Every figure we have published and then had to change is listed here, with what it was, what it became, and why. Nothing is edited away. If you have quoted one of our numbers, this page tells you whether it still stands.
A number that changes without notice is worse than a number that was never published, because somebody has already used it. On 7 August 2026 we found that one flaw — matching Korean titles by their text alone — had put wrong figures on seven pages and four articles at once, and that a second, unrelated flaw had been misreading a government table for as long as we had used it. Everything that came out of that is below.
We also publish how each number is made, and every data page names its source and says what it leaves out. This page is the other half of that: what we got wrong anyway.
Every correction here is filed under the mistake it came from. Counting them this way answers the question a list cannot: are we making the same mistake twice?
9 of the 36 come from one cause, and it is the one we wrote an article about. 7 other causes have happened more than once. If a number here starts climbing, that is the thing to hold us to.
12 of the 16 causes now fail the build if they come back. 4 do not, and we would rather show you that list than a sentence about taking accuracy seriously. A guard is a check that fails the build if this kind of mistake recurs. Null means we have not written one yet — it is a work list, not a claim of coverage. Which ones, and why →
The places were selected on Wikidata's country property (P17). That property records involvement as well as location: the Battle of Khe Sanh carries South Korea because South Korea fought there, and the Order of Friars Minor, two Korean Air crash sites, and the Antarctic Jang Bogo Station entered the same way. We re-tested every place against its coordinates instead, because a location does not follow a belligerent. The nineteen accounted for 4.61% of all reads in the panel. Our first attempt at this check also used the country property and reported no problem at all — a check that reuses the ruler being checked always agrees with it.
Our rule for deciding which Netflix titles are Korean matched title text against Wikidata labels in any language, so 23 titles written in Arabic, Hebrew, Cyrillic and Japanese script were counted as Korean. They are local works: Netflix does not translate titles in these lists — Squid Game holds 445 places across ten Arab markets under its Latin name, and only 10 of the 1,530 distinct titles on Egypt's chart are written in Arabic script. The rule now ignores non-Latin labels. Across the whole site this moved Korea's share of world chart places from 37,962 places to 37,750, both of which round to 7.7%.
The sentence was written by hand and never measured. 2005 is the year with the least regional detail, and the largest cell in it was the one we called empty. The page now reads the exception out of the data.
Eight titles were in a Korean panel although our own attribution query — the one that produces the shared/unambiguous split we publish — returned no Korean work for those exact names: Waterworld, Re/Member, Into the Storm, Wildflower, Feng Shui, Glorious Days, Long Live Love! and You and Me. This was not a judgement call; it was our list contradicting our own instrument. A check now fails the build if any panel title is in that state.
The peak-group table was computed on the fifty largest titles by hours. Inside that group every title has already had a long run, so run length looked irrelevant when it is in fact what put those titles in the group. Recomputed on all 235 titles, with medians beside means because one title holds 5.05bn of 23.72bn hours.
The panel is built on English Wikipedia pageviews, so we treated the missing acts as uncountable. Wikidata records the same acts whether or not English Wikipedia does, and we already query it for the roster. Counting with and without an English sitelink gives a floor on the blind spot.
Korean titles were matched by title text against a Wikidata list. 52 titles Netflix classes on its English-language charts — The Perfect Couple, Suits, The Circle, Hit Man among them — were counted as Korean. Netflix splits its charts by a title's primary language, so only the Non-English charts are counted now. Hand-checking the largest remaining titles removed 8 more.
Same cause as /watched. The page was published after the correction and has never shown the wrong totals, but it rests on the corrected set and says so.
KOSIS table DT_113_STBL_1025706 classifies on two levels — form of export above, type of company below — and returns only the lower name, so several rows a year read 'terrestrial broadcasters' and differ by code alone. We had taken the finished-programme row for three of the four company types while the fourth entered whole. Each type is now the sum of every export form published for it. The residual we described as unattributed did not exist; we had not added it up.
Companies that disclose headcount but not pay were left in the denominator, so their staff were averaged in as if they earned nothing — 456 people in content and 20,613 across the market. Pay is now weighted over disclosing companies only. Tenure, headcount and the gender split come from companies that disclose them and did not move.
