separately: if it still moves with the economy, the data will show it.

Why not a single bar

A national-conservative outlet, which puts the nation, tradition and religion first, may at the same time call for strong state intervention in the economy. On the map it stands near "State intervention" and near "Conservative". On a single bar where the economy weighs more, it would appear "left", as happened with our previous method.

More generally, in our measurements the economic and the social positions of the outlets correlate only weakly: the correlation coefficient is 0.39 (at 0 two quantities are unrelated, at 1 one fully predicts the other). An outlet's position on the economy says little about its position on social issues. A single bar would hide that difference.

The stance towards the government is measured separately

How an outlet treats the actions of the government of the day is another question, with other evidence (headlines and summaries of news stories, not opinion texts) and another display: a separate strip with the outlet's rating and a chart by month. The measurement starts afresh with every change of government and is never mixed into the map.

An outlet may stand close to the positions of the governing party and criticise the government, or stand far from them and support some of its policies. If the two counted together, the dot would move with every change of government without the outlet's views having changed.

How a dot gets to the right place

The process has nine steps. In none of them does an artificial intelligence model decide what is "left" or "right".

1. The list of issues

The basis is a published list of 46 specific issues of today's public debate on five axes (10 on each of the first three, 8 on each of the two trial ones), plus 6 issues we measure only for the topic comparison. For each issue we define in advance its wording and which side counts towards which end: support moves towards the end the table shows, opposition towards the other one. Each issue is written as one full sentence that a columnist can agree or disagree with. For each issue the list also keeps examples of what counts as "for" and what does not; the review of the list and the editorial desk use them, while the models that read the texts see only the sentence. The column "Topic comparison" shows in which topics of the comparison each issue also appears. The list is drafted by our team. Before publication, an artificial intelligence model reviews it under written rules and the project's editor approves it. It is published with a version number.

Economy: State intervention ↔ Free market (draft)

Society: Progressive ↔ Conservative (draft)

Europe & the West ↔ Sovereigntism (draft)

National issues: Dialogue ↔ Hard line (trial; first draft)

Institutions & system: Institutional ↔ Anti-system (trial; first draft)

On this axis the distinction from criticism of the government is the hardest point: a text that criticises a government decision does not count here, unless it questions the institutions or the system as a whole.

Off the axes: issues only for the topic comparison

Every issue off the axes adds readings and checks (steps 4 and 5). So we add such issues only for topics readers ask about.

2. Who speaks for the outlet

What counts are the unsigned editorials, the texts of the newsroom's columnists and the texts of regular contributors (those who published at least 5 texts in the outlet in 12 months). A column with a fixed name counts as one columnist, whoever writes it.

Who is what. For every byline with at least 5 texts, an artificial intelligence model decides under a written rule whether it is the editorials, a regular contributor, a guest or reporting. It is a different model from the checking model of step 5, and it sees the outlet's name, the bylines and the texts' headlines. The editor in charge can change any of its decisions. A byline not yet decided does not count.

We count columnists, not texts. An opinion text first says what its author believes. So that one very prolific columnist does not become "the outlet", each regular columnist counts once, however many texts they write. Editorials express the newsroom itself: they count together as one columnist, with double weight. In our measurements, in 5 of the 21 outlets with enough opinion texts a single columnist wrote more than half, and in one outlet all of them.

At least three columnists. We give an outlet a position on an axis from its opinion texts only if at least three columnists took positions on its issues, the editorials counting as one: for example the editorials and two columnists. With fewer, we show the columnists' positions under their names, not as the outlet's position.

No columnist decides alone. We compute the position again, removing in turn each columnist and the editorials. If removing one of them moves the dot outside its range of uncertainty, the position on that axis is not published as the outlet's; we say that it depends on one columnist.

Each outlet's page also shows its columnists separately, so that the reader can see how alike or different they are.

Occasional guest pieces, such as an article by a politician, do not move the dot: hosting a view does not necessarily mean adopting it. They appear separately, as the "range of guest views".

Headlines. News reports as texts are not read for the map: they rarely take a position, as our measurements show (in 985 judgements of news articles, only 1% found an economic or social position) and as international research does (see the Technical annex). Headlines, however, are written by the newsroom, and sometimes they take a position through the words they choose: the same event one outlet writes as "68 migrants" and another as "68 illegal migrants". So we tested headlines as a second, separate source, with checks of their own (steps 4, 6 and 8). The test failed (step 8), so the map relies only on opinion texts.

3. Collection

We always look at the last 12 months: at each update the new texts come in and the oldest go out (a "rolling" window).

Each text is recorded as it was when we read it. For every paragraph we keep a digital fingerprint, that is a short code that changes if even one word changes. So a later change to the article cannot silently alter the evidence.

We keep only what the method needs, always refer to the original and publish only short quotes.

The headlines, and the short summaries below them, are already in our data, from the outlets' news feeds. We use the headlines of stories that at least 4 outlets covered, so that they can be compared with each other.

