Koppenteller

How we counted

How the count was made, what the numbers do and do not show, and how sure they are.

21,574 headlines from fifteen Dutch news outlets, from January 2025 to November 2026. They mention someone 9,339 times where the headline shows whether that person is a woman or a man. 26% of them are women: about one in four. The margin runs from 24% to 28%. Everything was read by a language model. People check samples on an ongoing basis. Until the first agreement rate is published, Koppenteller is called a pilot.

Which headlines

The headlines come from the public news feeds (RSS) of NOS, nu.nl, AD, De Telegraaf, de Volkskrant, NRC, RTL Nieuws, Trouw, the FD, the Reformatorisch Dagblad, Hart van Nederland, Omroep Brabant, RTV Noord, L1 and Rijnmond. A feed is a list of a site's newest headlines. Most headlines come from copies of those feeds in the Internet Archive (archive.org). It holds about one copy per feed every ten days. RTL Nieuws and NRC were collected once, on 1 November 2026. For Trouw, the Reformatorisch Dagblad and Hart van Nederland no feed was saved. For them, the first thirty headlines on the front page were taken from the same kind of archive copies; the date is then the day the headline was on the front page. These eight sources were added on 6 November 2026. The articles themselves were not read.

What is in the set depends on which feeds were saved. For nu.nl, 24 feeds were saved, including sport. For NOS, 17. For De Telegraaf and de Volkskrant, only the front page. RTL Nieuws has one feed with everything, including showbiz. October 2026 is much bigger than the other months. RTL Nieuws is only in October and early November 2026. Sometimes the same headline was in two feeds of one site, and it now counts twice. This concerns a few dozen headlines. Without them, the main figure stays at 26%. In the next version of the data, they will be counted once.

That is why the main figure comes with a range. If every site counts equally, it is 28%. If every month counts equally, 26%. Without RTL Nieuws it is 25%, without the separate sports feeds 27%. All of these fall within the margin of 24% to 28%.

The main figure does depend on how much sport is in the feeds. Of the 9,339 people, 2,986 are in a sports headline, almost a third, and 15% of them are women. Without all sports headlines, 2,010 of the 6,353 people are women: 32%. That is more than without the separate sports feeds, because the general feeds also carry sports headlines. Showbiz works the other way: 44% are women there, and without showbiz the main figure is 24%. So which feeds of a site were kept helps decide where the figure ends up.

The headlines are in Dutch. On the English pages they stay in Dutch, because they are quotes.

Who did the reading

Every headline was read by a language model, following a fixed handbook. That language model is Claude, made by Anthropic. The work was split over several sessions of the same model. Each session read one part. For every person in a headline, it recorded:

The reading was done in two rounds: first 2,000 headlines, later the rest. The first round comes out more than three points lower than the 7,308 headlines after it, which come from the same sites and the same months (21% against 25% women). Those 2,000 headlines were therefore read again in November 2026, also by the language model. That gave almost the same result: 21.1% against 20.9%. So the difference is not caused by the reading. Where it does come from is not yet clear. In the first round, people whose gender is unknown were not recorded.

For the deeper analysis, more was read for each person. Does the headline describe someone through a partner, parent or child? Or by age or looks? Is the person quoted? And what job did the person have at that moment? Ministers and junior ministers were matched by name and date with who was in the cabinet at the time.

What “checked” means on this site: a second reading and targeted checks, also by a language model. In addition, people check samples on an ongoing basis. Where the site talks about sorting or checking, it means the work of the language model.

When someone counts

A person only counts if the headline itself shows whether they are a woman or a man, or if it is a well-known person. That can be a name that is clearly female or male. Or a word like mother, businessman or actress. Or a pronoun like he or she. Or someone like Femke Kok or Max Verstappen.

‘The judge’, ‘a top executive’, ‘the suspect’, ‘a Russian’ or ‘a student’ say nothing about gender. So they do not count. That applies to 20% of all people in the headlines, and to a third of the people in crime headlines. Headlines more often mention men without a gender. Because of that, the share of women among victims and suspects may come out a little too high.

