ChatGPT Is Sending Buyers to Pages That Don't Exist on Your Website

Short answer: AI assistants sometimes put links to your own website into their answers that point at pages you never built. In three weeks of tracked answers for one global hardware brand, in one market, we found 18 such URLs across 27 answers, and ChatGPT, Gemini and Claude all produced them. They were not random: the models copied a URL structure the brand uses in other markets and applied it where it does not exist. You can check your own site in five minutes by filtering your 404 logs for utm_source=chatgpt.com, and one redirect rule usually fixes the whole pattern. Ayzeo's new URL Health view now runs this check automatically for every link to your own site that your tracked AI answers expose, confirms broken links with two checks, suggests a possible fix and emails you a digest. It is included in the Pro and Enterprise plans.

1. The Link That Looked Perfect

We track how AI assistants answer buying questions for a global hardware brand. During a review call, the client spotted a link in a ChatGPT answer that looked perfect: the right domain, the right country store, the right page title.

It was a 404.

Nobody had noticed before, and that is the uncomfortable part. The link is not on the brand's website, so no crawler and no SEO audit ever requests it. It exists only inside AI answers, where the brand's team does not look and where buyers do.

2. What the Data Showed: 18 Invented URLs in 27 Answers

We went through every stored answer for that market. Three weeks, one market: 18 invented URLs across 27 AI answers. By invented we mean an address on the brand's own domain that returns a 404 and that the brand never published. ChatGPT, Gemini and Claude all produced them. Same shape every single time.

This is not one brand's bad luck. In 2025, Ahrefs checked 16 million URLs cited by AI assistants and found that the assistants send visitors to 404 pages 2.87 times as often as Google Search does. ChatGPT was the most frequent offender, with 1.01% of the URLs people clicked returning a 404.

3. The Pattern: AI Learns Your URL Structure

The shape is the interesting part. The brand's clearance stores in several other countries share one URL structure. The UK store uses a different one. The models learned the structure from the markets where it exists and applied it to the UK, where it does not. They even attached the real page's title to the invented link.

A made-up example of the same shape:

• Real page (200): example.com/uk/en/deals/black-friday/
• Link in the answer (404): example.com/uk/clearance-uk/en/deals/black-friday/

One extra path segment, borrowed from the way other markets are organized. It reads as plausible to a person and to a model, and it goes nowhere.

So this is not random hallucination. The models learn your site architecture. If your URL patterns are inconsistent across markets, AI fills the gap with pages you never built, and sends real buyers there.

4. Why These 404s Cost More Than Ordinary Ones

Every site has 404s. These are different in four ways.

• The visitor is at the end of a buying question. They asked which product to buy or where to get a deal, and the answer was your brand.
• The link arrives with a recommendation. The assistant has just told the reader your page is the one to open. A 404 spends that trust.
• It repeats. An invented address is not a one-off. In our data, 18 addresses appeared in 27 answers: the same invented link came back in answer after answer.
• Your usual tools cannot see it. Crawlers and site audits follow links on your site, and these links are not on your site. Only your server logs record them, and only after someone has clicked.

In volume this is still small. Ahrefs data puts AI assistants at about 0.25% of an average website's traffic. But these are visitors who arrive with a recommendation for your brand, and a dead end is the last thing they should find.

5. The Five-Minute Check in Your Logs

ChatGPT adds utm_source=chatgpt.com to the links in its answers, as OpenAI states in its publisher FAQ. That makes the visits it sends easy to find:

1. Open your server, CDN or hosting logs for the last 30 days.
2. Keep the requests that returned a 404 and whose address contains utm_source=chatgpt.com.
3. Group them by path. Each path is a page ChatGPT sent a person to that does not exist.
4. Look for shared shapes: the same extra segment, a folder that belongs to another market, a page name that exists under a different path.

In Google Analytics 4, the closest equivalent is to filter your page reports to the session source chatgpt.com and look for your 404 page's title.

Two limits to keep in mind. This finds only ChatGPT's clicks: other assistants do not always tag their links. And a link that nobody has clicked yet leaves no trace in any log, however many answers show it.

6. The Fix: One Redirect Rule

When the invented links share a pattern, one wildcard redirect covers all of them: point the invented pattern at the real pages. For the example above:

/uk/clearance-uk/* → /uk/*

Make it a permanent (301) redirect. Every AI mistake then becomes traffic instead of a dead end, and it keeps working for as long as the assistants repeat the invented address, which they do.

Three habits make the fix hold:

• Check each target before you ship the rule. The redirect should land a reader on the page the answer promised, not on the nearest category page.
• Fix the cause, not only the symptom. If your markets use different URL structures for the same kind of page, aligning them removes the gap the models fill.
• Keep a useful 404 page. A search box and links to your main categories rescue the visits no rule covers yet.

7. Now Automatic: URL Health in Ayzeo

Checking your logs once is useful. Doing it every week for every market is a job nobody has time for, and logs only show what someone already clicked. So we built the check into Ayzeo.