4. Reading

Opinion texts. Each text is read independently by two artificial intelligence models (programs that process written language) from different companies. They answer only three questions:

1. Which issues of the list does the author support or oppose?
2. Which way: for, against, or a mixed position (with elements of both sides)?
3. Which phrase shows it? It is quoted word for word and must be the author's own words.

The words of a politician quoted in the text never count as the outlet's position.

The model says only, for example, that the author holds that the state should raise wages. That this counts towards "State intervention" is set by the list of issues, not by the model, because language models themselves have measurable political preferences (see the Technical annex).

A text that touches no issue of the list counts nowhere. We do not filter texts by keywords before the models read them: a text can discuss housing policy without writing the word "rent".

Headlines. In two stages:

1. A model reads the headlines of each story and says which issues of the list the event concerns, if any. Most stories (crime, weather, international news unrelated to the list) concern none and stop there. To know what it misses, every month the checking model (step 5) judges again a random sample of the stories it rejected.
2. For the rest, the same two models as for the opinion texts independently read all the headlines of the story together, shuffled and without the outlets' names, and for each headline say whether it takes a position on the issue, for or against, and with which words.

The words are quoted verbatim and must belong to the headline itself or its summary: the words of a minister or of the opposition in quotation marks do not count. A headline that simply reports what happened takes no position, even if the news is good or bad for one side. Each outlet counts once per story and issue, however many headlines it wrote about it.

5. Review by another model

• When the two readings agree, the position is accepted.
• When they disagree, including when only one of them saw a position, a third artificial intelligence model, from another company, decides: "the checking model". It reads the whole text without the outlet's name, the author's or the readings' answers, decides what the author supports, and writes one sentence of reason and one phrase of the text, verbatim.
• Every month, the checking model judges again a random sample of the positions on which the readings agreed (10%, and at most 150 a month), because two models can make the same mistake too. For the first edition of the map it judges again one draw of 300 positions.
• The same holds for headlines: the checking model sees all the headlines of the story without the outlets' names.
• No person judges the positions one by one. The editor in charge can correct any decision, and every correction records who made it.

We publish the agreement rates, separately for opinion texts and for headlines.

6. Calculation

No artificial intelligence takes part here; it is simple arithmetic.

A columnist's position on an issue. For each columnist of the outlet (and for the editorials, together) and issue:

(texts towards one end − texts towards the other) ÷ all their texts on the issue

The value ranges from −1 to +1 and is 0 when the texts are evenly split. Mixed texts push towards neither side, but they count in the total.

The map's scale. On the map and on every page, an outlet's position is written on a scale from −3 to +3: three times the value of the calculation, with one decimal, so −0.93 is written −2.8. The calculation does not change; only the way the number is written does. −3 means that every text that counts supports the first pole on every issue of the axis, +3 the second, and 0 that the two sides balance. We chose this scale because the integers from −3 to +3 give the map a grid that is easy to read, and because it spans the same range as the grade of the stance towards the government, so readers meet one range of numbers across the site. The two remain different measures.

The outlet's position on an issue, from opinion texts. The average of the columnists who wrote on the issue, with a weight of 2 for the editorials and 1 for each columnist.

The outlet's position on an issue, from headlines. (stories where its headline took a position towards one end − stories where it took a position towards the other) ÷ all the stories of the issue where it took a position. An issue counts from headlines only if the outlet took a position in at least 3 of its stories, because a headline is weaker evidence than an opinion text.

Position on the axis. It is the average of the values on the issues of the axis the outlet wrote about, each issue weighing more the more texts it has: an issue with N texts counts N ÷ (N + 2). An issue with a single text counts a third, one with five texts about 0.7 and one with fifty about 0.96. So an issue with one text does not weigh as much as one with forty, and a favourite topic still cannot dominate: if an outlet writes fifty texts on taxes and five on housing, taxes count only 1.35 times as much as housing, not ten times.

Minimum evidence. An outlet is placed on an axis, from each source separately, only if it has taken a position on at least half of its issues: 5 of the 10, or 4 of the 8 on the trial axes. From opinion texts, three columnists are needed as well (step 2).

The two sources together. On each axis, the two sources are joined only if, among the outlets with a position from both, they rank the outlets in a similar order (ρ ≥ 0.60, over at least 8 outlets). Then an outlet's dot is the average of its two positions, with equal weight, and an outlet without opinion texts is placed from its headlines alone, with a hollow dot and the note "from headlines only". If an outlet's two positions are so far apart that their ranges of uncertainty do not overlap, we show not an average but both: the difference between what it writes in its editorials and what it chooses in its headlines is a finding too. If the two sources do not rank the outlets in a similar order, the map shows only the opinion texts and the headlines appear separately on each outlet's page.

Topic comparison. There we show the position on each issue separately, computed as above, with no average.

