Record Days

A page's normal level is the median of its daily views from 30 days before its record day to 22 days before it. The source file also offers the window 21 to 8 days before, and this site never uses it: for 45 of these pages traffic is already climbing by then. One definition of normal, used in every section.

214 of the 220 pages are drawn.6 are left out because their traffic was already high through the whole month before their record day, so there is no normal level to measure them against. Of the 214 drawn, 179 returned to normal, 32 were tracked for a full year afterwards and were still above twice their normal level at the end of it, and 3 had a record day too recent for the file to say either way. Those last 35 are the block at the bottom of the chart; their bars run past the axis and fade out.

The FIFA World Cup bar, at one day, is the least reliable number on the chart. A tournament runs longer than a month, so the window used for its normal level sits inside the tournament rather than before it. That makes its normal level too high and its bar too short. Its record day was 1,473,099 views. The Academy Awards and WrestleMania bars are short for the same reason.

These 220 pages were selected for having a huge day and then falling back from it within a week. So "traffic spikes and then falls" is how the set was chosen, not something this chart found. Nothing on this page rests on that: every number here is read either the day before the record day or more than 180 days after it, and the selection looks at neither.

On a screen narrower than about 640 pixels the shortest bars are drawn at a minimum width of 1.2 pixels, so one day and two days cannot be told apart there. Every other bar is drawn at its true length.

The day before means the one calendar day before the record day, not an average of several. Half of these pages got at least 40.9 times more views on the record day than they had the day before.

The upper field covers a single tenfold step of the scale, from 1,444,398 to 14,954,133 views. The lower field covers nearly six of them, from 5 views to 3,501,236. That difference in spread is what this chart is showing.

One page runs the other way. Antifa (United States) got 2,561,240 views the day before and 1,664,217 on the day. A page's record day here is the largest day the daily top-1000 ranking recorded for it, and for this one page its own daily series carries a larger day one day earlier.

A view is not a person. The source counts pageviews and does not identify readers, so one person loading a page ten times counts ten times.

Normal level is the same one used in every section: the median of a page's daily views from 30 to 22 days before its record day. The settled level is the median of days 300 to 340 after it. A page needs both windows to be drawn here, which is why 196 of the 220 are.

46 pages ended above twice their normal level and 14 above five times it.

5 of the 10 pages furthest to the left are scheduled events: a World Cup, two WrestleManias, a Royal Rumble and an Academy Awards. Those events run longer than a month, so the window used for their normal level sits inside the event rather than before it, and the level they are measured against is too high. Their squares sit further left than they should.

8 of these pages were renamed during the year measured here. Wikipedia keeps counting the old title, which after a move receives almost nothing, so those pages read as abandoned when what changed was the name. The two furthest left are both moves: J. D. Vance at 0.013 times its normal level and Charles, Prince of Wales at 0.022 times. They are drawn, because the move is real and dropping rows would change the count rather than correct it. Without all 8 of them, 85 of 188 pages end above their normal level, which is 45.2 percent rather than 43.9.

Squares are grouped into bins of equal width on this scale, and the number of bins follows the width of the chart, so the shape is coarser on a phone. The two counts either side of the line do not depend on the bins.

The day before is read as a share of the record day itself, so a page that got 1,000 views the day before a 1,000,000 view day counts as 0.1 percent. The quietest fifth were at 0.13% or less. The busiest fifth were at 29.7% or more.

The five groups do not fall in order. The busiest fifth settles at 0.58, slightly above the fourth group at 0.53. What holds across the five is narrower: the quietest fifth is the only group whose middle page ended above its own normal level, and the only group where most pages did (90 percent of it).

Cutting the same 196 pages at fixed thresholds instead of into fifths gives the same picture. Pages at under 2 percent of their record day the day before (n = 91) have a middle value of 1.3; pages at over 25 percent (n = 43) have 0.55. A Mann-Whitney test on those two sets gives z = 3.587, p = 3.4e-4.

The shapes are smoothed over 0.22 of a tenfold step, on one vertical scale shared by all five, so a group with more pages at one value stands taller. Nothing on this page rests on a bump in a shape. The medians, the counts and the test above are the numbers.

