Where do AI citation links actually go?

130,935 links, or 83.3% of verdicts, landed exactly on the cited address. Another 21,298, or 13.6%, resolved to a different address on the same website. 2,637, or 1.7%, resolved to a different website. 2,002, or 1.3%, were confirmed gone with a 404, 410, or a soft 404: a page that returns success but shows not-found content. 282, or 0.18%, ended on a bare domain page with no content.

The ranking matters for an audit. A same-site move usually preserves some of the click's value. A cross-site move, a dead page, or an empty domain page does not.

Link-rot research measures death: Pew found 38% of 2013 webpages unreachable a decade later, and Ahrefs found 66.5% of backlinks from the last nine years dead. In recently cited AI answers, death is rare and movement is common. 23,935 checked links, or 15.2% of verdicts, resolved to a different address, while 2,002, or 1.3%, were confirmed gone. That is 12 moved links for every dead one.

Our earlier study of the same index found 0.67% of recently cited pages returning 404 or 410. A marketer who audits only for broken links sees the smallest part of the problem.

Among domains with at least 100 checked cited addresses, apple.com led: 1,163 of its 1,269 checked addresses, or 91.7%, resolved somewhere else, mostly by dropping a language or region segment. microsoft.com followed at 82.0% and squareup.com at 82.0%, for the same reason. walmart.com reached 60.8% and wikipedia.org 53.1%, driven by other address changes such as host moves and homepage fallbacks.

A high moved share is not misconduct by the site. It means AI engines attach addresses that are not the address the site itself settles on, so anyone tracking citations by exact URL will under-count these domains.

7,621 links, or 4.8% of verdicts, resolved from a specific cited page to the site's homepage. amazon.com led with 180 homepage fallbacks, followed by arxiv.org with 167 and youtube.com with 140.

A homepage fallback has two causes we cannot fully separate. Some sites redirect a removed page to the homepage instead of returning 404, the classic soft-404 pattern. Other sites route an automated or cookie-less visit to the homepage as bot handling, and a person clicking the same link may still reach the page. Either way, the cited address does not deliver the cited content to a first-time automated visitor.

A deep link is a cited address with a path, such as a product or article page. A bare-domain link is a homepage-level address with no path. Of 65,713 deep links that resolved, 20,630, or 31.4%, ended on a different address. Of 89,157 bare-domain links, 3,305, or 3.7%, did. Deep links moved 8.5 times as often.

Domains are stable and pages are not. The more specific the address an engine cites, the more likely that address has a different current form.

5,958 links, or 3.8% of verdicts, differ from their resolved address only by a language or region path segment, such as /en or /en-us. This bucket explains most of the movement at the top of the leaderboard: 867 of apple.com's 1,163 moved addresses and 586 of microsoft.com's 751 are language cleanup.

The engine cited a real, working address; it is simply not the canonical one. For citation tracking this means exact-URL matching splits one page into several counted addresses, which is the same failure mode we measured for tracking parameters.

6,054 links, or 3.9% of verdicts, resolved to a genuinely different page on the same website: a renamed product page, a reorganized documentation section, a replaced article. A further 1,665, or 1.1%, resolved to a different host on the same site, such as a language subdomain or a support subdomain.

These are the moves a reader is least likely to notice. The site looks right and the content is close, but it is not the page the answer cited, and any anchor-specific claim can be gone.

2,637 links, or 1.7% of verdicts, resolved to a different website than the one cited. The examples read as a record of acquisitions, rebrands, and platform moves: mandiant.com now resolves to cloud.google.com/security/mandiant, apptentive.com to alchemer.com, openai.com/chatgpt to chatgpt.com, and riffusion.com to producer.ai.

In these cases the answer's citation names a source that, as a website, no longer exists in that form. The content may survive at the destination, but the attribution the reader sees does not match the site that loads.

2,284 links, or 1.4% of verdicts, were dead ends: 2,002 confirmed gone with a 404, 410, or soft 404, and 282 ending on a bare domain page with no content. Within the cross-site resolutions, 18 cited addresses resolved to a domain marketplace or for-sale page, including profound.ai, bluefish.ai, and rankscale.com. Eighteen addresses is a small tail, and we report it as a count, not a rate.

A cited address landing on a for-sale page is the sharpest version of the problem: the answer lends credibility to a domain that now belongs to nobody. It also shows that engines can keep citing an address after the site behind it has shut down.

91,073 of 248,227 checks, or 36.7%, ended without a verdict: 89,789 because the site refused the automated check with a 401, 403, 429, or a bot wall, 1,235 for inconclusive results such as unrendered pages or unreachable hosts, and 49 for empty results. Our study of access blocking covers the refusal side; every share in this report uses the 157,154 verdicts as its denominator, so the results describe sites that allow checks and may differ on sites that refuse them.

Two more limits. First, the checker also normalizes duplicate-route variants, so the language-cleanup bucket includes canonical rewrites alongside server redirects; we report it as its own labeled bucket rather than mixing it into movement claims. Second, the same-website versus different-website split uses a two-label root-domain heuristic with a list of common two-part country suffixes, so a small number of hosts can be misclassified. Checks ran between September 9, 2025 and August 29, 2026, and a link's state can postdate its most recent citation.

AI engines remember addresses, not pages. 16.7% of checked citation links did not land on the cited address, deep links moved 8.5 times as often as homepages, and moved links outnumber dead ones 12 to 1.

Keep any address that AI answers cite stable for as long as you can. When a page must move, redirect the old address to the equivalent page, never to the homepage. Publish one canonical address per page so engines do not cite regional variants. And when you audit AI visibility, resolve every cited link to its destination before you count it, or you will miss most of what changed.