Uncertainty. We compute the position again 1,000 times, each time on a random sample of the outlet's columnists and of each columnist's texts (and of the stories, for headlines), where the same item may be picked more than once (the technique is called the bootstrap, or resampling). If the position stays nearly the same, the estimate is stable; if it moves a lot, it is uncertain. The shaded ellipse (the oval shape) around the dot shows this range: on each axis it spans the interval where 9 in 10 of the 1,000 positions fall.

7. Reference points

On the same map we also show the Greek parties, placed as the outlets are: from their official texts (programmes, congress decisions, policy papers, speeches). The same two models read them, with the same list of issues and the same rules (steps 4 and 5), with one difference: only what the party itself supports counts, not what it describes or attributes to others. Each text is divided into sections, and each section counts as an opinion text does; a party counts as one columnist. Its position on an axis is calculated as an outlet's is, with the same weight for each issue (step 6), and is shown on the map from −3 to +3. A party appears on an axis only if its texts take a position on at least half of its issues, as an outlet must. The parties that fall short on either of the two axes you are looking at are named under the map.

So outlets and parties are measured with the same ruler: the issues of the list. Where an issue has almost only one side in the Greek public debate (support for the national health service, for example), outlets and parties alike lean towards it. That is why a position is best read against the other positions on the same map.

Until 12 October 2026 the map showed the parties where the CHES 2024 expert survey (Chapel Hill Expert Survey) places them: 609 political scientists placed 279 parties from 31 countries, 9 Greek ones among them. CHES stays as an outside check: on each axis, the order of the parties from their texts must agree with their order in CHES (ρ ≥ 0.70; see the Technical annex). For national issues CHES has no question, so this check is not made.

The parties help the reader "read" the map. That an outlet stands close to a party does not mean that it "belongs" to it.

8. The map's checks

Two checks by models. Two artificial intelligence models from two different companies, which are not the companies of the readings, judge the same 300 positions separately, neither seeing the other's answers or the readings'. The agreement of two models is not the agreement of two people: models that learned from similar texts can make the same mistakes. That is why we choose models from different companies and publish where they disagree. How far they agree with each other and with the readings is shown on the map page, with its current value. We measure the agreement with Cohen's kappa coefficient (κ), which removes the agreement that would arise by chance: 1 means full agreement, 0 agreement no better than chance. Our target is κ ≥ 0.70, a common threshold of satisfactory agreement.

Academics' views. Academics who study politics, the media and communication can place the outlets independently of us, along with each outlet's stance towards the government. No one places an outlet they write for regularly, and the answers are not linked to names, not even by us. Their view is an approval or a disagreement that the editorial team examines whenever it is given: where they disagree sharply with the map, the editorial team examines the outlet's position again. It is not, however, a check of the map. Up to version 1.5 it was two checks: the academics' order of the outlets had to agree with the map's order (ρ ≥ 0.70), and the stance towards the government they gave was compared with the grade we measure from headlines. We retired them on 5 October 2026: they cannot be measured without enough answers, the map is published without waiting for human approval, and the checking is done by artificial intelligence models under written rules, as described above.

Quality checks by experts. Experts anonymously check samples of the results and the data: an outlet's place on an axis with the quotes behind it, the position a model found in a quote, the judgement on a headline and the role of a byline. Each round is a fixed sample, and each expert marks every item “Correct”, “Wrong” or “Don't know”, with a note if they wish. The editors examine every “Wrong”, and the ones they accept go into the corrections log with their reason. A headline's or a quote's judgement the editors accept as wrong is corrected: the right judgement the editors write counts in place of the model's. A byline's role is corrected on the columnists' page. An outlet's place is made of its quotes and is never corrected by itself. The check is never a condition for publishing anything; its results, in numbers, are in the methodology.

Each axis separately. Agreement between the checking model and the readings is measured on 150 positions of each axis. An axis that does not pass is not shown, even if the others pass.

Headlines as a source. Before headlines are used, they pass a test with rules set before we see any results: the checking model characterises 150 headlines without seeing the outlets' names or the readings' answers, and its agreement with the readings must be κ ≥ 0.70; the model that picks the issues must miss at most 10% of the stories that concern an issue; and headlines must give enough outlets a position for the source to be worth it. If the test fails, the map relies only on opinion texts and we say so here. After the test, headlines are checked every month like opinion texts.

The test's result (3 October 2026). The test failed, so the map relies only on opinion texts. Of 400 random stories covered by at least 4 outlets, only 11 concerned an issue on the list. The model that picks the issues was not the problem: of 60 stories it turned down, only 2 concerned an issue, below the 10% limit. But across the 103 headlines of those 11 stories, the two reading models agreed on no position, and the checking model, on 58 headlines (all those either model gave a position, and 50 more), found a position in only one. With so few positions no outlet would gather enough stories per issue to be placed on the map, not even over a year. Outlets that publish no opinion of their own therefore appear with their stance towards the government and their coverage. A new test, with new rules set in advance, may follow later. Until then the two headline checks, the agreement with the checking model and the stories the issue picker misses, have nothing to measure; from version 1.6 they are not shown among the map's checks.

