How Often Do AI Assistants Link to Pages That Do Not Exist? We Checked 26,962 Links

Short answer: Of 26,962 checked links that AI assistants gave to the 74 brand websites in this study, 466 lead to a page that does not exist. That is 1.7%, and each one was confirmed by two separate checks. Over the 60 days to October 12, 2026, ChatGPT shows the highest rate at 2.0%, about one broken link in fifty. Claude (1.2%) and Google AI Overviews (1.1%) follow, and Perplexity and Gemini stay near zero. Document links are the weak point: 6.3% of PDF links are broken, eight times the rate of normal pages. Broken links also repeat: more than half of those shown in answers appeared in more than one answer, and the most-repeated one appeared in 629 answers. Every broken link we traced in the public web archive was once a real page. The assistants keep citing addresses that a site restructure removed months ago, in three country stores at once. The rest look almost right: they follow the site's URL patterns, and many sit one step away from a page that exists.

1. The Findings at a Glance

• 1.7% of checked links lead nowhere. 466 of 26,962 links to brand websites, confirmed broken by two separate checks at least 15 minutes apart.
• ChatGPT shows the most broken links: 2.0% of its links to brand sites over the 60 days to October 12, 2026, about one in fifty. Claude follows at 1.2% and Google AI Overviews at 1.1%. Perplexity (0.2%) and Gemini (0.1%) barely produce any.
• AI remembers your old website. Every broken link we traced in the public web archive was a real page before a restructure. The pages went away in late 2025 or early 2026, yet the old addresses still show up in answers in October 2026, in three country stores at once.
• The rest look almost right. They follow the site's usual URL patterns, and many sit one step away from a page that really exists. We could not find any past page behind them.
• PDF links break eight times as often as pages: 6.3% against 0.8%. Datasheets, catalogues and guides are where the broken links concentrate.
• Broken links repeat. Of the broken links shown in answers, 57% appeared in more than one answer. On average a broken link was shown in almost 11 answers; the most-repeated one appeared in 629.
• An early signal: links written without a live web search look riskier. For Claude, 6 of 36 such links were broken against 0 of 19 in answers where it searched the web. The numbers are small and cover one assistant, but they point the same way as the per-assistant table.

2. What We Measured

Ayzeo monitors how AI assistants answer the questions buyers ask, for the brands tracked on the platform. Every answer that includes a link to the brand's own website leaves a record: the exact address, the assistant, and the date. Since September 24, 2026, our link checker has been opening each of those addresses on the live site and recording what it finds.

The link check does not run for every brand on the platform, and the set grows as the check reaches more brands; at the study cutoff it stood at 74 brand websites. This study covers that segment, with every such link stored up to October 12, 2026:

• 45,483 distinct addresses on those 74 brand websites, collected from AI answers produced by ChatGPT, Claude, Gemini, Perplexity, Grok and Google AI Overviews.
• 26,962 of them checked with a clear result so far: the page opens, redirects, or is missing. The remaining 18,521 are still queued or answered with a timeout or a firewall block, which we never count as broken; the rates can move a little as those resolve.
• 466 confirmed broken: a 404, a 410, or a soft 404 (a page that answers 200 but behaves like a missing one), each confirmed by two separate checks at least 15 minutes apart.

The brands range from global hardware makers to single-market software companies. The example addresses in this post all point at lenovo.com, one of the world's most-discussed hardware sites. It is large and multilingual, which makes the patterns easy to see. Every example is a public address that anyone can verify, and each one was re-tested on the live site on October 12, 2026. The full rules are in the methodology.

3. One Broken Link in Fifty, for the Most-Used Assistant

The headline table. For links seen in answers over the 60 days to October 12, 2026:

The most-used assistant also shows the highest share of broken links. Perplexity and Gemini sit at the other end. They mostly link to pages they have just read, and a link copied from a live page rarely points at a missing one, which fits their near-zero rates. The gap between the assistants is one of the clearest results in the data.

