We loaded 4,706 pages that AI engines cited from brands' own sites
Does ChatGPT send people to broken pages on your website? The studies people cite looked at every link an engine gave out. We wanted the narrower version a brand cares about: when an engine cites your own site, does the page still work?
We took every answer Proofsource collected from ChatGPT, Claude, Perplexity and Google AI Overviews between October 4 and November 8, 2026, for the brands our users track. Any cited link on the brand's own domain, a subdomain or an alternate domain counted. That gave 18,169 citations of 4,706 different URLs, for 159 brands. On November 9 we requested each URL the way a browser would, followed redirects, and recorded where it landed.
789 URLs (16.8%) couldn't be checked: the site blocked or rate-limited the request. We left those out, which leaves 3,917.
87 of 3,917 cited pages were broken, 2.2%
We counted a page as broken in three cases: it returned "not found" (404 or 410), it redirected to the homepage, or its hostname no longer existed. 87 URLs met one of those (2.2%, 1.8% to 2.7%). 56 returned not found, 23 redirected to the homepage and 8 sat on a hostname that no longer resolved. We read every one by hand and dropped redirects that were fine, like a language prefix or a move to a login app.
Counted by citation rather than by URL, 308 of 16,160 checkable citations pointed at a broken page (1.9%).
29 of 159 brands had at least one
That's 18.2% of brands (13.0% to 25.0%). For most of them it was a small leak: in the median affected brand, 2.8% of its own-site citations hit a broken page. For 7 brands it was 10% or more, and for the worst, 45%.
Brands' own pages broke more often than the other sites these answers cited. In a random sample of 2,400 citations of other sites, 22 of the 2,070 we could check were broken (1.1%), against 1.9% for own-site citations.
ChatGPT had the fewest broken own-site links, Perplexity the most
Share of own-site citations that hit a broken page: ChatGPT 1.0%, Google AI Overviews 1.4%, Claude 2.3%, Perplexity 2.5%.
That's the reverse of the best-known study. Ahrefs found ChatGPT had the highest 404 rate of the assistants it measured, 2.38% of cited URLs. SE Ranking put ChatGPT at 1.22% against 0.56% for AI Overviews. Both counted every URL cited. We counted only the brand's own pages, so the numbers measure different things, and a gap of one point in our data is a few dozen URLs.
At least 25 of the 56 missing pages once existed
The common explanation is that engines invent URLs that look right. We found little of that. For 25 of the 56 not-found URLs, the Internet Archive holds a copy that loaded: product brochures and data sheets, a sustainability report, about and terms pages, case studies. The brand removed or moved them without a redirect.
The other 31 have no archived copy, which is common for small sites, so we can't prove where they came from. Most look like ordinary pages of the site: about, contact and course pages, blog posts, .html pages from an earlier build. One site moved its blog off a subdomain and its redirect sent old posts to a doubled path, /blog/blog/, that doesn't exist. Only one URL looked invented: it ended in a stray backslash.
The other two groups are plainly the brand's doing. 18 of the 23 homepage redirects came from one site that had paused and sent every page home. All 8 dead hostnames were subdomains the brand had retired.
What we'd do
- List every page on your site that AI answers cite and load each one. Fix the ones that fail before anything else on that list.
- Redirect each removed page to its closest live page, not the homepage. Google's own guide to site moves warns that many old URLs pointed at one irrelevant page, such as the home page, may be treated as a soft 404.
- Before a redesign or a move off a subdomain, export the cited URLs and map each one. Keep PDFs at stable addresses, or redirect them when you replace them.
How Proofsource handles it
Proofsource keeps every URL each answer cited, by engine and by question, including the pages on your own site. That's the list this check starts from.
Where this could be wrong
We checked once, on one day. A page that was down for an hour on November 9 counts as broken, and a page that broke after it was cited counts the same as one that was broken all along. The 16.8% we couldn't check could hide more broken pages, or fewer.
Calling a homepage redirect broken is our judgement, made by hand. The 159 brands are the ones our users chose to track, many of them small and many in India, so their sites change more often than a large company's might. This is five weeks and four engines.