The test with the text of news reports (5 October 2026). We also tested whether the text of news reports, not only the headline, shows the newsroom's position, for the outlets that publish fewer than 5 opinion texts a year. The two reading models read 10 news reports from each of these outlets, under three rules set in advance: agreement between the two models of κ ≥ 0.70, confirmation by the checking model for at least 80% of the positions, and a position for at least 5 outlets. News reports mostly relay what others say: of the 53 positions the models found, 45 were third parties' words in quotation marks. In the 268 reports both models read, the author's own position appeared in 6, and the two models rarely found the same one. The test failed on all three rules, so news reports do not place outlets on the map.

Readers. Readers' placements, from the rating form on each outlet's page, appear beside ours once there are at least 30 per outlet. They never change the position of the dot: they show how readers see the outlet, not what the outlet itself argues.

9. Publication and updates

• The map is published every night with that night's figures, with no check required. Each publication keeps the list of issues, the time window of the texts and the checks as they stood that night.
• The editorial team can withdraw a publication; the previous one is then shown, and the withdrawal is written in the public corrections log.

An example

Take an imaginary outlet, "Outlet X", and the issue "The state should raise wages, through the minimum wage or through collective agreements that apply to everyone, and not leave them to the market". Support counts towards "State intervention" (−1) and opposition towards the "Free market" (+1).

In the last 12 months, the editorials and three columnists of Outlet X took a position on the issue, in 9 texts:

• the editorials: 3 texts, all for, so (0 − 3) ÷ 3 = −1.00;
• columnist A: 4 texts, 2 for, 1 against and 1 with a mixed position, so (1 − 2) ÷ 4 = −0.25; the mixed text pushes towards neither side, but it counts in the total;
• columnist B: 1 text against, so +1.00;
• columnist C: 1 text for, so −1.00.

The editorials count double:

[2 × (−1.00) + (−0.25) + 1.00 + (−1.00)] ÷ (2 + 1 + 1 + 1) = −0.45

Had A written 40 texts instead of 4, in the same proportion, Outlet X's position would stay −0.45. Had we counted texts instead of columnists, it would be −0.29, almost A's position.

Outlet X took a position on 7 of the 10 economic issues with its editorials and three columnists, so it meets the 5-issue rule and the three-columnist rule. The average of its 7 values is −0.31, with an uncertainty of ±0.08: clearly towards "State intervention", but far from the end. On the map it reads −0.9, with a range from −1.2 to −0.7.

From its headlines, Outlet X took a position on the same issue in 5 stories, 4 times towards "State intervention" and once towards the "Free market": (1 − 4) ÷ 5 = −0.60.

The outlet's page shows the 6 texts for, each with the phrase that was counted and a link to the original, and in the same way the 2 against and the mixed one. So the reader can check every step for themselves.

When we show no position

• "Too few texts": the outlet took a position on fewer than half the issues of an axis. It appears as a dot on the map only if the rule holds on both axes the reader has chosen; otherwise the map says so, and its position on the other axes shows on its page.
• "Too few columnists": fewer than three of the outlet's columnists took a position on the axis (the editorials count as one). We show the columnists under their names, not a position of the outlet.
• "Depends on one columnist": without one of its columnists, the outlet's position would fall outside its range of uncertainty.
• "Publish news only": many news sites mostly report the news, without editorials or regular columnists. We tried to place them from their headlines and from the text of their news reports, without success (step 8). They appear with the strip of their stance towards the government and with their coverage data, that is which stories they cover.

We never place an outlet at the centre because we lack data. The centre is a position like the others: it means moderate or balanced positions, not that we do not know.

How we measure the stance towards the government

We examine how actions of the Greek government are presented in a headline and its own summary. Other headlines from the same story help resolve references. Their wording or judgements do not become evidence in the target headline.

The rule distinguishes four cases:

• Positive: foregrounds government success, benefit or a sound decision.
• Critical: foregrounds failure, harm or government responsibility.
• Neutral: describes an action without evaluation, or gives praise and criticism similar emphasis.
• Unrelated to the government: there is no supported connection to the Greek government, its actions or responsibility.

A positive headline means it foregrounds a success of the government, even when it reports a statement; it does not mean the outlet agrees.

Quotation marks do not make a statement neutral. Under this rule, a headline foregrounding “The government's plan failed” foregrounds criticism even when it identifies the speaker. This does not mean the outlet endorses the statement. The same rule applies to praise.

A routine meeting or announcement is not praise by itself. A disaster is not automatically criticism of the government either: the text must support a connection to its action or responsibility. All headlines in the same story can be neutral.