Ahrefs found the same direction in 2025, measured from the click side: across 16 million URLs cited by AI assistants, visitors were sent to 404 pages 2.87 times as often as from Google Search, with ChatGPT the most frequent offender at 1.01% of clicked URLs. Our study is narrower in scope and stricter in checking: it covers every link to the tracked brands' own sites that appeared in an answer, clicked or not, and checks each one twice.

4. AI Remembers Your Old Website

Where do the broken links come from? For the example site's most-cited broken links, the history is checkable. These links were once real pages. The assistants remember a version of the site that no longer exists.

ChatGPT answers about Lenovo's Premier Support service linked to these addresses, in answers about three different markets:

• www.lenovo.com/gb/en/premier-support → 404
• www.lenovo.com/fr/fr/premier-support → 404
• www.lenovo.com/de/de/premier-support → 404

These were real pages, with archive captures going back years. The public web archive shows all three answering as normal pages into late 2025, and the UK one as late as February 2026. Then the site restructured. Today's equivalent page in the UK store, Premier Support Plus, lives at www.lenovo.com/gb/en/services/support-services/premier-support-plus/, and the old addresses answer with a 404 instead of a redirect. Roughly eight months later, the assistants have not caught up: the three retired addresses appeared in the text of 59 stored answers, the latest on October 10, 2026.

The same pattern holds for the document links. news.lenovo.com/wp-content/uploads/2025/09/ThinkPad_P1_Gen_8_Datasheet_Final.pdf and ...ThinkPad_P16_Gen_3_Datasheet_Final.pdf were real datasheets, archived as live PDF files in September and December 2025. Today both are gone. They still appeared in the text of 28 and 46 answers respectively, the latest two days before this study's cutoff.

This changes what a site restructure costs. Your crawlers, your SEO tools and Google move on within weeks. AI assistants update their knowledge of your site much more slowly, and every answer built on the old version sends a buyer to a retired address, for as long as that lag lasts.

5. The Almost-Right Addresses

Not every broken link is an old page. Some match no past page we could trace, and many sit one step away from a page that exists. What they share: they follow the site's URL naming patterns so closely that they look real.

Claude, asked about Lenovo's outlet stores, linked to lenovo.com/de/outletde, lenovo.com/fr/outletfr and lenovo.com/gb/outletgb. The outlet stores are real, but they live one step deeper: www.lenovo.com/de/outletde/de/ opens, while the shorter form ends in a 404, and nothing redirects it to the store. Maybe the assistant shortened a real address, maybe it repeated a form that never worked; the public archive shows the short German form already answering 404 in October 2025. Either way, in three markets a reader lands on a missing page, one step from a working store. The same pattern appears on one brand's partner sites: dozens of broken document links carry one extra folder level, while the same file opens without it.

The catalogue link promotions.lenovo.com/uploads/market/e-catalogue-2026.pdf shows the same thing in a file name. The 2024 edition opens at exactly this kind of address. The 2026 one is broken, and we could not tell whether it ever existed.

Both stories, the remembered addresses and the almost-right ones, come down to one mechanism. AI assistants write a URL the same way they write the rest of the answer: from patterns. Nothing in that process checks whether the page still exists. Because a site's addresses follow consistent rules, the assistants learn those rules and write addresses that fit them, whether the page is live, moved, or never existed. An address that only looks right sounds exactly as confident as a real one. The breaks show up wherever your site moved pages without redirecting the old addresses.

6. Document Links Break Eight Times as Often

Across all links with a working-or-broken result, split by what the address points at:

Six in every hundred PDF links that AI assistants produced for these brands lead to a missing file. Two things plausibly add up here. A document that moves is seldom given a redirect from its old address, and document names follow strong conventions, so a broken document address reads exactly like a real one. For brands whose buyers ask technical questions, spec sheets, catalogues and guides, this is where the broken links concentrate.

7. A Broken Link Is Never Shown Once

A broken link in an AI answer would matter less if it were a one-off slip. It is not.

