Obility Editorial · · 4 min read
AI assistants send people to pages that never existed
A visitor clicks a link in a ChatGPT answer and lands on your 404 page. The link looked right: your domain, a folder you use, a slug that sounds like something you would publish. The page has never existed. Most teams only find out when someone forwards a screenshot.
Ahrefs measured how often this happens in a study published in September 2025, covering 16 million unique URLs cited by ChatGPT, Perplexity, Copilot, Gemini, Claude and Mistral. In its first test, Ahrefs took URLs that received visits referred by an assistant and flagged a page as a likely 404 when its title contained “404” or “not found”. For ChatGPT that was 1.01 percent of referred URLs, against 0.58 percent for Claude, 0.34 percent for Copilot, 0.31 percent for Perplexity and 0.21 percent for Gemini. URLs referred by Google ran at 0.15 percent. A second test checked the live status of cited URLs whether or not anyone clicked them: 2.38 percent of ChatGPT’s cited URLs returned a 404, against 0.84 percent of URLs in Google’s top 20 results.
Ahrefs says both figures probably undercount. Some error pages do not say “404” in their title, and a URL that never existed is often missing from a crawler database, so its status was never checked.
Two different failures produce the same 404
Ahrefs suggests two causes. Some cited URLs did exist once and were later moved or deleted without a redirect. A model answering from what it learned in training can still hold the old address. Others are true inventions that fit the pattern of the site. The most visited invented URLs on the Ahrefs blog were /blog/internal-links/ and /blog/newsletter/, which look exactly like pages Ahrefs would publish and do not exist.
The split matters because the fixes differ. A moved page is a debt you created: you changed a URL and left nothing at the old one. That has a standard fix. An invented URL is a prediction. The model has learned what sites like yours usually contain and filled in a path it expected to find.
Read the invented paths as a list of expectations
This is the part most advice skips. A list of 404s reached from AI answers is a short record of things a model believed your site should say. Take an illustrative B2B software company. If assistants keep sending visitors to /integrations/salesforce and the company has no such page, a buyer probably asked about that integration and the model expected the vendor to document it. A dead end at /pricing/enterprise suggests someone asked about enterprise terms. Each path is a question that reached an answer engine, involved your brand, and found nothing on your site to ground the answer.
That is not a reason to build a page for every invented path. Some are noise, and some ask for things you do not offer. Work through each one with two questions. Is there a real fact here that buyers need? Do we already have a page that states it? If the page exists under another URL, the problem is that the fact is hard to find or easy to misread. If no page exists and the fact is true, you have a content brief. If the fact is false, for example an integration you do not have, the useful page may be a plain statement of what you do connect with, because a model that guesses your URLs can also guess your feature list.
Redirect only where a clear replacement exists
Google’s Search Central documentation on soft 404 errors draws the line clearly. When a page has moved or has a clear replacement, return a 301 permanent redirect. When there is no replacement with similar content, return a 404 or 410 and make the error page useful: the same navigation as the rest of the site, links to popular pages, a way to report a broken link, and a real 404 status code. A page that says “not found” while returning a 200 status is what Google calls a soft 404.
So resist the reflex to point every invented URL at the homepage. A visitor who asked about enterprise pricing and lands on the homepage has been told nothing, and the redirect hides the signal you just found. Redirect the paths that have an obvious real equivalent, such as an invented newsletter URL to your actual signup page. Let the rest fall on a 404 page written for someone arriving from an AI answer: one line on what you do, then links to product, pricing and documentation.
Finding them takes an afternoon. Filter your analytics or server logs to visits whose referrer is an AI assistant, check the status of each landing URL, and keep the ones above a threshold. Ahrefs used more than 10 visits a month; pick a number that suits your traffic. Open each one before acting, since some will be real pages that failed for another reason.
Keep the size of the problem in proportion
Ahrefs is candid about scale. At its last measure, AI assistants accounted for 0.25 percent of an average site’s traffic, against 39.35 percent for Google. Dead links are a small share of a small share, so fixing them is a cheap cleanup, not a growth lever. The more durable value is the list of expectations, which costs nothing extra once you have the log.
A redirect also does not change what the model says. The next answer may invent a different path, or describe a feature you do not have without linking anywhere. Track what assistants say about your category separately from where they send people. In Obility, Agent Analytics keeps referral visits from answer engines apart from crawler requests and citation observations and looks at their landing pages, which is where invented URLs surface. Content Agents start a brief by identifying the existing page that should answer a question before creating another URL, which is the right default for most invented paths.