The source file keeps a page only if, a week after its record day, it is still visibly raised. One arm of that test is day seven divided by the median of days 21 to 8 before. For these 10 pages that comes out under 1, from FIFA World Cup at 0.14 to 98th Academy Awards at 0.99, because the window it is divided by sits inside the event rather than before it. Every one of the ten is a scheduled event: FIFA World Cup, WrestleMania 34, WrestleMania 33, Royal Rumble (2024), Tom Brady, 88th Academy Awards, 92nd Academy Awards, Peyton Manning, Beau Biden, 98th Academy Awards.

That is a fact about day seven, not about the record day. On the window this page uses, days 30 to 22 before, every one of those ten record days is between 2.67 and 122 times its own normal level.

Moving the window further from the event does not always help. For 6 of the ten the earlier window is no lower than the later one, and for 2 it is higher: the FIFA World Cup page was read 552,584 times a day 30 to 22 days before the final and 397,613 times a day 21 to 8 days before it, against 1,473,099 on the day. A tournament runs longer than a month, so there is no quiet reading anywhere in the 30 days this file carries.

The bottom ten are not a hand-picked opposite. They are the ten pages in the file with the least traffic the day before their record day, out of those that have a usable normal level that is not already climbing. Every one of the top ten sat above 1 percent of its record day through the later window and every one of the bottom ten sat below it, so the two blocks do not overlap on that reading.

This section is the reason every other section on this page uses days 30 to 22 before, and never the window the file offers.

Expected counts come from how many of each weekday the window holds, which is 31.4 to 31.5 of each across 4,062 days. A chi-square test on the seven counts gives 19.66 on 6 degrees of freedom, p = 0.00318.

Ordinary reading is measured on the same pages long after their event, on days 180 to 345. Each daily reading is divided by the median of the 7 days centred on it, so however fast a page is rising or falling, only the day of the week is left. That is 33,914 readings across 211 pages, and the seven values are Sun 1.080, Mon 1.041, Tue 0.988, Wed 0.970, Thu 0.960, Fri 0.954, Sat 1.008. Recomputing them over days 100 to 345 moves none of them by more than 0.01, and resampling whole pages 2,000 times puts the gap between the highest and the lowest weekday between 1.107 and 1.161.

The two weeks do not line up. Record days are highest on Monday and lowest on Tuesday; ordinary reading is highest on Sunday and lowest on Friday. Monday is a slightly busy reading day, by 4 percent, against a 66 percent excess in record days. Tuesday, the emptiest record day here, is an ordinary reading day at 0.9876.

Dates are UTC, as the source reports them, so an event that happens on a Sunday evening in the United States is recorded on the Monday. Nothing here explains why the counts fall this way, and nothing on this page claims a cause.

Dates are lumpy on their own, and this file already knows its weekdays are uneven. So the count is tested against a model that holds each record day's year, month and weekday fixed and redraws only which matching day of that month it fell on. Over 20,000 draws that model puts 25.4 pages in a group on average, and reached 47 or more 3 times. p = 2.0e-4.

The 19 dates are 12 pairs, 5 threes and 2 fours:

• 29 Feb 2016Leonardo DiCaprio · 88th Academy Awards
• 21 Apr 2016Prince (musician) · Chyna
• 9 Nov 2016Donald Trump · United States presidential election, 2016 · Electoral College (United States) · Melania Trump
• 16 Aug 2018Aretha Franklin · Atal Bihari Vajpayee
• 3 Feb 2020Shakira · Jennifer Lopez
• 4 Nov 2020United States Electoral College · 2016 United States presidential election
• 8 Nov 2020Kamala Harris · Joe Biden · Jill Biden · Beau Biden
• 8 Feb 2021Tom Brady · The Weeknd
• 9 Apr 2021Prince Philip, Duke of Edinburgh · DMX (rapper)
• 24 Feb 2022Ukraine · 2022 Russian invasion of Ukraine · Vladimir Putin
• 28 Mar 2022Jada Pinkett Smith · Will Smith
• 8 Sep 2022Elizabeth II · Operation London Bridge · Charles, Prince of Wales
• 18 Dec 2022Lionel Messi · Kylian Mbappé · FIFA World Cup
• 13 Jan 2023Lisa Marie Presley · Priscilla Presley · Riley Keough
• 13 Mar 2023Everything Everywhere All at Once · 95th Academy Awards
• 6 Nov 20242024 United States presidential election · 2020 United States presidential election
• 16 Nov 2024Jake Paul · Mike Tyson
• 3 Mar 2025Anora · Mikey Madison · 97th Academy Awards
• 20 Feb 2026Eric Dane · Alysa Liu