The front page data file was hand-made and could not be rebuilt, so it did not follow when the matching rule was corrected. It is now built by a script from the same rule as every other page.
Same matching cause. 42 titles were not Korean — among them two Witcher animations, I Am Legend and Elysium. The distribution the page describes did not change shape; it sharpened.
Same cause as the front page — a hand-made data file that could not follow a corrected rule. It is now built by a script.
Article corrections also appear at the top of the article itself, where anyone reading it will see them. They are repeated here so the record is in one place.
People were being de-duplicated by name alone, which merged different people who share a name. Keyed on name plus birth date, this panel holds 1,317 distinct people rather than 1,295, and 1,303 rather than 1,282 after removing 1 January. Every figure in this article moved a little: March 1.20x is unchanged, July went from 0.70x to 0.69x, and the noise test from 1.2% to 1.1%. March fullest and July thinnest still holds. The defect was found while building a wider panel of 9,249 Korean entertainers, where 288 names carried more than one birth date.
We widened this measurement from 30 people to 100 the same day. At 30, one person was read mostly through a title Wikipedia no longer uses; at 100 it is six, and all six turn out to have a rename in the move log. The figures in this piece are correct for the 30 measured here and are not edited. The fuller count, with all six named and the three different kinds of rename separated, is in Six Korean stars were renamed on Wikipedia. Readers followed none of them.
The Southeast Asia panel is 420 titles, not 421. This article said 421. The change is ours: on 25 August we removed three titles that were never Korean — Dangerous Liaisons (the French film Netflix released on 8 July 2022; it took no chart place in South Korea and one in all of Asia), Breathless (a Spanish series from 2024, charting in Spain, Argentina, Uruguay, Chile and Portugal) and One More Time (a Swedish film from 2023, charting first in Sweden). All three entered our figures through the same flaw this article is about: we match chart rows to Korean works by title text, so a work of any nationality sharing a title with a Korean one comes in with it. Between them they carried 473 chart places we had been counting as Korean. One of the three sat inside this panel, which is why the denominator moved by one. We now have a check that proposes such titles from their market pattern rather than waiting for a person to notice one; it proposes only, and each of these three was confirmed against Wikidata before it was removed. It proposes 153 more we have not yet worked through, and it does not catch every kind — a title where a Korean work and a foreign one of the same name are merged into a single row is invisible to it, because our market figures are not broken down by year.
This article said the rise stops at ten charting titles: actors with ten or more were looked up slightly less often than those with six to nine, and we called that a ceiling. It is not there. Two things changed. Three more weeks of chart data arrived, and a spelling fault in how we matched Netflix's title text to Wikidata was fixed — Netflix writes Escape From Mogadishu where Wikidata writes Escape from Mogadishu, and forty-four Korean titles were being dropped for that reason alone. The ten-or-more band went from twenty-three actors to thirty-nine, and its median went from below the six-to-nine band to well above it. The band was too thin to carry the claim we put on it, and we did not say so at the time. The median now rises at every step. What survives unchanged is the finding in the headline: the busiest actors are still not the most looked-up.
This article said there were nine kinds of mistake and that all nine failed the build. Both numbers were out of date, and the second was the worse of the two. Six corrections recorded on 15 August had never been given a cause, and reading them produced three kinds we had not named — a title set widened after publication, the word median computed two ways, and a measuring window short enough to make its own highest month look like a peak. A fourth kind, a Wikidata property that records who took part being read as where a thing is, had a cause but no entry in the guard table at all, which is why our own check counted it as neither guarded nor unguarded. There are fourteen kinds; ten fail the build and four do not, and the four are now named in the article rather than counted. The check behind this piece now also fails if a cause exists with no row in the guard table, which is the hole that let this go quiet.
Published first with 35 titles and 16 in the table. We widened the set to 59 the same day, and every count here grew: 26 titles timed rather than 16, nine halving within a month rather than five, nineteen returning above half rather than twelve. The median half-life did not move — it was two months at 16 titles and is two months at 26. The mean came down from 2.9 to 2.6. One title, Big Mouth, has now been watched for 26 months without falling below half, where previously none had held. The direction and size of the finding are unchanged; the counts are the new ones.