Groups (‘three men’) do not count in the shares. People who do not identify as a woman or a man get their own code. Those mentions do not count in the shares of women and men either.

All numbers are mentions: someone who is often in the news counts every time. For world leaders, the number of different people is given as well.

How sure

The margin of the main figure was calculated like this: the count was redone many times, each time with a different mix of sites and months (bootstrap). This takes into account that headlines came in clusters. For the parts and the comparisons with reality, the margin is calculated more simply, per mention (Wilson). There the real uncertainty can be somewhat larger. On this site, a difference from reality is only called a difference if it falls outside the margin.

Next to reality

Where possible, the figure from the headlines is shown next to an official figure. Sometimes that measures exactly the same thing. Then it says ‘in reality’. Sometimes it measures only roughly the same thing. Then it says ‘for comparison’.

Ministers and junior ministersin reality: who was in the cabinet at that moment · 57 of 103 mentions
55%39%
Different world leadersin reality: share of countries led by a woman · 8 of 78 leaders
10%13%
Victims of murder and manslaughter in the Netherlandsin reality: Statistics Netherlands (CBS) 2024–2025 · 21 of 44 cases, difference within the margin
48%37%
Company leadersfor comparison: Fortune 500, the largest American companies · 10 of 117
9%11%
Suspects, accused, convictedfor comparison: suspects of violent crime (CBS 2023); here fraud and scandals count too · 102 of 845
12%12%
Sports headlines about womenfor comparison: women in the Dutch team at the Paris Olympics (163 of 273) · 325 of 2,803
12%60%
Business headlines about care, education, childcarefor comparison: share of all workers in those sectors (CBS 2025) · 9 of 759
1.2%25%

For members of parliament, the comparison is with an expectation. Party leaders are in the news far more often than ordinary members. So we calculated how many women you would see if the news only looked at that position. Anyone who was a minister at the time only counts for the cabinet. (Caretaker ministers were allowed to be members of parliament for a while.)

Only 2% of mentions of world leaders are about a woman. That is because the same men keep coming back: half of the mentions are about Trump.

Next to other research

Is it really one in four? Other research arrives at the same figure, with a very different method. The largest is the Global Media Monitoring Project (GMMP). Since 1995, every five years, it counts who is in the news on one day, in dozens of countries at once. People read each story in full, not just the headline.

StudyWhat was countedWomenHere
GMMP 2025, worldEveryone in newspapers, radio and TV, 94 countries, 6 May 202526%26%
GMMP 2025, EuropeSame, 25 countries27%26%
GMMP 2020, NetherlandsEveryone in 6 newspapers, 3 radio and 3 TV bulletins, 4 news sites, 29 September 2020 (527 people)28%26%
GMMP 2025, NetherlandsNewspapers, radio and TV, one day (about 360 people)31%26%
GMMP 2025, sportPeople in sports news, world15%12%
Gender Gap Tracker, CanadaWho is quoted, 613,000 articles, 2018 to 2020, counted automatically29%31%

‘Here’ is the figure from the headlines: all people, the sports headlines, and who is quoted in the headline (the last one in the headlines of the first seven sources).

The methods differ a lot. The GMMP counts everyone in the whole story, on one day. Koppenteller counts only the headline, over almost two years. Headlines mostly feature people with a position or a result: the prime minister, a world leader, a football player. Those are more often men. That Koppenteller comes out a little below the Dutch GMMP figures fits with that. The Dutch GMMP samples are small, though: one day, a few hundred people. Koppenteller also only counts people whose gender is clear from the headline. The GMMP 2020 shows that such a choice can move the figure by about 2 points (28% without, 26% with the people whose gender was unknown).

The Dutch Media Authority (Commissariaat voor de Media) counts something else: who appears on Dutch TV, including presenters and guests. In news and current affairs that was 35% women in 2023. That is higher because presenters and guests count.

