There is a group of people an AI already recommended you to. They clicked. They hit a page that doesn’t exist.

They won’t go back and look again, and they won’t show up in any report you have — a URL like /blog/geo-guide-2026 is one you never built, so it never occurs to you to check it. This piece covers how to recover them, and why recovering them once doesn’t close the matter.

One in every 233 visitors an AI sends you hits a wall

In September 2025, Ryan Law and Xibeijia Guan at Ahrefs checked 16 million URLs cited by AI assistants across ChatGPT, Claude, Perplexity, Copilot, Gemini and Mistral.

Overall, 0.43% of the links AI assistants hand out return a 404, against 0.15% for Google Search — 2.87x. The spread between engines is wider:

Engine 404 rate on clicked URLs
ChatGPT 1.01%
Claude 0.58%
Copilot 0.34%
Perplexity 0.31%
Gemini 0.21%
Mistral 0.12%
Google Search (reference) 0.15%

ChatGPT is the worst of them: one in every hundred clicked URLs is dead. Counting every cited URL rather than only the clicked ones, it’s 2.38%.

0.43% sounds small until you look at the denominator. These are people an AI already recommended you to and who then decided to click — the highest-intent segment in the whole funnel. That traffic converts far better than ordinary organic search to begin with; Ahrefs’ own figure is 23x. What you’re losing isn’t 0.43% of pageviews. It’s 0.43% of qualified prospects.

The AI isn’t copying wrong — it’s inferring the URL

Dead links come from three places, and they call for different handling.

Some the model guesses from structure. It has seen how tens of thousands of sites lay out their URLs, knows blogs usually live under /blog/ and pricing usually sits at /pricing, and assembles one on the spot. If your site happens to be arranged differently — pricing at /plans, articles under /insights/ — it guesses wrong. That URL never existed.

Some are old URLs. You restructured, migrated a domain, or retired a campaign page, and the old version is still sitting in the model’s memory. These were real once.

And some drift in the retelling. The cited source is correct, but a character or two gets mangled while the answer is generated.

You can’t prevent the first kind, only catch it. The second is entirely self-inflicted and the most worth fixing. The third is rare, but it tends to look almost identical to a real URL, so it’s easy to dismiss as a typo when you’re scanning the list.

Finding them: only the 404s with traffic

The phrase that matters is with traffic. Every site’s 404 list runs to hundreds or thousands of entries, most of them scanners probing at random. What you want are the handful that real people actually landed on.

Three places to look, each covering something different:

Server logs are the most complete. Pull requests with a 404 status where the referrer or user agent maps to an AI source. The catch is that someone has to be comfortable reading logs.

GA4 comes next. If your 404 page fires an event, you can see which paths recorded a page_view and cross-reference parameters like utm_source=chatgpt.com. Bear in mind that many AI sources arrive without a referrer and get bucketed as direct, so what GA4 shows you is a floor, not a total.

Search Console takes the least setup and is the least precise. In the Page Indexing report, under “Not found (404),” look for paths you’re certain you never built but which are shaped like real URLs. It can’t isolate AI sources, but it needs no configuration.

Run all three, take the union, then keep only the ones with actual clicks in the last 90 days. You’ll usually be left with somewhere between a dozen and a few dozen. Not more.

Fixing them: map each one, don’t dump them on the homepage

Point every dead URL at the most relevant real page with a 301.

The tempting shortcut is to send all of them to the homepage with a single rule. Don’t. Someone arriving from an AI answer has a specific question in hand — they want to know how your pricing is calculated. Drop them on the homepage and they have to start looking again, and most of them leave at that step. You’ve recovered nothing.

For the ones that don’t map to anything, hold off on deleting them. They’re topic signals: the AI believes you should have this page, which means it’s what it was reaching for when answering a class of question. Rather than redirecting to something unrelated, consider actually writing the page.

Then there’s the 404 page itself. Even with everything mapped, some will slip through. A 404 page shouldn’t only say “page not found” — give it a search box, links to your main sections, and a way to contact you, so the people who land there get a second chance.

Preventing them: treat a published URL as permanent

  • Don’t restructure URLs for aesthetics. Of all the sources of dead links, this is the least necessary.
  • If you must change them, always leave a 301 — and leave it permanently, not for three months. A model’s memory outlasts your patience.
  • Don’t delete retired campaign and pricing pages, redirect them to the current equivalent.
  • Name URLs the conventional way: /pricing, /blog/, /contact. Paths everyone else uses are the ones a model guesses correctly. This is one of the few places where following convention beats having a style of your own.

After you’ve recovered them once

The work above fits in an afternoon, and what it recovers is pure gain — one of the very few moves in GEO that pays back immediately.

It isn’t a one-off, though. Every model update, every change you make to your site structure, every new article that gets cited will grow a fresh crop of invented URLs. Leave the list alone for three months and it rebuilds itself, with no notification of any kind. A 404 happens quietly; nobody writes in to tell you they hit a dead link on your site.

What separates sites here is whether someone is watching consistently. It’s the same shape of problem as the AI traffic attribution blind spot: the work isn’t hard, you just won’t remember to do it every month.

If you’re unsure how many people your site is dropping right now, we can run a scan first.

Further reading