The same holds for announcing or describing a measure, a benefit, a programme or a project (what it provides, for whom, how much, when) and for a government goal, plan or promise: they count as neutral unless the headline or its own summary praises or criticises them, or presents them as a success or a failure. That a measure gives money or benefits someone is not enough by itself to make a headline positive. Nor are independent authorities and public bodies, such as AADE or DYPA, the government by themselves when they do their usual work. The rule was made explicit on these points on 11 October 2026 (version 3), because the model saw praise in plain announcements more often than the check did.

The planned route to a public measurement is:

1. Story selection: at least 6 outlets cover the same event. Material clearly unrelated to the Greek government is skipped under written rules, with a sample of skipped material reviewed.
2. Headline comparison: the model sees the distinct headlines together, with their own summaries and without outlet names. The government roster corresponds to the event date. For each headline it writes its judgement and one sentence explaining it.
3. Check by a second model: a second artificial intelligence model, from another company, judges 330 headlines from a new sample under the same rules. The 300 are headlines the first model has already judged; the second sees neither its answer nor the outlet's name, and for each judgement it writes its reason and the words of the headline that support it. From these we measure how far the two models agree, with a target of κ ≥ 0.70. The other 30 come from stories the selection left out: of these, at most one may concern the government. Only judgements of the model and the rule under check count. The check is done by a model, not by people.
4. Calculation: each outlet counts once per story. With more positive than critical relevant headlines the story counts as positive, with more critical ones as critical, otherwise as neutral: when no headline takes a side, or when the two sides have as many headlines each. Neutral headlines do not cancel a headline that takes a side. Each story belongs to one month, the month it was first seen. The rating of a month requires at least 20 stories in that month in which the outlet's headlines took a side.

Published before the check. The route provided for 14 days of trial operation of the final rule without being shown, and for the check of step 3, before publication. By decision of the editorial team, the measurement is published without waiting for the 14 days and the check. Until the check has passed, the figures are provisional and change every night with the new judgements. For the measure published today, the check passed on 11 October 2026. The state of the check and its result are shown at the top of the page How Protagoras works. If the check does not pass, the measurement is withdrawn and the groups return to the editorial estimate.

Only the stories in which the outlet's headlines took a side count. The grade, from −3 to +3, comes from the ratio of positive to critical stories: ±1 when one side has at least twice as many stories as the other, ±2 four times, ±3 nine times, with the sign of the side that leads, and 0 below twice. The index (positive − critical) ÷ (positive + critical) ranges from −1 to +1, and the limits correspond to absolute values of 1/3, 3/5 and 4/5. The limits were set before the measure was published; changing them is a new version of the method. How many of the outlet's stories took a side is shown next to its grade.

The second check of every side. Every headline the model judges positive or critical is read again by another model, from another company (today Claude Haiku 5.5), with written instructions and format checks, on the same group of headlines the first model saw and without knowing its judgement. The side counts only when the other model gives the same side. If it finds the headline neutral, of the other side or not about the government, or has not answered yet, the headline counts as neutral. An editors' correction counts as before. The other model judges up to five headlines per call and always sees all the story's headlines. If one of its answers is rejected twice by the format checks, its headlines count as neutral; for September and October this concerned 17 of about 12,000 headlines. Third parties' statements count as the rule says, with one exception, the next.

The speaker check. For every side that counts, the same other model (today Claude Haiku 5.5) says whose statement or act carries it: the outlet itself, a Greek speaker or body, a foreign government or an official of the party that governs a foreign country, an international body, another foreign source, or that it cannot tell. When a foreign government or an international body carries it, the headline counts as neutral, as it does while the model has not answered. The Greek government's own act with a foreign or international party, such as a repayment, a Greek resolution that was adopted or a meeting, counts as usual, as does a Greek speaker on a foreign matter. The model judges ten sides per call, on the title and summary the first model saw, and quotes word for word what shows the source. Before it was used, it was tried on every side that counted and checked by hand: in its first version about three in four of what it left out were right, and its instructions were corrected three times, mainly so that the government's own acts count. If one of its answers is rejected twice by the format checks, or it does not answer for a side in three passes in a row, the side counts as usual; for September and October this concerned 50 of 6,555 sides.

What changed on 22 October 2026. In a check of 240 headlines with a side that counted, 15% of the critical and 21% of the positive ones were plainly wrong: headlines about a foreign government, headlines in which the government speaks about the opposition instead of being judged, plain reports read as taking a side, and once the opposite side. By decision of the editorial team the second check above now applies. In the same check the plain errors fell from 15% to 3% of the critical and from 21% to 7% of the positive headlines. In the measurement of 22 October, September kept 3,894 of its 7,086 stories with a side and October 2,434 of 4,299, and the outlets with a grade became 54 of 63 and 47 of 57. A stricter version, which also leaves out every side resting on a third party's statement, was examined: it kept 13% to 16% of the stories with a side and tilted outlets towards positive, because criticism of the government reaches headlines mostly as a statement of the opposition; it was not adopted. Before it was published, the new measure was computed beside the old one from the same judgements: on 22 October 2026 it gave 10 pro-government, 31 neutral and 20 critical outlets, against 10, 42 and 17 under the previous rule; 19 outlets had fewer than 20 stories with a side, against 11.