The sites

As they came in, the sites differ a lot: from 18% at nu.nl to 38% at RTL Nieuws. That is mostly because of the feeds. At nu.nl, 68% of the people in the headlines are athletes. At RTL Nieuws, 52% are in showbiz. If you compare the same topics, all sites are between 22 and 27%, and the margins overlap. So this count says nothing about which newsroom is better or worse. News from the press agency ANP counts for the site that published it.

SiteHeadlinesPeopleWomen
NOS2,80694022%
nu.nl2,9352,11418%
AD1,07984526%
De Telegraaf41432226%
RTL Nieuws83473938%
de Volkskrant1,05745626%
NRC1634326% (small)
Trouw1,52759725%
FD92818717%
Reformatorisch Dagblad1,67854229%
Hart van Nederland81430641%
Omroep Brabant1,02449531%
RTV Noord2,18453428%
L11,75034837%
Rijnmond2,36187028%

People: everyone whose headline shows whether they are a woman or a man. Small: fewer than fifty people, where a percentage says little. The numbers follow what was saved from each site, not what the site published. You can check them in the open data, column bron.

What this does not show

Names and privacy

The card shows the headline for every figure. All names are blacked out, including those of well-known people and people in public office, and including first names and initials. The point is the ratio of women to men, not who it is. The names are not in the site's source code either. Where the article is still online, the headline links to it.

A script blacks out the names. All headlines were then checked again for missed names and partial names (5 and 11 November 2026). This is done with a script and a language model, and may miss something. Do you still see a name? Email astrid@missingmeasure.com. See also the privacy statement.

The headlines were only read to answer these questions. They were not used to train a language model. The articles are not published.

Earlier research

Koppenteller builds on earlier research. Since 1995, the Global Media Monitoring Project has measured worldwide every five years who is in the news. In 2015, Shor and others showed where the shortage of women in newspapers mostly comes from: the top of society. Whoever has power is in the news (‘A Paper Ceiling’, American Sociological Review). Koppenteller finds the same for the cabinet and world leaders. The Canadian Gender Gap Tracker continuously counts who is quoted in the news.

Koppenteller also looks, for each person, at why they are in the headline and how they are described. And people with a job are compared with who held that job on the day of the headline.

Sources

Per day

The count above is a fixed version: headlines from January 2025 to early November 2026, taken from archive copies. It stays as it is. Since 16 November 2026, Koppenteller also collects headlines itself from the feeds of the fifteen media, a few times a day. Those headlines are counted per day, on the page Per day.

A day runs from midnight to midnight, Dutch time, by the moment the headline appeared in the feed. Every night an AI language model reads all headlines of the previous day, with the same handbook and the same instructions as the fixed count. It counts who is in the headline and whether that person is a man or a woman, in the same way as above. The result is online by seven the next morning at the latest. A person does not check it in advance. A person does check a sample of one hundred headlines every week. What the person changes is in the data and counts from then on.

One day is a small count with a wide margin. On an ordinary day it is just over a thousand headlines and about six hundred people whose gender shows in the headline. The margin is then about four to five percentage points either way. Per medium the count is much smaller still: for NOS or a regional broadcaster it is a few dozen people a day. That is why a medium with fewer than thirty people gets no percentage. One day says little about a medium. What counts is what happens over weeks and months.

A feed only shows the newest headlines. If a lot of news appears between two collection rounds, a headline can be missed. The log keeps track of where that may have happened; this is listed with each day in the data. The data per headline is available per day as CSV on the page Per day, under CC BY 4.0. It contains no headlines or names, only the reading.

Open data

All the counts are available as CSV on the open data page, with every column explained. That way you can check every number on this site yourself.

Spotted a mistake?

Is a headline, a number or a classification wrong? Email astrid@missingmeasure.com. Corrections will appear here, with a date.

Version 19 November 2026. Change: the check by people is ongoing, with samples. Version 14 November 2026. Change: added an explanation of sport and of the two reading rounds; Koppenteller is now called a pilot; the language model is named; the duplicate headlines are explained; the data page has a citation line. Version 11 November 2026: all names are now blacked out, including those of well-known people.