• 460 of the 466 broken links appeared in answers to questions we still track. Of those, 262 (57%) appeared in more than one answer.
• On average, a broken link was shown in almost 11 answers.
• The most-repeated broken link appeared in 629 separate answers.

This is what makes the lag expensive: once an address is in the assistants' picture of your site, the same address keeps coming back, answer after answer, for every person who asks a similar question. And the reader who receives it is not browsing. They asked which product to buy or where to get support, the assistant named the brand, and the link arrived inside that recommendation. A 404 at that exact moment wastes trust the brand has just earned.

One more early signal, stated with its sample sizes. When Claude wrote a link into an answer without running a live web search, the link was more often broken: 6 of 36 such links. Where it searched the web first, 0 of 19 were broken. Samples this small cannot settle the question, but they point the same way as the per-assistant table.

8. What Brands Should Do

Nobody can edit an AI answer. But every finding above has a practical response.

• Treat redirects as part of any restructure, and keep them for years. The retired addresses above were still filling answers months after the pages went away. A permanent (301) redirect from each old address to its new page sends every reader of those answers to the live page instead of a dead end, for as long as the assistants keep citing it.
• Find the broken links AI is already showing for you. The quickest free check: filter your server or CDN logs for 404 responses whose address contains utm_source=chatgpt.com, the parameter ChatGPT adds to the links in its answers, and group the results by path. This finds only the ChatGPT links people clicked. It misses the links nobody has clicked yet, and the assistants that do not tag their links.
• Redirect the almost-right forms too. When the real page differs from the shown address by one folder level, as with the outlet stores above, one redirect rule per pattern covers every variant. Check every target first: the redirect should land on the page the answer promised.
• Treat documents like pages. PDFs broke eight times as often in this study. When a datasheet or catalogue moves, redirect its old address the way you would a product page.
• Keep a useful 404 page. A search box and links to the main categories help the visitors no redirect covers yet.

For the brands tracked on Ayzeo, this check now runs continuously. URL Health collects every link to your own site that shows up in your tracked AI answers and checks each one on the live site. It confirms broken links with two checks and suggests a possible fix: for 99 of the 466 broken links in this study, the checker found a working page on the same site that matches the broken address. It then emails a digest. URL Health is included in the Pro and Enterprise plans, with nothing to set up.

9. Methodology

For readers who want to check our work, these are the rules behind the numbers above:

• Population: every distinct link to a tracked brand's own website that appeared in the stored AI answers of the study set. The study set is the brand websites where our link check was active during the study window, a segment of the platform, not every brand tracked on it; at the study cutoff it stood at 74 brand websites. A brand's own website means its domain, its subdomains, and other domains of the same brand. We count links in the answer text, pages the assistant visited, and pages it consulted; links to third-party sites are out of scope. That gives 45,483 distinct addresses, from answers stored up to October 12, 2026.
• Checking: each address is opened on the live site by our link checker (AyzeoLinkCheck), which identifies itself, paces its requests and follows robots.txt, the file where a site states what automated tools may request. Redirects are followed to their final page. Checks ran from September 24 to October 12, 2026.
• Broken means confirmed broken: a 404, a 410, or a soft 404 (a page that answers 200 but behaves like a missing one, judged against how the site answers made-up addresses), confirmed by two separate checks at least 15 minutes apart. Timeouts, server errors, firewall blocks and robots.txt refusals are recorded as unreachable and never counted as broken. At the study cutoff, 26,962 links had a clear result, including 11 awaiting their second confirmation; every percentage uses those 26,962 as its base, and the PDF-versus-page table uses only the links that are either working or broken.
• Per-assistant rates count links seen in answers during the 60 days to October 12, 2026, so every assistant is measured on the same window. A link shown by two assistants counts for both, so the rows do not add up to the overall totals.
• Grok appears in the data with 36 checked links and none broken, too few to rate, so it is left out of the table.
• How the examples were labelled: a broken link does not automatically mean the address was never real, and we do not claim to know any single address's full history. Where the public Internet Archive or the site itself shows the page once existed, we label the link a moved or retired page; every example we traced turned out to be of that kind. A link that matches no capture we could find, or that sits one folder level from a working page, is presented as almost right; we do not claim it never existed. Some broken links fit neither description cleanly and stay unclassified. These are links to brands' own sites in brand-related answers, not a sample of all links in all AI answers.

