AI assistants send visitors to 404 pages 2.87 times more often than Google Search, according to a September 2025 Ahrefs analysis of 16 million URLs. A separate study from SE Ranking, published in December 2025 and built on 145,463 URLs cited by ChatGPT, found its broken-link rate running at nearly double Google's AI Overviews. For a brand that has spent months earning mentions in AI answers, that gap is not a rounding error. A citation pointing at a dead page delivers nothing to the reader who clicks it, and the same retrieval behavior producing those 404s is quietly deciding whether an AI engine can find and cite your content at all.
How much worse are AI citations, exactly
Ahrefs measured the gap two ways. Tracking real visitor traffic referred from each assistant, Ahrefs' analysis found ChatGPT-referred sessions landing on a 404 page 1.01% of the time, against 0.15% for Google Search. Claude's outbound links hit 404s at 0.58%, Perplexity at 0.31%, Gemini at 0.21%, and Mistral at 0.12%.
Checking the URLs each assistant actually cites in its own answer database, rather than what visitors click, widened the gap further. ChatGPT's cited-URL 404 rate rose to 2.38%, against a Google search-results baseline of 0.84%. SE Ranking's separate study, based on 145,463 URLs pulled from 100,000 ChatGPT prompts, put ChatGPT's overall broken-link rate at 1.22% against 0.56% for Google's AI Overviews, roughly twice as likely to fail. Ninety-one percent of those failures were plain 404s rather than server errors, and even so, 97.55% of ChatGPT's citations still resolved successfully.
ChatGPT sends visitors to 404 pages 2.87 times more often than Google Search, whether you measure by real visitor traffic or by the URLs the assistant cites directly.
Why ChatGPT is the worst offender
The volume explains part of it. Search Engine Journal's analysis of Alli AI crawler data, covering 24.4 million requests across 69 sites over 55 days in early 2026, found OpenAI's ChatGPT-User crawler generating 3.6 times as many requests as Googlebot, and 3.8 times as many once GPTBot is counted alongside it. Yet ChatGPT-User's own fetches succeeded 99.99% of the time, with an 11-millisecond average response, a cleaner hit rate than Googlebot's 96.3% success and 84-millisecond average, which the report attributes to Googlebot still re-crawling legacy URLs that no longer exist.
That is the paradox. The crawler's live fetches are nearly flawless, but the citations ChatGPT eventually shows a user are dead more often than Google's are. The failure point sits after the crawl, when the model pulls a URL from a stale index entry, a cached snippet, or an outright hallucinated URL, a real domain stitched to a path that was never live. It fits a pattern we've measured before: AI engines already skew toward citing fresher pages than Google does on average, which means an old, uncrawled link sitting in an engine's index gets penalized twice, once for being stale and again when it quietly breaks.
Redirect chains: AI crawlers follow them, just not the way you'd expect
GPTBot, ClaudeBot, and PerplexityBot all follow standard 301 and 302 HTTP redirects, according to a 2026 review of AI crawler behavior by crawlability tool Urllo. What none of them do is execute a JavaScript redirect or a meta-refresh tag, the same rendering gap documented for AI crawlers generally: a crawler can fetch the script that would fire the redirect in a browser, but it never runs it, so the visit simply dead-ends. Redirect handling is separate from access control. What these bots are and aren't allowed to fetch in the first place is covered in our breakdown of what robots.txt actually blocks for each one. None of the major AI crawlers publish a documented hop limit, unlike Google's roughly five-hop tolerance, and Urllo's own guidance is to treat anything past two hops as a liability rather than wait to find out where an undocumented limit sits.
That same SE Ranking dataset adds a wrinkle. URLs ChatGPT actually cites carry a redirect only 0.79% of the time, compared with 5.75% for Google's organic results and 2.85% for AI Overviews. That is not evidence ChatGPT handles redirects more gracefully. It is evidence that when a cited URL has moved, ChatGPT is more likely to fail outright, landing in the 404 numbers above, than to follow the chain through to wherever the page now lives.
The crawl-budget number no citation report shows you
Every stat above measures what a user or the answer database actually sees. Urllo's crawl-level log analysis found a bigger, hidden number: ChatGPT spends 34.82% of its total crawl requests hitting pages that return a 404, against 8.22% for Googlebot. That figure has nothing to do with what gets cited. It is the share of a crawler's own crawl-to-refer ratio that produces no retrievable content at all, because the index it is working from still carries URLs your site stopped serving months ago. A citation-tracking tool will never surface that number, because by definition it only sees the answers that made it out the other side.
The redirect and 404 hygiene GEO actually requires
None of this calls for new GEO tooling that most technical SEO teams don't already own. It calls for pointing what they have at AI user agents specifically, the same way it has long been pointed at Googlebot:
Audit for redirect chains longer than one hop with Screaming Frog or Ahrefs Site Audit, and collapse each to a direct 301 into the live URL.
Point internal links and XML sitemap entries straight at final destinations instead of through a redirect. A crawler that follows a chain just to find a sitemap URL is one hop closer to giving up.
Replace any JavaScript or meta-refresh redirect with a server-side 301. A crawler that can fetch a script without running it will never see where that script meant to send it.
Pull GPTBot, ClaudeBot, and PerplexityBot user agents out of server logs and check what share of their requests return a 404 or 5xx, the way log analysis built for AI crawlers is designed to do.
Before blocking a spike of bot traffic hitting dead URLs, confirm it is actually GPTBot or ClaudeBot and not a spoofed crawler impersonating one. Blocking the real thing over a false positive costs you the citation you were trying to protect.
Server log analysis remains the only method that catches the 34.82% figure. None of the citation-tracking dashboards on the market today audit what a crawler hits before an answer gets written, only what eventually shows up inside one.
Frequently Asked Questions
Why do AI assistants send users to broken links more often than Google?
Because most AI citations come from an index or cache rather than a live, real-time fetch. Google recrawls and revalidates URLs constantly, while AI engines often cite a URL that was valid when it was last indexed, or one a model produces as a plausible-looking hallucination, so the failure only shows up when a reader clicks.
How much more likely is ChatGPT to cite a 404 page than Google?
Ahrefs found ChatGPT sending visitors to 404 pages 2.87 times more often than Google Search across both traffic data and its own citation database, and SE Ranking separately measured ChatGPT's broken-link rate at roughly twice that of Google's AI Overviews.
Do AI crawlers like GPTBot and ClaudeBot follow redirects?
They follow standard server-side 301 and 302 HTTP redirects, but none of them execute JavaScript redirects or meta-refresh tags, and none publish a documented limit on how many redirect hops they will follow before giving up.
How many redirect hops is too many for AI crawlers?
There's no official limit to test against, so GEO practice treats two hops as the ceiling, with one hop as optimal. Google tolerates roughly five hops, but AI crawlers are widely assumed to have less patience given how much of their crawl budget already goes to waste on dead URLs.
How can I tell if AI crawlers are hitting broken links on my site?
Filter your server logs for the GPTBot, ClaudeBot, and PerplexityBot user agents and check what share of their requests return a 404 or 5xx status. That crawl-level view catches problems citation-tracking tools miss, since those tools only see what eventually appears in a chat answer.