The vertical scale runs from 1,444,398 to 14,954,133 views, which are the same smallest and largest record day the second section prints. Every step up it is a fixed multiple, not a fixed number of views.

A page is in this file only if its own biggest day reached the daily top 1,000 for that day, so two pages sharing a date here means both were among the most-read pages on that date, not that nothing else happened.

A page's level here is not the normal level used in the other sections. It is the median of days 330 to 400 after the record day with the fortnight either side of day 365 taken out, so the day being measured cannot lift the level it is measured against. One page's reading is that day's views divided by that level, and every number in this section, including every control, comes out of that one calculation.

46.2 percent of the pages read at least twice their own level on the day. Running the identical calculation at 200 days (0.9648 times, 1.9 percent), 240 days (1.0129 times, 6.3 percent), 270 days (0.9648 times, 0.5 percent), 300 days (0.9984 times, 7.4 percent), 330 days (0.9671 times, 7.5 percent) finds nothing of the kind. A sign test on 177 of 199 gives p = 1.4e-31, and resampling the pages 20,000 times puts the middle value between 1.6 and 2.085.

13 of these pages are titled for something that happens again a year later, an election, a final, a ceremony, so their day 365 may be the next one rather than a memory of the last. They read 2.0532 times. The other 186 read 1.8409 times, with 89.2 percent above their own level, so the finding does not rest on them. Dropping the four pages of the shape described in no section here and the eight renamed pages leaves 187 pages at 1.8527 times.

199 of the 220 pages can be measured: the rest have a record day too recent for the file to hold a year after it. 196 of the 199 have every day of the window. The largest reading is Diogo Jota at 82 times its own level, the smallest Pete Hegseth at 0.42 times. Traffic is still falling at day 365 for some of these pages, which pulls the surrounding level up and the reading down, so this is the conservative direction.

For each page, the traffic above its normal level is fitted twice over the same days: once against the day number, which is an exponential decay, and once against the logarithm of the day number, which is a power law. Only days above the page's own normal level can be used, and a page needs 10 of them inside a window to be fitted there. 182 pages clear that at both 60 and 340 days, and those same pages are what every point on this chart is measured on.

The reason the answer moves is that the exponential is the wrong shape. Over 60 days the power law fits better on 139 of the 182 pages (76.4 percent, median R² 0.835 against 0.713); over 340 days, on 167 (91.8 percent, 0.492 against 0.223). A curve that is steep early and shallow late has no single halving time, so the number a fit reports is set by where the window stops.

The pages of this file were selected partly on the shape of their fall, which if anything favours the exponential. So the same test was run on a second, separate set of 88 events pulled by hand with no shape test of any kind: the power law fits better on 66 of them, 75 percent, median R² 0.802 against 0.706. That set counts all traffic rather than human readers only, so its numbers are not mixed with anything else on this page.

The flat band is read without any model: the first day a page's traffic above normal falls to half, a tenth and a twentieth of the record day's, which is day 1, day 3, day 5 for the middle page, and the half-life each of those implies — 1 days, 0.9 days, 1.16 days. Something that really decayed exponentially would give the same answer at all three depths, and this does. The fitted curve is above that band at every window on the chart, by a factor of 84 at the widest.