Published without naming the method or its limits. The check described here is a jackknife — leave one observation out, recompute, repeat — introduced by Quenouille in 1949 and named by Tukey in 1958. That matters, because the literature already establishes what it can and cannot do for a median: removing one observation barely moves a median, so the jackknife understates how much a median varies, and Miller showed in 1974 that it is not consistent for the median at all. The consequence runs one way. A swing this check finds is real, because a method that understates variation found it anyway. A swing it does not find is not evidence of stability — it may simply be one this method cannot see. We had been publishing that second case as a result, with the words 'this median holds' on two pages. Those pages now say what the check actually establishes, and this article now names the method, its origin and its known failure for medians. Measuring stability properly would need a delete-d jackknife or a bootstrap (Efron 1979); we have built neither, and we will not imply otherwise.
The two medians here were computed by taking the upper of the two middle values rather than their average. Both samples are even-numbered, so this mattered: the managers' median was published as 82.4% when the middle two values are 80.3% and 82.4%, and the players' as 50.6% when the middle two are 50.5% and 50.6%. They are now 81.35% and 50.55%. Taking the upper middle value is a defensible definition, but every other figure we publish uses the average of the two, and a reader comparing two of our pages has no way to know the word meant different things on each. The gap between managers and players, which is the finding, is unchanged. We attached our leave-one-out check to both figures at the same time: removing any single person moves the players' median not at all and the managers' by 0.03 times its own size, so both hold.
This article was first published with a 48-month window and five measured titles, reporting a median peak of 2.2 times the surrounding months and a median floor afterwards of −6.7%. Re-fetching the same titles over 72 months moved six of the peaks, because a window that is too short makes its own highest month look like a peak. Squid Game's true peak is October 2021 at 2,440 per million, not January 2025 at 2,046; All of Us Are Dead peaked at 1,268 in February 2022, not 186 in July 2025; Vincenzo at 772 in June 2021, not 158 in April 2024. Three of the five titles in the original table were sitting on peaks that were not their peaks. The wider window also exposed the real limit, which the first version missed entirely: for a new title there is no floor before the wave at all, because the encyclopaedia article is created when the title arrives. The article was rewritten around that.
Second correction, same day. We widened the title set from 35 to 59 to answer a different question, and the answer to this one moved with it. The measured set went from six titles to nine, and the median change in floor after a wave went from −15.1% to −4.5% — from a fall worth reporting to something close to nothing. The median peak moved from 1.6× to 2.2×. Three of nine floors now end higher rather than two of six. Nothing was wrong with the earlier figures; they were the median of six titles, and six was too few for a median to hold still. The finding this article can now support is weaker and is stated as such.
As first published, this article explained the fall in travel articles by saying that questions with one settled answer — the visa rule, the transit card, which airport — have been draining out of encyclopaedias while pages you browse for pleasure have not. No figure in the article was wrong, but that sentence was not measured before it was written, which is the thing this newsroom exists not to do. We have since split the travel articles into those two halves and measured: the settled-answer half fell 33.5% and the browsing half 29.2%, a gap of 4.3 points against the 18.6 points that separate travel articles from culture articles. The kind that fell least of all was the nightlife district page at −10.7%, which is on the browsing side — the direction the explanation predicts, at about a quarter of the size it would need. The explanation is now reported as one we tried and could not keep, with the table behind it, and the article no longer offers any account of why the travel pages fell.
Published with Malta in the first table at a median breadth of 71. Malta is no longer in it: Croatia is, at the same 71. Thirteen titles were being counted as Korean that are not, and two of them charted in Malta. The test that found them is where a title charted. Undercover took 48 chart places without one of them in Korea, 16 of them in the Netherlands and 6 in Belgium — it is the Belgian series, and Wikidata knows no Korean work by that name. The other twelve are the same shape: a title that never reaches Korea's own chart, has no Korean work of that name, and peaks in exactly the country the foreign work comes from. Removing them takes 84 chart places out of 37,750, which is 0.22%, and moves three figures here: Japan's median breadth from 17.5 to 18, the under-5% band from 62 to 63, and the range inside that band from 49 to 48.5. The direction of the article is unchanged. Two further titles met two of the three tests but were kept, because a Korean work does share their name and we cannot tell which one charted: Keys to the Heart and Life Is Beautiful.
