How often do AI citations lead to broken pages?

Scrunch reports that citation activity for a source falls by half in about four to five weeks. A 2026 academic study found 5% to 18% citation URLs that did not load across deep-research and academic tasks. Those studies measure citation loss or a broader set of failed responses. This study isolates strict 404 and 410 results on recently cited AI search pages.

In the Parse index, 93.7% of pages returned a 2xx success and 4.2% returned a 3xx redirect. Other 4xx responses accounted for 1.3%, strict broken responses for 0.67%, and 5xx server errors for 0.17%. A citation can disappear even when its page still loads.

ChatGPT Search had 418 broken pages among 19,942 eligible pages, a 2.1% rate. Google AI Mode had 598 among 134,116, a 0.45% rate. The ChatGPT Search rate was 4.7 times the Google AI Mode rate.

A broken-link audit should keep engines separate. The aggregate rate hides a large platform difference and does not show whether the engine created the URL, retrieved an old URL, or cited a page that the publisher later removed.

Among the eight largest classified source groups, reference pages had a 0.99% broken-page rate. SaaS pages were at 0.76%, general web pages 0.66%, news 0.53%, finance 0.51%, editorial 0.44%, e-commerce 0.37%, and social pages 0.10%.

Source type changes the baseline, but no large group reached a 1% rate. Use the group rate to prioritize review, not to assume that a specific page is broken.

We required at least 100 eligible pages and at least five broken pages before displaying a domain rate. En.wikipedia.org had 10 broken pages among 180 checked pages, or 5.6%. Finance.yahoo.com had six among 183, or 3.3%.

The threshold removes small-domain percentages built on a handful of pages. These rates identify where to inspect URLs. They do not establish why the pages returned 404 or 410.

Rootly's page about reducing alert fatigue was cited 34 times across nine prompts before its qualifying check returned 404. Time's page about job-search networking was cited 19 times across six prompts. Two more Rootly and UP42 pages followed with 18 and 17 citations.

The page list is what you can act on. A broken page cited repeatedly deserves attention before a one-off URL because it can affect more answers and more buyer questions.

Among the eligible pages, 579 had more than one check and at least one 404 or 410 during the window. Of those pages, 534 were still broken on the latest check and 45 returned a different HTTP status.

A prior broken result is not permanent proof. Recheck the exact URL before redirecting or rebuilding it, then keep the latest result as the decision record.

The current engines cited 1,808,782 pages in all. We excluded 1,547,242 without a check in the window, 14,431 whose latest check had no HTTP status, and 94,796 checked before their latest recorded citation. That left 152,313 unique pages as the base for every rate here.

This is measured in the Parse index, not a live-state claim or a random sample of the web. The seven-day sensitivity check found 395 broken pages among 62,276 checked pages, or 0.63%, close to the headline rate. The result applies to eligible checked pages and should not be extrapolated to pages Parse did not check.