Skip to content
Home Projects Typhoon Analysis
🌀 Disaster Analysis Python Data Visualization 2014-2020 Dataset

There Is No Such Thing As What The World Is Reading

One week, 18 countries, 14 language editions of Wikipedia. The median pair of countries shares 2.81% of its most-read articles, 60 of 91 pairs of language editions share none at all, and India's entire top list sits on English Wikipedia.

2.81%
Median overlap between two countries' top lists
60 of 91
Language pairs sharing no top article at all
100.0%
Of India's most-read list is English Wikipedia
6 of 24
Countries the API will not report at all
Key Takeaways

Across 2026-08-24 to 2026-08-30 the 18 countries that Wikimedia will report on shared a median 2.81% of their most-read articles, and 16 of 153 country pairs shared nothing. Between language editions it is starker: a median of 0.0%, with 60 of 91 pairs sharing no article at all.

  • India's 140 most-read entries over the week were 100.0% on English Wikipedia and 0 on Hindi Wikipedia. Nigeria is also 100.0%, the Philippines 97.83%.
  • Japan is the opposite pole at 0.0% English and 100.0% Japanese; France reads English for 2.14% of its list. This is a split, not a spectrum: 6 countries are above 80%, 9 below 10%, and only 3 anywhere between.
  • On 2026-08-15, 13 language editions had 12 different most-read articles. Hindi Wikipedia's was India's Independence Day; Italian Wikipedia's was Ferragosto, Italy's own public holiday on the same date. Neither shows any sign of the other.
  • Per-country view counts are rounded by Wikimedia on purpose so that reading cannot be traced to individuals. Every country figure here is approximate by design, and a check asserts none is ever labelled exact.
  • 6 of the 24 countries asked for return no data at all — Russia, Egypt, Vietnam, Turkey, Pakistan and Bangladesh. That is 25.0% of the sample, and it is not a random quarter.

A Quarter Of The World Cannot Be Measured

Before any finding, the limit that shapes all of them. The per-country endpoint answers 404 for six of the 24 countries asked for, with the message that the country is not loaded. This is not a fetch failure and no amount of retrying changes it.

Entries returned per country, of a possible 140

Countries with no data

6 of 24

Russia, Egypt, Vietnam, Turkey, Pakistan and Bangladesh — 25.0% of the sample, and between them roughly a billion people.

Countries reporting

18

Everything else on this page describes these, and says so rather than saying “the world”.

Shortest list

15 of 140

Nigeria answers, but only this many articles cleared the privacy floor across the week. A smaller audience gets a shorter list as well as a rounder one, so the data is thinnest exactly where it would be most interesting.

The Median Pair Of Countries Shares Almost Nothing

Overlap is the share of articles two countries both had in their top 20 on any day of the week — a Jaccard index over sets, not a comparison of rankings, because rank order is noisy and the question is whether the same things are being read at all.

Overlap between country pairs, most to least similar

Median overlap

2.81%

Across all 153 pairs of reporting countries.

Pairs sharing nothing

16

Not one article in common across seven days. Japan shares nothing with Italy, Poland, Thailand or Argentina.

The most similar pair

21.21%

United States and United Kingdom, sharing 28 articles. That is the ceiling: the two countries with a shared language, shared media and shared celebrities still disagree about four fifths of what they read.

Between Languages It Is Worse

The same measure on the 14 language editions directly, using their exact view counts rather than the rounded per-country ones. If there were a global conversation this is where it would show up.

Overlap between language editions, most to least similar

Median overlap

0.0%

The median pair of language Wikipedias shares no most-read article at all.

Pairs sharing nothing

60 of 91

65.9% of pairs, over a full week and a top 20 each day.

The most similar pair

7.69%

English and French. Even the closest two editions overlap by less than a thirteenth.

Who Reads In English

For each country, the share of its most-read entries that sit on English Wikipedia rather than on any other edition. The shape of this distribution is the most interesting thing in the dataset.

Share of each country's most-read list that is English Wikipedia

India

100.0%

All 140 entries across the week, and 0 on Hindi Wikipedia — which exists, and has hundreds of thousands of articles.

Japan

0.0%

The other end. 100.0% Japanese, and France reads English for 2.14% of its list.

A split, not a spectrum

6 / 3 / 9

Countries above 80% English, between 10 and 80, and below 10. Only 3 sit in the middle, and Indonesia at 47.14% is the clearest of them.

One Date, Two National Holidays, No Shared Awareness

2026-08-15 is Independence Day in India. It is also Ferragosto, a public holiday in Italy. Both show up clearly — each in exactly one language edition and nowhere else.

Most-read article per language edition, 2026-08-15

Hindi Wikipedia

32,679

Views on Independence Day (India), its most-read article that day.

Italian Wikipedia

55,302

Views on Ferragosto, Italy's own 15 August holiday, and the top article there.

English Wikipedia

987,162

Views on an unrelated biography — 30.2 times the Hindi article, on India's own national holiday. 13 editions produced 12 different top articles.

What This Data Cannot Tell You

Five limits. Two of them are unusual enough to be the most interesting thing about the source.