On 22 May 2017, 22 pages each had the biggest traffic day in their history. Their record days run from 1,054,667 to 1,066,589 views — a largest over smallest of 1.0113, a coefficient of variation of 0.26 per cent, and 23,300,230 views between them. Every one of the 22 is a country or a historical state, and every one was thrown out by the shape test:

Norway · The Bahamas · Russian Empire · Prussia · Grenada · Korea · Ancient Rome · Kingdom of Poland (1385–1569) · Saint Kitts and Nevis · Niue · Kingdom of Prussia · Moldavian Democratic Republic · Second Hungarian Republic · People's Republic of Benin · Restoration (Spain) · Kingdom of Laos · Saint Lucia · Malayan Union · United Kingdom of Portugal, Brazil and the Algarves · Couto Misto · Korean Empire · Serbian Despotate

How many. Dates are lumpy on their own, so the count is tested against the same model the seventh section uses: each page keeps its year, its month and its weekday, and only which matching day of that month it fell on is redrawn. Over 20,000 draws that model puts 96.8 pages on a shared date against the 139 observed, and its biggest single date holds 6.95 pages on average and 13 at its most extreme. It never reached 22. p = 5.0e-5 for both.

How close. Sharing a date is a coincidence; agreeing to a hundredth is not. Draw 22 pages at random from the same 442 and their record days spread a median 6.7402 times; the tightest of 20,000 draws is 2.0204, and none reached 1.0113. The file's peaks also crowd towards the bottom of the ranking, so the same question was asked without any null at all: the tightest 22 record days that can be assembled from anywhere else in the 442 spread 1.0272 times, and they sit on 22 different dates.

The research record for this project claimed every other multi-page date spreads at least 1.14 times, median 2.24. Both numbers were wrong and are corrected here. 1.141 is the tightest other date carrying three or more pages (25 May 2022); three pairs are tighter than 22 May 2017 is, which is what pairs do when there are 33 of them. The median over the other 48 dates is 1.5549, and the 19 shared dates the seventh section draws spread a median 1.5918.

A second reading agrees and does not look at size at all. The shape test records each page's traffic on day seven against day three. 43 of the 442 pages are at 0.15 of day three or less by day seven, and 22 of those 43 are this one date. So the 22 pages are not only the same size, they are the same shape.

The other date whose pages agree is 18 Jan 2018: JSON Web Token, HTTP cookie, Access token and Session token, within 50 per cent of each other, all four thrown out.

The limits. 29 pages anywhere in the 442 fall inside the value band these 22 occupy, so 7 pages on other dates are that size too; the band is busy, not empty. A thrown-out page has a peak, a date and three test readings in the source file and no traffic history at all, so no page drawn in pink here can appear on any chart on this page whose axis is days. And what produced this is not identifiable from this extract: a crawler, a mirror and a redirect all leave the same marks.

All 49 dates, with the number of pages on each and how far apart they are:

• 24 Nov 20152 pages, within 18% · 2 thrown out
• 11 Jan 20162 pages, within 562% · 1 thrown out
• 29 Feb 20163 pages, within 41% · 1 thrown out
• 21 Apr 20162 pages, within 292%
• 9 Nov 20164 pages, within 255%
• 6 Dec 20162 pages, within 71% · 1 thrown out
• 15 Jan 20172 pages, within 8% · 2 thrown out
• 17 Mar 20172 pages, within 52% · 2 thrown out
• 22 May 201722 pages, within 1% · 22 thrown out
• 23 May 20172 pages, within 57% · 1 thrown out
• 2 Nov 20172 pages, within 0% · 2 thrown out
• 18 Jan 20184 pages, within 50% · 4 thrown out
• 19 May 20183 pages, within 312% · 2 thrown out
• 16 Aug 20182 pages, within 7%
• 12 Nov 20182 pages, within 53% · 1 thrown out
• 15 Apr 20192 pages, within 152% · 1 thrown out
• 31 Aug 20192 pages, within 26% · 2 thrown out
• 19 Dec 20192 pages, within 18% · 2 thrown out
• 27 Jan 20202 pages, within 17% · 1 thrown out
• 3 Feb 20204 pages, within 124% · 2 thrown out
• 10 Feb 20202 pages, within 66% · 1 thrown out
• 1 Jun 20202 pages, within 18% · 1 thrown out
• 4 Nov 20202 pages, within 3%
• 8 Nov 20205 pages, within 343% · 1 thrown out
• 8 Feb 20212 pages, within 59%
• 9 Apr 20212 pages, within 27%
• 24 Feb 20223 pages, within 58%
• 28 Mar 20222 pages, within 20%
• 25 May 20223 pages, within 14% · 3 thrown out
• 10 Jun 20222 pages, within 22% · 2 thrown out
• 7 Jul 20222 pages, within 64% · 1 thrown out
• 8 Sep 20224 pages, within 790% · 1 thrown out
• 6 Nov 20222 pages, within 87% · 1 thrown out
• 18 Dec 20223 pages, within 110%
• 13 Jan 20233 pages, within 174%
• 13 Mar 20232 pages, within 28%
• 21 May 20232 pages, within 96% · 2 thrown out
• 28 Jan 20242 pages, within 46% · 1 thrown out
• 4 Jun 20242 pages, within 29% · 1 thrown out
• 15 Jul 20242 pages, within 324% · 1 thrown out
• 6 Nov 20242 pages, within 8%
• 13 Nov 20242 pages, within 186% · 1 thrown out
• 16 Nov 20243 pages, within 15% · 1 thrown out
• 3 Mar 20253 pages, within 81%
• 8 May 20252 pages, within 54% · 1 thrown out
• 10 Sep 20253 pages, within 1006% · 2 thrown out
• 20 Feb 20262 pages, within 82%
• 1 Mar 20263 pages, within 322% · 2 thrown out
• 14 Jun 20262 pages, within 24% · 1 thrown out

The snap. Take the last day a page was still being read at event scale — above 20 times its own normal level and at least 10,000 views — and divide it by the day after. Attention that is real is still coming down, so that ratio is small: over 207 pages the middle value is 1.78, three quarters are under 2.54 and nine tenths under 6.09. Three pages are at 817.9, 204.4 and 81.8, and the fourth largest anywhere in the file is Charles, Prince of Wales at 11.5.

The dropout. A separate reading that never looks at the end of an event: a single day at least 10 times below both the day before it and the day after it, where both of those are above 5,000 views. An audience cannot vanish and return overnight. Exactly 3 pages in the file have such a day — Dulce María (4), Question mark (2), Cook's Country (1) — and nobody else has one.

The two readings name 4 pages between them and agree on 2. Index (statistics) sat at 44 views a day, held 20 straight days at event scale running 128,405 to 2,372,030, and read 157 the next day. Cook's Country sat at 97 views a day, read 57 two days before its record day and 833,826 one day before it, and 876 on day 15. Question mark read 522,905 on day one, 1,142 on day two and 638,607 on day three.

Dulce María is named by the dropout reading and not by the snap, and the reason is worth stating: her traffic is still above 20 times her own normal level on day 400, the last day this file covers, so there is no day after to divide by. She is one of 4 pages in that position; 3 more never reach event scale at all, which is why the snap is read on 207 pages rather than 214. Her four dropout days are 353, 377, 394, 397, past the right edge of this chart. What the chart points at instead is her steepest single day inside it: day 47, from 33,034 views to 1,529.

The same reading applied to the other three points at the same day the snap does for two of them and not for Question mark, whose steepest day inside this window is its dropout on day 2 — 457.9 times in one day, larger than the 81.8 it falls when it finally stops on day 48. The two readings disagree about which of its days matters most, which is why both are kept.

These four are in every other chart on this page and move none of them. The middle return time is 28 days with them and 28 without. The share of pages settling above their own normal level is 43.9 per cent with them and 43.8 without.

What caused this is not identifiable from this extract and nothing here names a cause. An automated crawl, a mirror and a redirect pointed at a title all draw the same rectangle in a pageview file, and nothing in the census separates them. What is reported is the shape, which two unrelated readings agree on and which no readership in the file has.

214 of the 220 pages are drawn. The other 6 have no quiet month before their record day to measure a normal level against, which is the same six the first chart leaves out. Two pages have days the source never returned and are drawn as broken lines. A page reading below its own normal level is drawn on the bottom line rather than under it, so the floor of this chart is a floor and not a reading.