Correction
What changed on 29 October 2026
Corrected 29 October 2026, after an independent audit. Three 2022 Nobel laureates in economics were missing because the Nobel Prize API gives the wrong year for that prize; they are now in. The curve comparison published on 28 October fitted only days above the baseline, which made a biexponential win almost by default (it reported 90%); every day now counts, and the selector was checked on simulated data. The half-life compared a smoothed series with a single raw day; it now compares like with like. 65 deaths reported days late are counted as delayed spikes rather than no spike. The Oscar baseline fell after the nominations; it now ends a week before them. Earlier figures are kept as sensitivity rows. The full list is in docs/decisions.md, entries 26 to 33.
The question
When someone becomes news, how fast does the world stop looking them up?
A death or a prize sends a burst of readers to a Wikipedia article. This study measures that burst for every person with an English article and at least 30 Wikidata sitelinks who died between January 2016 and October 2025 (1,905 after exclusions), and for the winners of six Nobel Prizes, five Academy Awards, the Ballon d'Or and the Booker Prize in the same years (178). For each event it reads daily views from 60 days before to a year after, from Wikimedia's pageview API, counting user agents only.
The event list was fixed and committed before a single pageview was fetched, so nothing about the results could shape who is in it. 21 deaths and 5 awards drop out because the article was written less than 30 days before the event and has no baseline; among them are George Floyd and Mahsa Amini, whose deaths made them known. 65 more drop because the article was renamed inside the window, which splits its views across two titles.
Finding 1
Half the spike is gone in about two days
On the median death the page draws 206 times its usual traffic on the peak day, which is the day after the death in 44% of cases. Smoothed over a week, half the extra traffic is gone 2 days after the peak (middle half 2 to 3 days). The averaged curve loses nine tenths of its excess in 5 days and sits at 1.96 times baseline a month later and 1.09 times in the last month of the year, before jumping to 1.58 times on the anniversary. Awards lose nine tenths in 4 days, and climb again in the last week of the year as the next prize comes round. 1,839 of 1,905 deaths have a visible spike (peak at least twice the baseline) on days 0 to 3. Another 65 (3.4%) have a delayed spike: nothing on days 0 to 3, then a day above twice the baseline on days 4 to 30 (Leonard Cohen, Sam Shepard and Steve Ditko among them), most often because the death became news days or weeks after it happened. 1 death has no spike at all.
Exhibit 1
Median daily views as a multiple of each event's baseline
deathsawardsbaseline
scripts/analyze.py.
Finding 2
A fast fade and a slow tail fit most events
Each event's daily views after the peak were fitted three ways, with every day counted: the baseline plus an exponential decay, plus a power law, or plus a biexponential (two exponentials, the form Candia et al. 2019 found over years in attention to songs, films and biographies). The biexponential has the lowest AIC for 76% of 1,839 deaths and 64% of 177 awards; the power law wins 14% of deaths and the exponential 9%. Where the biexponential wins, its fast part has a median time constant of 3.0 days and carries all but 1.8% of the initial excess; the slow part's median time constant is 169 days. On simulated series the selector names the true model 87% of the time for a pure exponential, 95% for a power law and 100% for a biexponential, and it calls 13% of pure exponentials biexponential, so the real share is a mild overstatement. The model comparison first published on 28 October dropped every day at or below the baseline before fitting; on the same simulations that method called 100% of pure exponentials biexponential, which is why it reported 90% biexponential wins.
Exhibit 2
Which model fits each event best
biexponentialpower lawexponential
| True model | Noise 0.15 | Noise 0.30 | First method, 0.15 | First method, 0.30 |
|---|---|---|---|---|
| Exponential | 87% | 82% | 0% | 0% |
| Power law | 95% | 88% | 82% | 73% |
| Biexponential | 100% | 99% | 95% | 91% |
How often the selector names the true model on 150 simulated series per row (scripts/simulate_selection.py). "First method" is the comparison published on 28 October, which fitted only days above the baseline.
Finding 3
A death leaves a page a little busier than before
A year after a death the median page draws 1.08 times its old traffic, and 59% of pages sit above their baseline. Award winners settle at a median 4.20 times, but the prizes differ: Nobel laureates, with a median of 20 views a day beforehand, keep 8.81 times that; Booker winners 2.56; Ballon d'Or winners 0.87. For Academy Award winners the answer depends on which 'before' you use: against the 53 days ending a week before the nominations it is 0.58, against days -150 to -90 before the ceremony 1.03.