The country counts are rounded on purpose

The API field is literally views_ceil. Wikimedia rounds per-country figures so that reading behaviour cannot be traced back to individuals, which is the right decision and makes every country number here approximate by design. A check asserts that no country row is ever labelled exact. The per-language counts are unrounded, which is why the two tables are kept apart rather than merged.

A quarter of the sample is simply absent

Russia, Egypt, Vietnam, Turkey, Pakistan and Bangladesh return 404. Any claim about global reading built on this source is really a claim about the countries Wikimedia publishes, and the ones it does not publish are systematically not the wealthy English-speaking ones.

A pageview is a request, not a read

There is no dwell time, no scroll depth and no way to tell a person who read an article from one who bounced off it. all-access also includes automated traffic that Wikimedia could not classify; a user-only filter exists per project but not per country, so neither table uses it, because comparability matters more here than precision.

Wikipedia is not the internet

It is one encyclopaedia, unusually open about its numbers. Countries where people get reference information from a search summary, a messaging app or a walled platform will look quieter here than they are, and that bias also is not random.

One week is one week

Seven days from 2026-08-24 to 2026-08-30, plus 2026-08-15 as a separate named case study that is not mixed into the weekly figures. A different week with a World Cup final or an election in it would move every overlap number up, and the direction of that bias is knowable while its size is not.

Method

One fetcher, nine CSVs, and more work in the filtering than in the analysis.

Namespaces come from MediaWiki, not from a pattern

Main pages, searches and category listings are navigation, not reading, and the main page alone is usually the largest single entry in every country — keeping it would make every country look alike for a reason unrelated to interest. Matching them by pattern does not work: Special:Search and Spider-Man: Brand New Day are the same shape, and there are 288 distinct colon-prefixes across the 98 Wikipedia editions that appear. The namespace names are fetched from each edition's own API instead, which is exact.

Two bugs that filter caught only in a browser's worth of detail

A regex over letters missed the Hindi, Vietnamese and Thai spellings of Special:Search, because Python's \w excludes the combining marks those scripts use. And the first namespace fetch kept only namespaces numbered above zero — but Special is namespace -1, so the one that mattered most was excluded. Both are recorded in the fetcher rather than quietly fixed.

Only Wikipedia, not all of Wikimedia

The per-country endpoint spans every Wikimedia project, and its wiktionary, wikibooks and commons entries are almost entirely Special:RecentChanges traffic on tiny editions. 684 such entries were dropped, counted in the coverage file rather than discarded silently.

Overlap is a set measure

Jaccard over the set of articles each country or edition read during the week, not a rank correlation. Rank order moves around day to day for reasons that are not interesting; whether two populations read the same things at all is the question.

The holiday is a case study, not a data point

2026-08-15 sits in its own table and is excluded from every weekly figure. Picking a national holiday and then reporting it inside an average would be choosing the result first.

Licensing

The pageview datasets are released CC0. Article titles are content and carry the editions' own CC BY-SA terms; they appear here as data points rather than as article text.

Key Findings & Summary

  • The median pair of reporting countries shares 2.81% of its most-read articles, and 16 of 153 pairs share none.
  • Between language editions the median is 0.0%, with 60 of 91 pairs — 65.9% — sharing no article at all.
  • India reads English Wikipedia for 100.0% of its top list and Hindi Wikipedia for 0 of 140 entries. Japan reads English for 0.0%.
  • It is a split rather than a spectrum: 6 countries above 80% English, 9 below 10%, and 3 in between.
  • On 2026-08-15, 13 editions had 12 different top articles — India's Independence Day in Hindi, Ferragosto in Italian, and in English a biography that outdrew the Hindi article 30.2 times over.
  • Per-country counts are rounded by design, and 6 of 24 countries return no data at all.

Sources & Citations

Every figure on this page traces to one of these, through a CSV in data/global-wikipedia/. Each is checked against its source query on every build.

Wikimedia Analytics: top articles per countryPrimary

The most-read articles in each of 24 countries across every language project, 24-30 August 2026, top 20 per day after filtering navigation pages — Counts are returned as views_ceil — deliberately rounded so per-country reading cannot re-identify anyone — so every country figure here is approximate by design. Six of the 24 countries asked for return 404: the API says the country is not loaded.

Wikimedia Analytics: top articles per projectPrimary

The most-read articles on each of 14 language Wikipedias, same window, plus 15 August 2026 as a named case study — These counts are exact, unlike the per-country ones. all-access includes automated traffic Wikimedia could not classify; no user-only filter is applied, because one exists per project but not per country and the two tables have to stay comparable.

Wikimedia Foundation data licensingPrimary

Licence terms for Wikimedia analytics data — The pageview datasets are released CC0. Article titles are content and carry the projects' own CC BY-SA terms; this analysis publishes titles as data points, not article text.

Have Data But Need Help?

If you have raw data and want to transform it into meaningful insights like this analysis, let's connect. I can help with data cleaning, visualization, and actionable recommendations.

Have a dataset you'd like analyzed or need a mini AI project as a starter guide? Send me your suggestions!

Connect on LinkedIn Dev.to Articles 💬 Messenger

Prefer the plain-English version?

I wrote a companion post — “I Wanted To Know What The World Reads. There Is No Such Thing.” — in simple, everyday words.

Read the blog post →