The screen panel this article measures was 235 titles and is now 234, and the shares move with it: the 17 titles that reach half the attention are 7.3% of the panel rather than 7.2%, and the ten largest hold 56.2% rather than 57.0%. The article was right about the numbers it was given; the rule that produced them was wrong. Our test for whether a Netflix title is Korean matched the title text against Wikidata labels in any language, so 23 titles written in Arabic, Hebrew, Cyrillic and Japanese script were counted as Korean works. They are local titles. Netflix does not translate titles in these lists — Squid Game holds 445 chart places across ten Arab markets under its Latin name, and only 10 of the 1,530 distinct titles on Egypt's chart are written in Arabic script. The rule now ignores non-Latin labels. The comparison this article makes is unchanged in direction: music is still the more concentrated of the two, and still more concentrated once Squid Game is removed from the screen side.
Recomputed again. Eight titles came out of the panel because our own attribution query returned no Korean work for those exact names — Waterworld, Re/Member, Into the Storm, Wildflower, Feng Shui, Glorious Days, Long Live Love! and You and Me. The catalogue falls from 405 to 397 and every figure and table here moved with it: the all-six share rises from 44.4% to 45.1%, the single-market share falls from 31.9% to 30.7%, and the middle rises from 23.7% to 24.2%. This article carried the old total for about forty minutes after the data changed because no automated check covered it. There is one now.
This article said twelve figures changed on 7 August and that all twelve came from one flaw. Both were wrong. Eleven changed that day — the twelfth was a correction made on 6 August to a different article, which we swept into the count. And three of the eleven had nothing to do with the title-matching flaw: two were separate errors this article's own text already described as unrelated, in a KOSIS table read on one level and a pay figure with the wrong denominator. Eight came from the flaw. The counts are now computed from the corrections record, where every entry carries its cause, and a check fails our build if the article and the record disagree. An article about miscounting that miscounted is the worst possible place for this, which is why it is at the top of the page rather than the bottom.
This piece is itself the correction. Our /staying-power page carried the heading 'Reaching number one doubles the hours. It adds no weeks.' Both halves came from a table computed on the fifty largest titles by hours — a group selected on the outcome being explained. Recomputed on all 235: the ratio is 4.3× on medians rather than 2.5× on means, and run length differs sharply rather than not at all. The page now reports the whole panel, shows medians beside means, and shows the banded comparison.
Published against a 405-title panel; the panel is now 397. Within an hour of this piece going out we found eight titles that our own attribution query said no Korean work carries — Waterworld, Re/Member, Into the Storm, Wildflower, Feng Shui, Glorious Days, Long Live Love! and You and Me — and removed them. Every figure in this article moved: films are 58.1% rather than 59.9%, series 23.6% rather than 23.9%, and the review queue is 20 one-country titles rather than 25. Wildflower was named here as a clear miss and is now gone from the panel rather than queued in it. The direction of the finding did not change, which is the least interesting thing about it: the piece argued our film numbers are the soft ones, and seven of the eight removals were films (167 down to 160) and one was a series (238 to 237).
Our /data page said of acts with no English Wikipedia article: 'they are invisible — and we cannot say how many there are.' That was true of the panel and false of what we could reach: Wikidata records the same acts and can be counted directly. The page now carries the figures below. Nothing in the attention panel changes; what changes is that the size of its blind spot is now stated instead of waved at.