And later the same day. By decision of the editorial team the speaker check above now applies too. It answered for 6,505 of the 6,555 sides that counted since 1 September, and 249 of them left: 185 positive and 64 critical, because foreign praise is far more common in headlines than foreign criticism. The measure then gave 9 pro-government, 29 neutral and 21 critical outlets, against 10, 30 and 21 without this rule; 21 outlets had fewer than 20 stories with a side, against 19.

What changed on 14 October 2026. Until then every relevant story counted, and the index was (positive − critical) ÷ all, neutral ones included, with limits at 0.15, 0.40 and 0.70. In each story the stance that prevailed in the outlet's headlines counted, so a plain update evened out a headline that took a side. About two in three stories about the government are plain reports, so an outlet's group also depended on how many such stories it publishes. By decision of the editorial team the rule above applies now. Before it was published, the new measure was computed beside the old one from the same judgements: on 14 October 2026 it gave 10 pro-government, 41 neutral and 18 critical outlets, against 6, 52 and 16 under the old rule.

The editors' corrections. When the editors accept that the judgement on a headline is wrong, for example after an expert's “Wrong”, they write the right judgement and its reason. The correction counts in place of the model's judgement, shows on the headline with its reason and goes into the corrections log. It holds for the rule the headline was judged under and for the headline as it was written: when the rule or the headline's wording changes, the new judgement counts again. The figures we publish are the model's judgements, as the other model confirms them, together with these corrections; the check on 300 headlines and its κ measure the judging model alone.

The rating on an outlet's page. The rating an outlet's page shows, in its header, on the stance card, in "Why" and on the share image, is computed from all the outlet's measured stories together, from September 2026 on and within the government in office. The positive, neutral and critical stories of every month are added up, and the rating comes from the sums of the positive and the critical ones, by the same rule. Since each story belongs to one month only, none counts twice. At least 20 stories with a side in all are required; until then the page shows the editorial estimate, if there is one, and otherwise how many stories have counted so far. The months appear separately, in the page's chart and in the list of stories. The coverage bars on stories use the group of a single month, as explained below, so an outlet's group in the bars can differ from the rating on its page: while no month has 20 stories with a side yet, the bars show the editorial estimate, while the outlet's page already shows a rating from all months together.

On each outlet's page, below its stance, the page shows what it rests on: how many stories counted as positive, neutral and critical, the calculation, and examples of headlines with the model's short explanation. The full list shows every story that counted, all months together or each month on its own, and anyone who finds a mistake can dispute it. The explanations are written by the model and may contain mistakes; we show only those that quote words of the headline itself or of its own summary.

The strip and the chart by month describe stance towards that particular government. The measurement starts again when the government changes. The index does not judge whether praise or criticism is justified.

The groups and the coverage bars

While the monthly measurement is published, groups follow the grade: pro-government with at least twice as many positive as critical stories, critical the other way round, and neutral when neither side reaches twice the other. At least 20 stories with a side are required. Without a sufficient new measurement, a group can be retained for up to two months within the same government; the outlet is then marked “not measured”.

A story's bar shows how many outlets in each group covered it. It does not say that those particular headlines are all positive, neutral or critical. The group describes the outlet; a headline judgement describes that particular text. The usual coverage mix is shown alongside it for comparison.

Blind spots

A blind spot is a story that almost only the pro-government outlets covered, or almost only the critical ones. We measure who covered it, never how.

1. Which stories. Those covered by at least 3 outlets with a group, 24 hours after the first report, so that the other side has had time to write. Each outlet counts once; outlets whose article collection is not working do not count as "absent".
2. The rule. If the outlets of one group that covered it are so few that this would happen by chance less than 1 time in 20, given that group's share of usual coverage, the story is flagged. With none, it is a "complete absence"; with a few, "marginal". Never for the neutral outlets: their absence says nothing.
3. Re-check. Before it is published, we search all the articles of the three days around the event for a text from the "missing" group that we had not linked to the story. If one is found, the story is not published.
4. Judgement. An artificial intelligence model first judges whether the card tells the reader something political (a lottery draw or a crime report from abroad does not). Whatever slipped through is hidden by the editor in charge or, on the editor's instruction, by a second artificial intelligence model, always with a stated reason; every hide records who made it. Hidden cards are recorded and counted: the reason "there is coverage we did not find" is the only one that shows a mistake of ours.

Blind spots stay in the archive with the data of the moment they were found, even if other outlets covered the story later.

Transparency, corrections, right of reply

We publish the list of issues with all its versions, every position per text with the phrase that supports it and a link to the original, the formulas of the calculation and the history of the map's editions.

Every outlet can dispute, with evidence, a position we attribute to it: for example, that the phrase belongs to a third party and not to the author, or that the author is not a regular contributor. Every such request is examined by the editor in charge, not by a model, and where we were wrong we correct it.