URL Health collects every link to your own site from the AI answers Ayzeo tracks for your prompts and checks each one on your live site. For every link it shows:

• The status: 200, a redirect with the page it lands on, 404 or 410, soft 404 (a page that answers 200 but behaves like a missing one), or unreachable.
• How often AI showed it: the number of answers, the models, and when it was first and last seen.
• Where it appeared: in the answer text a reader could click, or only among the pages a model visited or consulted. The default view shows the links in the answer text, because those are the ones people click.
• A possible fix: "Did you mean ...?", a working page on your site that matches the broken link, with a redirect rule to copy, one wildcard rule when several links share a pattern.

A link counts as broken only when two separate checks agree. Timeouts, server errors and firewall blocks are shown as unreachable and never counted as broken, so a bad night on your server does not turn into a false alarm.

Around the view:

• An email digest lists the broken links shown in answers, weekly by default or after every full scan, grouped by model. When Google Analytics is connected, a broken link that received visits from AI assistants while it was broken shows how many, the number that makes the case to your web team.
• Badges mark broken, rechecking and redirected links in AI Sources & Destinations, in the answer view and in Prompt Insights.
• The monthly report includes the number of broken links shown in AI answers.
• MCP: the get_url_health tool gives your AI assistant the same list, so it can draft the redirect rules for you.

URL Health is included in the Pro and Enterprise plans. There is nothing to set up: the first check starts within about an hour of your next scan, or the next morning (Berlin time) if the scan ends at night.

8. What URL Health Does and Does Not Do

• It checks links to your own site only: your project's domain, its subdomains and other domains of the same brand. Links to other websites are never checked.
• It never changes your site. A possible fix is a pattern match that opened correctly when we checked it, not knowledge of what the assistant meant. Your team verifies it and sets up the redirect.
• It sees the answers Ayzeo tracks for your prompts, not every answer every assistant gives. A broken link can also reach buyers through a question you do not track, so your logs stay a useful complement.
• It is a polite visitor. The link checker identifies itself as AyzeoLinkCheck, paces its requests and follows your robots.txt. The details for your web team are on ayzeo.com/bot.

9. Frequently Asked Questions

Q: Why do AI assistants invent URLs on my domain?

A: They produce links the way they produce text: from patterns. When a URL structure appears across many of your pages or markets, a model can apply it to a page that does not follow it. The invented address looks plausible because it is built from your real ones, and the models even attach the real page's title to it.

Q: How can I check my own site for invented AI links without any tool?

A: Filter your server or CDN logs for requests that returned a 404 and whose address contains utm_source=chatgpt.com, the parameter ChatGPT adds to the links in its answers, and group them by path. This finds the links people clicked from ChatGPT. Other assistants do not always tag their links, and a link nobody has clicked yet leaves no trace.

Q: What is the fastest fix for an invented link?

A: A permanent (301) redirect from the invented address to the real page. When several invented links share a pattern, one wildcard redirect covers all of them. Check each target before you ship the rule.

Q: Can I stop ChatGPT from showing the wrong link?

A: Not directly, because nobody can edit an AI answer. The redirect makes the link work whatever the answer says, and consistent URL structures across your markets make the invented pattern less likely in the first place.

Q: What does Ayzeo URL Health check?

A: Every link to your own site in the AI answers Ayzeo tracks for your prompts: links in the answer text, pages the models visited and pages they consulted. It records whether each link opens, redirects or is broken, confirms broken links with two checks, suggests a possible fix and emails you a digest.

Q: Does a timeout or a firewall block count as a broken link?

A: No. Only a 404, a 410 or a soft 404 confirmed by two separate checks counts as broken. Timeouts, server errors and firewall blocks are shown as unreachable and are checked again later.

Q: Which plans include URL Health?

A: The Pro and Enterprise plans. It needs no setup: the first check starts within about an hour of your next scan, or the next morning (Berlin time) if the scan ends at night.

Key Takeaways

• AI invents links on your own domain: 18 invented URLs across 27 answers in three weeks, for one brand in one market, from ChatGPT, Gemini and Claude alike.
• The invented links follow your architecture: models copy URL structures from markets where they exist and apply them where they do not, down to the real page title.
• Five minutes in your logs finds ChatGPT's share: filter 404s for utm_source=chatgpt.com and group by path.
• One redirect rule usually fixes a whole pattern: point the invented pattern at the real pages, and verify each target first.
• URL Health keeps checking for you: every link to your own site in your tracked answers, broken links confirmed twice, a possible fix, and a digest by email.

AI visibility is not just about being mentioned. It is about what happens after the click.

If you check your logs and find these, we would be curious what your numbers look like: hello@ayzeo.com.

Sources

• Ayzeo platform data (September 2026): three weeks of stored AI answers for one tracked brand in one market, from ChatGPT, Gemini and Claude; 18 URLs on the brand's own domain that returned a 404, across 27 answers. The brand is not named, and the example addresses in this post are made up.
• Law, R. (2 September 2025). New Study: How Often Do AI Assistants Hallucinate Links? (16 Million URLs Studied). Ahrefs.
• OpenAI Help Center. Publishers and Developers: FAQ, on the utm_source=chatgpt.com parameter in ChatGPT links.