The State of Broken Links 2026: we checked 6.5 million of them

Between 15 April and 31 August 2026, Broken Link Scan checked 6,501,595 links across 886 websites. 517,430 of them were dead — 7.96%, or roughly one link in thirteen.

Every figure on this page comes from links we requested ourselves, not from an index or an estimate. The method, including what we threw out and why, is at the bottom. The data is free to quote.

The headline numbers

The web is not uniformly broken — it is split

The single most misleading way to read this dataset is to take the 7.96% and picture every site carrying one dead link in thirteen. That is not what the data looks like at all.

The median site has a dead-link rate of 0.84%. Half of all sites are cleaner than that. A third — 327 of them — had no dead link at all on the day we looked. The average is dragged up by a minority in genuinely bad shape.

So the useful sentence is not "8% of the web is dead". It is this: two sites in three carry at least one dead link, most carry very few, and one site in five is past 10% — the point where a visitor starts running into them by accident.

Why one site in five gets that bad

Nothing in the dataset says why, and we are not going to invent a cause we did not measure. What we can say is where the dead links sit, and it is rarely where people look first.

• Outbound links rot on somebody else's schedule. A resource you linked to in 2019 can disappear without anything on your site changing.
• Migrations leave a long tail. A redirect map covers the pages someone remembered; the links inside old articles are the ones nobody re-checks.
• Images count. A missing image is a dead link that no reader reports and no CMS flags — it simply renders as a gap.
• The bigger the site, the longer the tail. Rates above 10% cluster in sites with years of archives, not in new ones.

What we counted as dead

A link counts as dead when a real request for it fails: a 4xx or 5xx response, a DNS failure, a connection that never completes. We request every link ourselves rather than reading a status from an index.

Three things are deliberately not counted as dead, because each one produces false alarms at scale:

• 429 (rate limited) — that is the server asking us to slow down, not a broken link. Counting it cost us a customer once; it is now skipped and retried.
• Bot challenges — a page behind a challenge answers with an error to us and with content to a human. Sites showing this pattern are flagged, not scored.
• Timeouts on the site's own pages — a slow server is a slow server. An internal link that times out is retried before anything is concluded.

Method, and what we excluded

The corpus is every completed scan run through Broken Link Scan between 15 April and 31 August 2026 that found at least one link, grouped by site so that a site scanned ten times counts once. That is 970 sites before exclusions.

We excluded 84 sites whose dead-link rate came out at 50% or higher — 69,807 links, 1.06% of the corpus. A site where half the links fail is almost always a measurement artefact rather than a ruin: a firewall answering every request with an error, a bot challenge, an entire host down during the crawl. Keeping them would have pushed the headline figure from 7.96% to 8.64%, and that difference would have been noise, not signal. We would rather publish the smaller number and be right.

Two limits worth stating plainly. First, this is a self-selected sample: these are sites whose owners chose to check them, which likely skews towards sites someone already suspected had a problem. Second, sites vary enormously in size, so the per-site median and the global rate answer different questions — we have given both rather than picking the flattering one.

Using these figures

Quote any of them. We would appreciate a link back to this page so readers can see the method, but there is no permission to ask for. If you want a cut of the data we have not published — by site size, by response code, by month — write to [email protected] and we will pull it.

We intend to re-run this every six months. The interesting number will not be 7.96%; it will be whether it moves.