This is the first Ayzeo Research study, built on the data the platform collects for its dashboards. If you are a journalist or researcher and want a specific cut of the data, write to hello@ayzeo.com.

10. Frequently Asked Questions

Q: How often do AI assistants link to pages that do not exist?

A: In this study, 1.7% of 26,962 checked links to brand websites led to a missing page, each confirmed by two separate checks. By assistant, measured over the 60 days to October 12, 2026, ChatGPT showed the highest rate at 2.0%, about one broken link in fifty, followed by Claude at 1.2% and Google AI Overviews at 1.1%. Perplexity and Gemini stayed near zero.

Q: Are broken AI links removed pages or invented addresses?

A: Every broken link we traced in the public web archive was once a real page. The assistants keep citing addresses a restructure moved or retired, months after the pages went away and in several country stores at once. The rest match no past page we could find. They follow the site's URL naming patterns and often sit one step away from a page that exists.

Q: Why do AI assistants show links that do not work?

A: They write links the same way they write text: from patterns. Nothing in the writing of an answer checks that the page still exists. So retired addresses keep appearing long after a site restructure, and an address that merely fits the site's naming patterns can get through without matching any page.

Q: Which links break most often?

A: Document links. 6.3% of PDF links in the study were broken, against 0.8% of normal pages, roughly eight times the rate. A document that moves is seldom given a redirect from its old address, and document names follow strong conventions, so a broken document address reads exactly like a real one.

Q: Does a broken link in an AI answer matter if nobody complains about it?

A: Yes, because it repeats. 57% of the broken links shown in answers appeared in more than one answer, a broken link was shown in almost 11 answers on average, and the most-repeated one appeared in 629. Each reader it reaches has just been told that this page is the one to open.

Q: How can a brand find and fix these links?

A: Filter server or CDN logs for 404 responses whose address contains utm_source=chatgpt.com to find the links people clicked from ChatGPT. Then set a permanent (301) redirect from each broken address to the real page; one wildcard rule often covers a whole pattern of broken addresses. Ayzeo URL Health runs the check automatically for every link to your own site that shows up in your tracked AI answers, confirms broken links twice, suggests a possible fix and emails a digest. It is included in the Pro and Enterprise plans.

Key Takeaways

• 1.7% of AI links to brand sites lead nowhere: 466 confirmed broken links out of 26,962 checked, across 74 brand websites. Over the 60 days to October 12, 2026, ChatGPT's rate was 2.0%, about one in fifty.
• AI remembers your old website: every broken link we traced was a real page before a restructure. The pages were live until late 2025 or early 2026, and their old addresses still show up in answers in October 2026.
• Others are almost right: addresses that follow the site's URL patterns, often one step from a page that exists, with no past page behind them that we could trace.
• PDFs are the weak point: 6.3% broken against 0.8% for pages. Redirect moved documents like moved pages.
• Broken links repeat: 57% of those shown in answers appeared in more than one, and the most-repeated one appeared in 629 answers. Every repeat sends a buyer to a dead end with your brand's name on it.
• Redirects are the main fix: a 301 from each retired or almost-right address, kept for years, turns the answers that still cite it into traffic instead of dead ends.

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

Sources

• Ayzeo Research (October 2026): 45,483 distinct links to the 74 brand websites in the study set, collected from AI answers stored on the Ayzeo platform; 26,962 checked with a clear result between September 24 and October 12, 2026; 466 confirmed broken. Methodology above. Example addresses re-tested live on October 12, 2026, and their history checked against the Internet Archive's Wayback Machine.
• 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.
• Internet Archive Wayback Machine captures of the example addresses, including the retired Premier Support pages (last archived answering between November 2025 and February 2026, depending on the store) and the ThinkPad datasheets (archived as live files in September and December 2025).