For the first edition of the map, the outlets are not notified seven days before publication. Every outlet can, at any time after publication, ask for a correction or dispute any position we attribute to it.

Every correction is recorded in a public corrections log: what changed, when and why.

Who does what.

Every model decision is stored with the model that took it, the rule, the reason and the phrase that supports it. No person checks these decisions one by one, and we do not say that the measurements were confirmed by people.

Limits

• There is no "right answer" to compare with. Another reasonable method might give somewhat different positions.
• The issues shape the axes. With other issues, the axes would measure something slightly different; that is why an artificial intelligence model reviews the list under written rules, the project's editor approves it and we publish it.
• Outlets write about different issues. That is why issues count equally and we require positions on at least 5 of the 10. The cost: an issue with two texts counts as much as one with fifty; this shows as a larger ellipse of uncertainty.
• The twelve-month window is slow. A sudden change of line, for example after a change of editor, will show gradually.
• A headline is weaker evidence than an opinion text. It shows a choice of words, not an argument. That is why headlines count separately and are joined with opinion texts only where they agree with them.
• An outlet with few columnists is seen through them. That is why we count columnists, require at least three and show each one separately.
• Models make mistakes, for example with irony or with who is speaking in a text. That is why there are the review without the outlet in view and the monthly random check (step 8).
• The check is itself done by a model, from another company: two models can make the same mistake, and their agreement is not confirmation by people. In the only comparison with a person, on 289 headlines, the artificial intelligence check agreed 71% of the time (κ 0.57) and found positive or critical presentation more often than the person did. These headlines are not part of the check's sample.
• The map does not see the choice of stories, that is what an outlet highlights or leaves out. That shows in the coverage data.
• More axes do not mean more precision. The additional axes rest on fewer texts, so they place fewer outlets and with greater uncertainty.
• The axes are not independent of each other. Among the Greek parties, for example, Europe and Institutions move together (0.77). The map notes it when the two axes the reader chose move together among the outlets.

Technical annex

Formulas

Convention: on each axis, the pole named first is −1 and the second +1 (State intervention −1, Free market +1; Progressive −1, Conservative +1; Europe & the West −1, Sovereigntism +1; Dialogue −1, Hard line +1; Institutional −1, Anti-system +1). On the map −1 is left or bottom and +1 right or top. On the map and on every page, a position x is written as 3·x with one decimal, that is on a scale from −3 to +3; the formulas and thresholds below refer to the value from −1 to +1. For columnist v of outlet m (the editorials count as one) and issue i in the twelve-month window, n(+), n(−) and n(0) are their texts with a position towards +1, towards −1 and mixed. For headlines, e(+), e(−) and e(0) are the stories where the outlet's position on the issue was towards +1, towards −1, or tied (headlines of it towards both sides, in equal number).

s(v,i)   = [n(+) − n(−)] / N,   where N = n(+) + n(−) + n(0)
sγ(m,i)  = Σ w(v)·s(v,i) / Σ w(v)
           (columnists with N ≥ 1; w = 2 for the editorials, 1 for each columnist)
sτ(m,i)  = [e(+) − e(−)] / E,   where E = e(+) + e(−) + e(0), and counts only if E ≥ 3
xγ(m,a)  = Σ b(i)·sγ(m,i) / Σ b(i),   where b(i) = N(i) / [N(i) + 2]
           (issues of axis a with a position, N(i) the outlet's texts on issue i;
            placement only if k ≥ K/2, where k is the issues with a position and K is
            the number of issues of the axis: 10 or 8, and if at least 3 columnists took a position)
xτ(m,a)  = the same with the sτ(m,i)
x(m,a)   = [xγ(m,a) + xτ(m,a)] / 2   when the axis joins the two sources and
           the ranges of uncertainty of the two positions overlap
x(p,a)   = the xγ of party p, from the sections of its official texts
           (the party counts as one columnist; placement only if k ≥ K/2)
g(m,t) = (p − c) / (c + p + u)
         (p, c, u: stories with a positive, critical or neutral outlet stance in month t;
          −1 = all critical, +1 = all positive)
R²(a)  = 1 − Σ(y − ŷ)² / Σ(y − ȳ)²
         (y: the positions of the CHES 2024 parties on the candidate axis a;
          ŷ: the prediction of a linear regression on the existing axes)

The topic comparison shows each issue's sγ(m,i) (and sτ(m,i), where headlines are joined with opinion texts), not x(m,a). For issues off the axes, s measures support for the wording (+1 = all for).

The Greek parties in CHES 2024

The positions of the 9 Greek parties of CHES 2024, on the scale from −1 to +1 of our axes (three times larger on the map's scale: SYRIZA, for example, −1.1 on the economy and −2.1 on society). They are not shown on the map; the order of the parties that comes from their official texts is checked against them.

