Korean entertainment writing sorts groups into generations — first, second, third, and lately a sixth. The boundaries are quoted as if they were established, but no institution sets them and no database records them. They are a convention. So we did the one thing a convention allows: we took 182 Korean acts whose founding year we could verify, put them against how much they are read in English, and asked whether anything actually changes at the lines.
Two things came out. Median reading does climb with each generation — that part holds. But the generations overlap by 86% to 100%, so knowing an act's generation tells you almost nothing about that act. And the years where reading genuinely breaks are not the years the labels put the break.
| Generation | Acts | Median reads | Middle half runs | Median member share |
|---|---|---|---|---|
| Before the labels (pre-1996) | 9 | 35,481 | 33,456 – 39,768 | 78.8% |
| First (1996-2002) | 28 | 63,586 | 16,492 – 213,131 | 72.8% |
| Second (2003-2011) | 43 | 116,555 | 44,761 – 555,719 | 80.0% |
| Third (2012-2017) | 69 | 114,924 | 22,302 – 246,896 | 72.2% |
| Fourth (2018-2022) | 28 | 124,132 | 39,586 – 698,757 | 46.9% |
| Fifth (2023-2025) | 5 | 263,840 | 142,483 – 354,299 | 2.5% |
Reads are the twelve months 2025-09 ~ 2026-08, English Wikipedia, act page plus member pages added together. Median, not average — one act can be read a hundred times more than its neighbour, and an average would hand that act the whole generation. Suspected sub-units are held out of this table; 126 of the 308 acts we could date are sub-units, and they read like people rather than like groups because the sub-unit's own name is the part nobody knows.
A median is a single point. What matters for any one act is the spread, so here is how much the middle half of each generation overlaps the middle half of the next.
| Pair | Acts | Middle halves overlap |
|---|---|---|
| Before the labels → 1st | 9 · 28 | 100% |
| 1st → 2nd | 28 · 43 | 86% |
| 2nd → 3rd | 43 · 69 | 90% |
| 3rd → 4th | 69 · 28 | 92% |
| 4th → 5th | 28 · 5 | 100% |
| 5th → 6th | 5 · 0 | not measured |
Overlap is the shared part of the two middle-half ranges, divided by the narrower of the two. 100% means one generation's middle half sits entirely inside the other's. At these numbers the label is a description of when an act started, not a description of how much it is read.
This is the test the labels have to pass. Line the acts up by founding year, and for each pair of consecutive years compare the median reading. If the generation boundaries are real, the biggest jumps should sit on them.
Of the 16 year-to-year gaps we could measure, 2 are generation boundaries. Among the 2 largest jumps, 0 are a boundary — chance alone would have given 0.25.
| Year begins | Median moves | Change | Acts either side | On a boundary? |
|---|---|---|---|---|
| 2011 | 518,907 → 57,247 | 9.064× | 5 · 8 | — |
| 2019 | 369,279 → 43,059 | 8.576× | 7 · 9 | — |
| 2001 | 25,376 → 130,973 | 5.161× | 4 · 6 | — |
| 2007 | 96,357 → 446,433 | 4.633× | 5 · 4 | — |
| 2018 | 109,879 → 369,279 | 3.361× | 8 · 7 | yes |
| 2009 | 629,181 → 241,224 | 2.608× | 5 · 11 | — |
| 2015 | 80,075 → 208,217 | 2.6× | 10 · 11 | — |
| 2013 | 32,694 → 75,834 | 2.32× | 18 · 8 | — |
The boundaries we could measure land at position 5 of 16 (2018, 3.361×) and 11 of 16 (2012, 1.751×) in that ranking. The two largest breaks are elsewhere: 2011 at 9.064× and 2019 at 8.576×. Both are a single year off a boundary, which is worth saying plainly — it is not evidence that nothing changed, it is evidence that the line is drawn a year from where the reading moves.
A gap is only counted when both years hold at least four acts, because a year with two acts in it moves on one act. 16 of the year pairs cleared that; the rest are left out rather than filled in.
Our set contains no acts in Sixth (2026-). That is not a claim that no such groups exist — Korean outlets are naming them right now. It is a statement about what is measurable: an act founded in 2026 has had no twelve-month window to be read in, and English Wikipedia has usually not written it down yet. The label arrives first, the record arrives later, and for a while there is nothing to count.
In the first table the fifth generation's median member share is 2.5%, far below every older generation. Read carelessly that says fifth-generation fans care about the group and not the people. It says nothing of the kind.
Our first explanation for it was wrong, and the correction is worth printing. We assumed the members had no English articles yet. They mostly do — the fifth-generation acts carry a median of 4 members with an article, against 2 for the third generation. The articles exist. They are barely read.
Here is the whole column, all 5 acts, so nothing hides behind a median.
| Act | Founded | Members with an article | Member share |
|---|---|---|---|
| KiiiKiii | 2025 | 1 | 0.0% |
| TWS (group) | 2024 | 1 | 0.2% |
| Hearts2Hearts | 2025 | 4 | 2.5% |
| Illit | 2024 | 4 | 3.2% |
| Kiss of Life (group) | 2023 | 4 | 47.0% |
4 of the 5 send under 5% of their reading to members, and two have only one member with an article at all. Whether that is because the encyclopaedia has not caught up, or because attention on very new groups genuinely sits on the group name, we cannot separate — not with five acts and one twelve-month window. So we do not claim either. What we will not do is print 2.5% as a fact about fans of new groups.
It is not a ranking of generations. The bins are a press convention, not a fact, and this file exists to test them.
It cannot say a generation label is wrong. It can say whether the years where reading breaks are the years the label puts the break.
Founding year is not always debut year. Wikidata records inception; some acts trained for years before releasing.
The boundaries in the first column are the ones Korean and English coverage tends to use. We wrote them into the collector so they can be read and argued with rather than inferred from our tables. If you would draw them elsewhere, the year-by-year gaps above are the part that does not depend on our lines.
We measured a different thing on nearly the same acts and got the same shape: how long a group stays looked up after its peak does not vary by debut era either — the median fall from a group's best month to half of it is about two months whether the group started in 1995 or 2020. Two separate measurements, both finding that the era an act belongs to explains less than the label implies. And for where a group's reading lands once you have it, the group page or the member pages is the split this page reuses.
The generation labels are quoted like facts, but nobody sets them. We put 182 Korean acts against English Wikipedia reading: the bands overlap 86% to 100%, and in both boundary years we could measure, the real break sits one year away.
8 September 2026
Founding years — Wikidata SPARQL, property P571 inception. Where an act carries several, the earliest is used. Of 380 acts, 308 had a usable year and 72 did not; the 72 are left out of every table rather than placed in an "unknown" row. Founding year is not always debut year — some acts trained for years before releasing — and we do not have release dates for all of them.
Reads — Wikimedia Pageviews REST API, English Wikipedia, all-access, agent=user (bots excluded), monthly, 2025-09 ~ 2026-08. The current month is excluded because it is not finished. Reading was not re-collected for this page; it is the same measurement as /member-vs-group, joined by Wikidata identifier. Counted 2026-09-08 10:58 KST.