In short
Of 2,165 Google Drive links in research papers we checked, 620 no longer open (28.6%). Pew found 38% of web pages from 2013 gone a decade later, and half the links in US Supreme Court opinions no longer show what the court cited. Persistent identifiers such as DOIs held up best in every study here that compared them.
Key figures
Two kinds of decay
The studies below measure two different things, and each figure says which. Link rot is a link that no longer resolves, so the reader gets an error, a missing page or a sign-in wall. Content drift, which some authors call reference rot, is a link that still loads but shows something other than what the author pointed to. A few studies count both together, and those are marked.
The link returns an error, a missing page or a sign-in wall.
The link still loads but shows something different.
The study counted the two together.
A plain label names what a study measured when it was neither, such as archive copies of dead links.
Google Drive links in research papers
Our own study, published 13 October 2026. We requested 2,207 distinct Drive files and folders cited in 2,107 open-access papers in PubMed Central, once each, on 13 October 2026. The full method, data and limits are in our Google Drive link study.
- 620 of the 2,165 links that could be judged no longer open, or 28.6% (95% range 26.8% to 30.6%).
- 403 of them (18.6% of all links) were deleted, and Drive answered "Page not found". The other 217 (10.0%) sent the reader to a sign-in or access-request page.
- The failure rate climbs with the age of the citing paper, as the chart below shows.
- Drive's older link format, with IDs starting 0B, met a sign-in wall in 51 of 110 cases (46.4%). Current-format links met one in 166 of 2,024 cases (8.2%).
- 16 links failed as printed because typesetting had turned a hyphen in the Drive ID into a dash. With a plain hyphen restored, 9 of the 16 opened.
- As a benchmark, 247 of 249 Zenodo links and DOIs in the same papers still resolve (99.2%), including all 166 Zenodo DOIs. The two failures are on Zenodo's test site.
Show these figures as a table
| Published | Links checked | No longer open | 95% range |
|---|---|---|---|
| 2013-2018 | 102 | 78 (76.5%) | 67.4% to 83.6% |
| 2019-2020 | 181 | 72 (39.8%) | 32.9% to 47.1% |
| 2021-2022 | 617 | 213 (34.5%) | 30.9% to 38.4% |
| 2023-2024 | 672 | 171 (25.4%) | 22.3% to 28.9% |
| 2025-2026 | 593 | 86 (14.5%) | 11.9% to 17.6% |
| All links | 2,165 | 620 (28.6%) | 26.8% to 30.6% |
The web at large
8 figures
25% of pages no longer accessible
A quarter of the pages Pew sampled from 2013 to 2023 could not be reached in October 2023. In 16% of cases the page had gone from a site that still worked, and in 9% the whole domain had stopped working.
#pew-pages-2013-2023 (link to this figure)38% of 2013 pages not available in 2023
Older pages were far more likely to be gone. 38% of pages that existed in 2013 could not be reached a decade later, against 8% of pages from 2023.
#pew-2013-pages-gone (link to this figure)1 in 5 2021 pages no longer accessible
About one in five pages from Pew's 2021 sample could not be reached two years later.
#pew-2021-pages-gone (link to this figure)35.3% still alive in 2023
Of 27.3 million URLs the Internet Archive first saved between 1996 and 2021, 35.3% still returned a working page in 2023. Of the rest, 25.1% returned a 4xx error and 39.6% failed on DNS, timed out or hit a server error. The group reports these as preliminary results.
#odu-archived-urls-alive (link to this figure)Over 50% of deep links died within a year
In the same set, more than half of deep links stopped working within a year of first capture, and 4% lasted longer than ten years. Home pages fared better, with 10% gone within a year and 20% still alive after more than 20 years.
#odu-deep-links-first-year (link to this figure)66.5% of links to the sampled sites rotted
Ahrefs found that at least 66.5% of the links pointing to 2,062,173 websites since January 2013 had rotted, and 74.5% were lost once temporary errors are counted. This measures links that vanished from the linking page, so it describes lost backlinks rather than dead destinations.
#ahrefs-backlinks-rotted (link to this figure)47.7% of lost links were on dropped pages
Of those lost links, 47.7% sat on pages that had dropped out of Ahrefs' index and 34.2% had been taken off pages that still existed. Redirects accounted for 5.99% and "not found" errors for 4.11%.