How they are derived:

• Economy: the CHES lrecon scale (0 to 10), as (x − 5) ÷ 5.
• Society: the GAL-TAN scale (0 to 10), as (x − 5) ÷ 5.
• Europe & the West: the average of three questions. The stance towards European integration and towards the EU's powers in foreign policy (support scales from 1 to 7) are converted as (4 − x) ÷ 3, so that +1 is the sovereigntist end. The relationship with Russia (0 = more defence, 10 = more trade and diplomacy) is converted as (x − 5) ÷ 5.
• Institutions & system: the average of the question whether "the people" or elected representatives should decide and of the emphasis on anti-elite rhetoric (both 0 to 10, as (x − 5) ÷ 5). The second measures how strongly a party uses such rhetoric, not a policy position; so it is used only as a check.
• National issues: CHES has no matching question.

ELAS and Hope for Democracy, both founded in 2026, MeRA25 and New Left are not included in CHES 2024. We read their official texts as we read the other parties', and they appear on the map by the same rule (step 7); they do not count in the check of the parties' order against CHES, since CHES does not place them.

Uncertainty

• A two-stage bootstrap for opinion texts: from the outlet's V columnists, V are drawn with replacement, and for each columnist drawn, as many of their texts with at least one position are drawn, with replacement; each text brings all its positions. For headlines, the stories are drawn. Where the two sources are joined, each repetition draws both; 1,000 repetitions.
• The one-columnist check: for each columnist, the position is recomputed without them. If it falls outside the interval from the 5th to the 95th percentile of the full estimate, the axis is marked "depends on one columnist".
• The ellipse is centred on the dot, with semi-axes of half the interval from the 5th to the 95th percentile on each axis. It is not a region holding 90% of the 1,000 positions on both axes together; such a region would be larger.
• The ± on an axis is half the width of the interval from the 5th to the 95th percentile.

Thresholds

Versions

The list of issues has a version number; every addition, removal or rewording of an issue makes a new version. Each version is reviewed by an artificial intelligence model under written rules and approved by the project's editor before publication. Each publication of the map states the version of the list and the time window it used. All versions stay in a public history. This page is version 1.8 (draft) of the methodology, of 13 October 2026. Version 1.8 places the parties from their official texts, with the same issues and the same calculation as the outlets; CHES 2024 stays as a check (step 7). Since 13 October we also read ELAS, Hope for Democracy, MeRA25 and New Left, which CHES does not cover. Version 1.7 gives each issue a weight that grows with its texts («Position on the axis»): an issue with a single text no longer weighs as much as one with forty. Version 1.6 retires four checks: the comparison of the map with academics' placement of the outlets, the comparison of the stance towards the government with theirs, and the two headline checks, since headlines are not used (step 8). Version 1.5 wrote every position on the map on a scale from −3 to +3 and showed the parties where CHES 2024 places them; the reading of their programmes stays as a check. On 5 October 2026 the page “How Protagoras works” became a short explanation with figures, and the detailed rules stayed here. The method did not change; descriptions that were not true were corrected, among them: which model decides which columnists count (step 2), what the ellipse shows (Uncertainty), an editors' review when an outlet moves outside its ellipse, which does not happen, that the additional axes appear only once they pass their checks, that headlines are not used as a source for the map, and how the stance towards the government is published.

Sources

• Chapel Hill Expert Survey (CHES) 2024. 609 political scientists place 279 parties from 31 European countries (economic axis, GAL-TAN, European integration, Russia, "people or representatives" and more). Data and a codebook of the variables: https://www.chesdata.eu/ches-europe/ · Rovny, J. et al. (2025), Electoral Studies 97, https://doi.org/10.1016/j.electstud.2025.102981
• Katsanidou, A. & Otjes, S. (2016). "How the European debt crisis reshaped national political space: The case of Greece". European Union Politics 17(2), 262-284. https://research.rug.nl/en/publications/how-the-european-debt-crisis-reshaped-national-political-space-th/
• HelpMeVote. The voting advice application of the Aristotle University of Thessaloniki; the Greek precedent for placing parties by specific issues. https://helpmevote.gr/info/index.html
• Budak, C., Goel, S. & Rao, J. M. (2016). "Fair and Balanced? Quantifying Media Bias through Crowdsourced Content Analysis". Public Opinion Quarterly 80(S1), 250-271. The news reports of major American outlets show little partisan slant; the slant is much stronger in opinion texts. https://academic.oup.com/poq/article/80/S1/250/2223443
• Gentzkow, M. & Shapiro, J. M. (2010). "What Drives Media Slant? Evidence From U.S. Daily Newspapers". Econometrica 78(1), 35-71. A newspaper's slant is measured from the words it chooses for the same topics; the use of headlines rests on this. https://doi.org/10.3982/ECTA7195
• Rozado, D. (2024). "The political preferences of LLMs". PLOS ONE 19(7), e0306621. Most conversational language models show left-of-centre preferences in political orientation tests. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0306621