| Event | n | Baseline views/day | Peak | Half-life, days | After week one | Lift |
|---|---|---|---|---|---|---|
| Death | 1,905 | 202 | 206x | 2 | 29% | 1.08x |
| Nobel Prize | 110 | 20 | 3012x | 1 | 46% | 8.81x |
| Academy Award | 50 | 9,378 | 40x | 2 | 23% | 0.58x |
| Ballon d'Or | 9 | 20,300 | 4x | 2 | 85% | 0.87x |
| Booker Prize | 9 | 76 | 222x | 3 | 69% | 2.56x |
Exhibit 3
Views a year later as a multiple of the baseline
deathsawards
A regression of log lift on event type, log baseline, occupation group and year (OLS, HC3 errors, n = 2,083, R² = 0.28) puts awards 194% above deaths at the same prior traffic (p < 0.001); with the Oscar baseline as the spec defined it the figure is 167% above (p < 0.001). Among deaths, with age added (n = 1,902, R² = 0.33), each extra year of age at death goes with 2.0% less lift (p < 0.001). The same regression for log half-life explains little (R² = 0.09).
Finding 4
Which prize matters more than what the person did
Among deaths, the share of the first year's extra views that arrive after the first week runs from 23% (business) to 33% (writing and journalism), and median lift from 0.97 times (business) to 1.22 times (music). The gap between Nobel and Oscar winners is wider than any gap between occupations.
Exhibit 4
By event type and, for deaths, by occupation
Checks
What moves the numbers, and what does not
Bot leakage: capping any day above 50 times the median of its six neighbors touches 31 events and leaves every half-life unchanged; median lift with the cap is 1.08 for deaths and 4.20 for awards. Renames: 65 events were dropped because the article moved to a new title inside the window, and 26 articles moved later are read under the title they had at the time. Peak window: searching for the peak over days 0 to 30 instead of 0 to 3 moves it later for 163 deaths and leaves the median half-life at 2 days. Smoothing and definition: the table below gives the half-life every way it was measured. Day of week: the median page's busiest weekday draws 1.42 times its quietest, which is why every series is adjusted before measuring.
| Half-life measure | Deaths, median days | Deaths, one day | Awards, median days | Awards, one day |
|---|---|---|---|---|
| Half-life, smoothed (7 days) against smoothed peak (main) | 2 | 11% | 2 | 33% |
| Half-life, smoothed (3 days) against smoothed peak | 2 | 33% | 1 | 62% |
| Half-life, smoothed (14 days) against smoothed peak | 6 | 0% | 6 | 0% |
| Half-life, daily views against daily peak | 1 | 58% | 1 | 77% |
| Half-life, as in the spec (smoothed against daily peak) | 1 | 95% | 1 | 99% |
| Half-life, peak searched over days 0 to 30 | 2 | 21% | 2 | 39% |
Limits
What this cannot show
English Wikipedia only, and a pageview is someone looking a person up, which is not the same as caring about them. The 30-sitelink floor decides who counts as famous and leans toward people known in many countries; it is today's count, so a few people clear it because of the death itself. Some automated traffic gets past the user filter. A death or prize can coincide with other news about the same person, a film release or a scandal, and nothing here separates the two. Awards season raises an Oscar winner's traffic for months, so any "before" window is a choice: the Oscar lift is 0.58 against the 53 days ending a week before the nominations, 0.38 against the spec's window after them, and 1.03 against days -150 to -90 before the ceremony. The award sample is small (178 events, 9 of them Ballon d'Or and 9 Booker), so differences between prizes are descriptive.
Method
How it was built
- Events
- Wikidata SPARQL for deaths (run per half-year) and Academy Awards; the Nobel Prize API for laureates and announcement dates; each year's English Wikipedia article for Ballon d'Or and Booker dates. Queries in
queries/, list indata/events.csv. - Pageviews
- Wikimedia REST API,
per-article/en.wikipedia/all-access/user/daily, days -60 to +365, titles resolved through Wikidata at fetch time; page moves inside the window found from revision summaries. - Measures
- Baseline is the median of days -60 to -8 (for the Academy Awards, of the 53 days ending a week before the nominations), peak the maximum of days 0 to +3, half-life the days from the smoothed peak until the centered 7-day mean of excess views falls below half of it, lift the median of days +335 to +365 over baseline. Day-of-week factors from the baseline window.
- Models
- Baseline plus an exponential, power-law or biexponential decay, fitted to log(1 + daily views) on every day after the peak; AIC from the residual sum of squares; the selector checked on simulated series. The browser runs the same code:
assets/fit.jsis a port ofscripts/fits.py, checked against it byscripts/check_fit_parity.py. - Regressions
- OLS with HC3 errors of log half-life and log lift on event type, log baseline, occupation group and year; for deaths, also age at death. Full table in
tables/regressions.csv. - Prior work
- Candia, Jara-Figueroa, Rodriguez-Sickert, Barabási and Hidalgo (2019), The universal decay of collective memory and attention, Nature Human Behaviour 3, 82–91.
- Built with
- Python (pandas, NumPy, statsmodels, matplotlib); plain JavaScript and inline SVG for the charts. Every number on this page is written from the analysis outputs by
scripts/export_numbers.py.