#ahrefs-why-links-lost (link to this figure)18% of tweets no longer publicly visible
Of nearly 5 million tweets posted in March and April 2023, 18% were no longer publicly visible by mid-June 2023.
#pew-tweets-gone (link to this figure)Wikipedia and government pages
4 figures
11% of reference links no longer accessible
Of just over 1 million reference links on 50,000 English Wikipedia pages, 11% led to pages that could no longer be reached.
#pew-wikipedia-references (link to this figure)54% of pages with a dead reference link
54% of those Wikipedia pages had at least one dead reference link.
#pew-wikipedia-pages (link to this figure)21% of pages with a broken link
On about 500,000 US government web pages, 21% carried at least one broken link. Every level of government was at 14% or above, and city pages fared worst.
#pew-government-pages (link to this figure)6% of links no longer accessible
Followed one by one, 6% of the links on those government pages led to pages that no longer existed, with internal and external links failing at similar rates.
#pew-government-links (link to this figure)News
8 figures
5% of links no longer accessible
On 500,000 pages from 2,063 news sites, 5% of links to other sites no longer worked.
#pew-news-links (link to this figure)23% of pages with a broken link
23% of those news pages carried at least one broken link.
#pew-news-pages (link to this figure)25% and 26% busiest and quietest fifth of sites
Size made little difference. 25% of pages on the busiest fifth of news sites had a broken link, against 26% on the quietest fifth.
#pew-news-site-size (link to this figure)25% of deep links completely inaccessible
A quarter of deep links in New York Times articles from 1996 to mid-2019 were completely inaccessible.
#nyt-deep-links (link to this figure)72% of deep links from 1998 had rotted
Older Times links failed far more often, with 6% gone from 2018 articles, 43% from 2008 and 72% from 1998.
#nyt-links-by-year (link to this figure)53% of articles with a rotted deep link
53% of Times articles that contained deep links had at least one that no longer worked.
#nyt-articles-with-dead-link (link to this figure)13% of working links had drifted
In a hand-checked sample of 4,500 Times links that still loaded, 13% pointed to content that had changed significantly since publication. The share was 4% for 2019 articles and about 25% for 2009 articles.
#nyt-content-drift (link to this figure)36% and 13% Sports and The Upshot
Links in the Sports section had rotted at about 36%, against 13% in The Upshot. The authors note that the sections' links differ in age, which accounts for part of the gap.
#nyt-sections (link to this figure)Law and court opinions
13 figures
50% do not produce the cited information
Half of the 555 links in US Supreme Court opinions from 1996 to 2011 no longer showed the material the court cited, in a 2013 check.
#scotus-reference-rot (link to this figure)Over 70% do not produce the cited information
In the Harvard Law Review, the Harvard Journal of Law and Technology and the Harvard Human Rights Journal, more than 70% of cited links no longer showed the cited material.
#harvard-journals-reference-rot (link to this figure)26.8% of Harvard Law Review links held the cited material
Checked by hand, only 26.8% of sampled Harvard Law Review links still held the cited material, with 29.9% for the Human Rights Journal and 34.2% for the Journal of Law and Technology.
#harvard-journals-hand-check (link to this figure)60% to 67% returned a working page
Only 60% to 67% of links in those three journals returned a working page at all.
#harvard-journals-working-page (link to this figure)63.6% returned a working page
Of the 555 Supreme Court links, 63.6% returned a working page, 24.5% returned "not found" and 10.1% got no answer from the server.
#scotus-status-codes (link to this figure)78.5% of working links showed the cited material
A working page did not mean the cited text was still there. Of the 353 Supreme Court links that loaded, 78.5% held the cited material, while 8.2% were the site's own "not found" page and 9.1% redirected elsewhere.
#scotus-working-not-cited (link to this figure)29% of cited links invalid
An earlier count found 29% of 430 links cited in Supreme Court opinions from 1996 to 2010 no longer worked.
#scotus-1996-2010-invalid (link to this figure)No clear link between a link's age and whether it works
In that study, whether a link failed showed no clear link to how long ago it was cited, to its file type or to whether the source was a government site.
#scotus-no-age-pattern (link to this figure)68% of the dead links were in the Internet Archive
The Internet Archive held a copy of 68% of those dead Supreme Court links.
#scotus-archive-copies (link to this figure)30.4% had left their original address
Of 579 law and policy documents archived by the Chesapeake Digital Preservation Group in 2007 and 2008, 176 (30.4%) were no longer at their original address by March 2011.