The export figures in this article were wrong when it was published, and the error was in how the source table was read rather than in the arithmetic. KOSIS table DT_113_STBL_1025706 has a two-level classification — the upper level is the form of export (finished programmes, format sales, video and DVD, time blocks, support for overseas Korean broadcasting, other) and the lower level is the type of company. The rows it returns carry only the lower-level name, so five or six rows in each year are all labelled 'terrestrial broadcasters' and differ only by code. We had taken the finished-programme row alone for three of the four company types, while independent production companies, which the survey does not split by form, entered at their full value. Three of the four figures were therefore a fraction of the right number and were being compared against one that was not. Each company type is now the sum of every export form the survey publishes for it. Where the survey also publishes its own subtotal — eight cells across 2022 to 2024 — our sum matches it to within $1,000, and the four types now add to the published annual total exactly in 2024 and within $2,000 in every other year. Terrestrial broadcasters were 81.5% of exports in 2012, not 65.7%, and 7.2% in 2024, not 3.8%; their peak was $278.5m in 2016, not $212.1m in 2014. The article's finding is unchanged and slightly larger than stated: the terrestrial share fell further than we had published, and the 2024 figure is still about a third of the peak. The claim that a residual was left unattributed in the source was also wrong — nothing was missing from the source, we had simply not added it up.
The Netflix figures this article quotes were our own published numbers, and they were wrong. Korean titles had been matched by title text against a Wikidata list, which pulled in titles Netflix classes on its English-language charts — The Perfect Couple, Suits, Hit Man and others that are not Korean — along with same-name collisions inside the Non-English charts such as Teach You a Lesson (China) and Hunger (Thailand). The catalogue is 236 titles and 23.7 billion hours, not 294 and 27.7 billion, and Squid Game is 21.3% of it rather than 18.2%. The ten biggest titles are 41.3%, not 36.4%. The pageview side of this article is unaffected — it comes from Wikimedia and was re-derived unchanged — and the comparison the article makes is unchanged in direction and slightly wider: one title still holds a fifth of the viewing while the most looked-up performer holds 3.1% of the attention. The correlation of −0.16 is computed only from pageview and title-count fields and does not move.
Every count in this article was overstated when it was published. Korean titles were matched by title text against a Wikidata list, and English titles collide across countries, so titles that are not Korean were counted — among them The Perfect Couple and Run Away (United States), Hunger (Thailand), Forgotten Love (Poland) and two Witcher animations. Netflix splits its global charts by a title primary language, and we now use that classification to drop any title Netflix places on its English-language charts, plus a hand-checked list of same-name collisions inside the Non-English charts. The catalogue falls from 448 titles to 405. Every figure and table below has been recomputed. The shape of the finding did not change — it sharpened: the all-six share rises from 43.3% to 44.4%, the single-market share from 29.9% to 31.9%, and the thin middle falls from 26.8% to 23.7%. The country lists carry no language field of their own, so this correction is only as good as the global chart classification it borrows.
Every total in this article was overstated when it was published. Korean titles were matched by title text against a Wikidata list, and English titles collide across countries, so 52 titles Netflix classes on its English-language charts were counted as Korean — among them The Perfect Couple, Suits, The Circle and Hit Man, none of which are Korean. Netflix splits its charts by a title's primary language, so a Korean title belongs on the Non-English chart; we now count those only. Hand-checking the largest remaining titles removed seven more, including Teach You a Lesson (China), The Empress (Germany), Forgotten Love (Poland) and Hunger (Thailand). The catalogue falls from 294 titles to 236 and the total from 27.7bn hours to 23.7bn, a cut of 14%. Every figure and table below has been recomputed on the corrected set, which raises Squid Game's share from 18.2% to 21.3%. One example in the closing section, Run Away, was one of the wrongly-matched titles and has been replaced.
A verification note said the regional figures sum exactly to the published total in all 20 years. They do not: 2017, 2021 and 2022 are off by one thousand dollars, which is rounding in the published table and at most 0.0002% of the year. Every other year is exact. No figure in the article changes; the check was overstated and is now stated as measured.
We tagged every changed figure with the kind of mistake that produced it. One kind accounts for nine of the thirty-six. Four kinds still have no test, and this piece names them rather than reporting the coverage.
8 August 2026 · corrected
The weekly charts carry nine columns and none is country of production. Match titles by name, as everyone must, and 17% of Korean viewing sits on names another country also uses. We measured the gap and publish its size.
7 August 2026