#chesapeake-2011 (link to this figure)160 of 579 had failed by March 2010
In the same sample, 48 links had failed by March 2008, 83 by March 2009 and 160 by March 2010.
#chesapeake-by-year (link to this figure)Over 32% of state-government links had failed
By 2010, more than 32% of the state-government links in that sample had failed, along with 25% of .gov links and more than 22% of .org links.
#chesapeake-by-domain (link to this figure)21 of 29 links from 2016 cases reported OK
A law librarian's small check of Supreme Court of Canada citations found 21 of 29 links from 2016 cases worked, against 3 of 17 from 2011 cases. Treat it as an informal spot check.
#canada-supreme-court-spot-check (link to this figure)Academic references
16 figures
1 in 5 2012 articles suffer from reference rot
About one in five science, technology and medicine articles from 2012 suffered from reference rot, meaning a reader can no longer revisit the web pages they cite as they were. The share was about 5% for articles from 1997 and more than 10% for 2005.
#klein-reference-rot (link to this figure)13% to 22% of 2012 web references rotten
For articles from 2012, 13% of web references in arXiv papers had stopped working, 22% in Elsevier journals and 14% in PubMed Central.
#klein-2012-articles (link to this figure)18% to 41% of 2005 web references rotten
For 2005 articles the figures were 18%, 41% and 36%.
#klein-2005-articles (link to this figure)34% to 80% of 1997 web references rotten
For 1997 articles they were 34%, 66% and 80%.
#klein-1997-articles (link to this figure)76% of comparable references had drifted
Of 241,091 scholarly web references that could be compared, 76.35% now show different content from the archived version closest to the citation date.
#jones-content-drift (link to this figure)About 30% have a representative archived copy
Only about 30% of over a million scholarly web references have an archived copy that matches what the author saw.
#jones-archived-copies (link to this figure)21.4% returned no live content
21.4% of the references that do have such a copy return nothing on the live web.
#jones-no-live-content (link to this figure)9.3 years median lifetime of a published URL
URLs in research abstracts from 1996 to 2010 had a median lifetime of 9.3 years, based on 17,110 URLs in Web of Science abstracts.
#hennessey-median-life (link to this figure)69% were available on the live web
69% of those URLs still worked, and 65% had a copy in at least one of the two web archives checked.
#hennessey-still-working (link to this figure)3.7% lower chance of availability for each year of age
The chance that a URL still works fell by 3.7% for each year of its age.
#hennessey-per-year (link to this figure)29.3% persistently inaccessible after ten years
Of 2,530 links in astrophysics papers published in 2015, 742 (29.3%) could not be reached in 2025. This is a preprint that has not been peer reviewed.
#astrophysics-2015-links-2025 (link to this figure)About 18 years effective half-life of a link
The authors put the half-life of those links at about 18 years. This is a preprint that has not been peer reviewed.
#astrophysics-half-life (link to this figure)12.7% and 14.6% consistently unavailable by October 2018
In an earlier check of 2015 astrophysics papers, 12.7% of links in Astronomy and Astrophysics and 14.6% in The Astrophysical Journal were down by October 2018.
#astrophysics-2015-links-2018 (link to this figure)23% to 53% of cited URLs invalid
In May 2000, 23% of URLs cited in computer science papers from 1999 no longer worked, rising to 53% for papers from 1994.
#computer-science-papers-2000 (link to this figure)23% could not be reached
Of 11,437 links in 664 doctoral theses deposited at Concordia University from 2011 to 2015, 23% could not be reached in March 2016.
#doctoral-theses-links (link to this figure)492 of 990 active links showed content drift
In a sample of those theses, 492 of 990 working links showed content that had changed.
#doctoral-theses-drift (link to this figure)Research software and code
6 figures
27.8% not accessible at the published URL
27.8% of 36,702 biomedical software tools could not be reached at the address given in the paper that introduced them.
#research-software-unreachable (link to this figure)41.9% and 17.5% before 2012 and from 2012 on
41.9% of tools cited before 2012 were unreachable, against 17.5% of tools cited from 2012 on.
#research-software-by-age (link to this figure)97.6% of GitHub links accessible
Tools hosted on GitHub were reachable 97.6% of the time, and on SourceForge 93.0%, against 70.3% for tools hosted elsewhere.
#research-software-github (link to this figure)93.98% still available on the live web
Of links to code hosting platforms in 2.6 million arXiv and PubMed Central papers, 93.98% still worked, and 68.39% of the repositories had a copy in Software Heritage.
#code-hosting-links-papers (link to this figure)9.1% returned 404 Not Found
9.1% of distinct links in source code comments, from 25,925 GitHub projects, returned "not found" in 2018.
#code-comment-links (link to this figure)37% of links to rarely linked domains decayed
For links to rarely linked domains the share was 37%.
#code-comment-rare-domains (link to this figure)Research data and file-sharing links
11 figures
32.6% of artifacts on temporary drives unavailable
Of research artifacts shared alongside software engineering papers from 2017 to 2022, 32.6% of those stored on temporary drives such as Dropbox and Google Drive had gone. The figure was 6.4% on GitHub, 7.1% on archiving services such as Zenodo and 11.8% on personal home pages. The study's temporary-drive group includes other services too.
#artifacts-temporary-drives (link to this figure)4.8% to 29.8% from 2022 papers to 2017 papers
Across those artifacts, the share that had gone ran from 4.8% for 2022 papers to 29.8% for 2017 papers.
#artifacts-by-year (link to this figure)17% lower odds a year that the data still existed
When researchers asked the authors of 516 ecology and evolution papers from 1991 to 2011 for the underlying data, the odds that it still existed fell by 17% for each year of the paper's age.
#datasets-odds-per-year (link to this figure)23% of datasets confirmed as extant
Only 121 of those 516 datasets (23%) could be confirmed to still exist.
#datasets-confirmed (link to this figure)5.4% of data links no longer available
Of 2,166 supplemental data links in Caltech's repository, 5.4% no longer worked in May 2023.
#caltech-supplemental-data (link to this figure)1.7% and 5.9% of DOIs and URLs lost
DOIs held up better than plain web addresses there, with 13 of 744 DOIs lost (1.7%) against 79 of 1,342 URLs (5.9%).
#caltech-doi-vs-url (link to this figure)75.1% and 90.1% of URLs and DOIs retrieved automatically
For data named in PLOS ONE data availability statements from 2014 to 2016, an automatic check in June 2021 could retrieve 75.1% of resources behind URLs and 90.1% behind DOIs.
#plos-data-automatic (link to this figure)78% and 98% of URLs and DOIs led to the data
Checked by hand, 78% of a sample of 350 URLs and 98% of 350 DOIs led to the data.
#plos-data-hand-check (link to this figure)72% and 45% at their address, and confirmed working
Of 927 bioinformatics web services published from 2003 to 2009, 72% were still at their published address when checked in 2009 and 2010, but only 45% could be confirmed to work as described.
#bioinformatics-services-2009 (link to this figure)25.7% not reachable
25.7% of 2,396 bioinformatics web tools published from 2010 on were unreachable in 2020. Half of the tools from 2010 still worked, against more than 90% of those from 2019 and 2020.
#bioinformatics-tools-2020 (link to this figure)How the figures were chosen
- Each figure comes from the study, paper or report that measured it, and none from an article about the study. We read each one in the source's own text on 14 October 2026 and kept the sentence it came from.
- Figures we could not find in the primary source were left out, even when they appear widely elsewhere. Examples include a 51.5% figure for economics journals behind a blocked download, and a 60% figure for UK parliamentary links that the paper never states in a sentence.
- Each figure is labelled by what it measured. "No longer resolves" means the link returns an error, a missing page or a sign-in wall. "Content changed" means the link still loads but shows something different. "Both" means the study counted the two together. A figure that measured something else, such as whether a dead link has an archived copy, carries a plain label saying so.
- Studies differ in what they sample, when they check and how they count a failure, so figures from different studies should not be compared as if they measured the same thing.
- Preprints, preliminary results and informal checks are marked as such where they appear.
- Where a study does not say when it checked its links, the figure shows the year the study was published and says the check date isn't stated.
- Our own Drive study is the one source here we ran ourselves. Its method, limits and full data are on the study page and in the Zenodo dataset.
How to cite this page
This page
Boreland, M. (2026). Link rot statistics 2026. NudgeHost. https://www.nudgehost.com/research/link-rot-statistics
A single figure
To cite a single figure, cite the study it comes from. Each one is linked above.