Do AI Engines Cite Pages That No Longer Exist? 3.3% of Pages Cited Between July and October Now Return 404 or 410, Against 0.4% Cited This Month

Most pages AI engines cite still load, and a small, steady share do not. In the Machine Relations Index answer ledger, we drew 1,600 cited pages, one per website, and fetched each on 2026-11-10. Of the 1,043 measurable pages whose most recent citation fell between July and October, 34 (3.3 percent, 95 percent interval 2.3 to 4.5) now return 404 or 410. Of the 271 measurable pages cited in the first nine days of November, 1 (0.4 percent, 0.1 to 2.1) does.

The rate does not climb with age. It sits between 2.9 and 3.8 percent for each of July, August, September and October, so a page last cited in July is about as likely to be gone as one last cited in October. Two groups never failed: none of the 54 pages cited by two or more engines, and none of the 32 pages cited six or more times, returned 404 or 410. Both groups are small.

This is a count of what exists today, not of what a citation caused. A page that dies cannot be cited again, so the recent group is partly filtered to pages that are alive. It is not a decay rate, and it does not show that a missing page cost anyone a citation.

What was counted #

• Population. Cited addresses in the Index's buying-question panel: successful runs from 2026-07-06 to 2026-11-09, five engines (ChatGPT, Claude, Gemini, Google AI Mode, Perplexity) through their APIs, runs that name a brand in the prompt excluded. This is the same 201,266 cited-URL events, 79,206 address strings and 23,729 runs as the page-counting study. It is not the consumer ChatGPT or Claude app, and Google AI Overviews are left out.
• Sample. Twelve social and open publishing hosts were removed (Reddit, YouTube, youtu.be, LinkedIn, Facebook, Instagram, X, Twitter, TikTok, Quora, Medium, Pinterest). One address was then drawn at random, with a fixed seed, from each remaining website, so a large publisher counts once, and the sample describes websites that get cited, not citations weighted by how often they are cited. Each address was filed by the date it was last cited, and 320 websites were drawn per group: July, August, September, October, and 1 to 9 November. That is 1,600 pages.
• Fetch. On 2026-11-10 around 23:15 UTC each address was requested once by plain HTTP with a desktop browser user agent, a 15 second limit, and up to five redirects followed. The ChatGPT tracking tag utm_source=openai was removed from the address first and fragments were dropped. Certificate errors were not treated as failures. The 39 pages that were not an ordinary load were requested a second time, and every one returned the same answer.
• Measured and unmeasured. A page counts as measured if the site answered with anything other than a block (401, 403, 429, 451), a server error (5xx), a timeout or a connection failure. 1,314 of 1,600 pages (82.1 percent) were measured. The other 286 (17.9 percent) were unmeasured, not zero: 174 returned 403, 47 failed to connect, 42 returned 429, 11 timed out and 12 returned a server error. A page we could not reach may be gone or alive, and a site that blocks automated reads is not a dead page.
• Gone. A page is gone if it returned 404 or 410 after redirects.

The result by when the page was last cited #

Of the 34 gone pages from July to October, 28 returned 404 and 6 returned 410, a deliberate "removed" answer. The one November page returned 404. Intervals are Wilson 95 percent intervals on the measured pages. The July to October group and the November group do not overlap.

Three further July to October pages loaded but were not what the engine cited: two redirected from a deep article to the website's homepage (the page is gone and the site hides it) and one returned 200 with the title "Article not found". Counting them, 37 of 1,043 (3.5 percent) were missing. The headline figure uses only 404 and 410 so that it needs no judgment call. The soft-error check also flagged a sports site's archive page whose title contained "oops" inside the site name; it was loaded and is counted as alive.

Who the gone pages are #

• Cited once, not often. 22 of the 689 measurable pages cited in only one run (3.2 percent) and 12 of 322 cited two to five times (3.7 percent) were gone. None of the 32 cited six or more times, and none of the 54 cited by more than one engine, were gone.
• Spread across engines, too few to rank. Among pages cited by one engine only (July to October): ChatGPT 9 of 216, Claude 4 of 145, Gemini 2 of 149, Google AI Mode 3 of 43, Perplexity 16 of 436. The intervals overlap, and the page claims no engine ranking.
• Ordinary content. The gone pages are company blog posts, product pages, a credit union PDF, a dental membership page and community answers. Nine of the 34 still carry the utm_source=openai tag in the ledger, which is how ChatGPT writes every address it cites, not a sign of a ChatGPT-specific problem.
• Redirects are common and were not classified. 78 of the 1,314 measured pages (5.9 percent) reached their content through at least one redirect. The study did not separate a trailing-slash or www. change from a move to a new address.

What this changes #

• Keep a cited address alive, or redirect it to its replacement. About one in thirty pages engines cited in the first four months is now a 404 or 410. A removed or renamed page that an engine has cited is a page the engine can no longer send a reader to.
• A redirect to the homepage is not a replacement. Two of the pages in the sample were removed that way, and a crawler or a person gets no answer to the question the page used to answer.
• Check before you count citations. If you track cited URLs, test them periodically. Roughly 3 percent of a month-old or older list will not load, and 18 percent of a list drawn from the open web will not answer an automated check at all.
• Citations are not a standing asset. Of the pages engines cited in the last nine days, 1 in 271 was already gone, and about 1 in 30 of those cited earlier is gone now. A citation list is a snapshot that goes stale. A dead cited page is not citation decay, which is a fall in an entity's observed citation rate between comparable measurement windows; this study does not measure that.

What this page does not show #

• A decay rate. Dead pages cannot be re-cited, so the November group is filtered toward live pages. The flat 2.9 to 3.8 percent across July to October may reflect that filter as much as the passage of time.
• That a page died after it was cited. We did not fetch the page when the engine answered. Some of these addresses may have been dead, or mistyped by the engine, at the time. The 0.4 percent in the newest group is the closest this study gets to that, and it is one page.
• Any effect on citations. The ledger cannot say that a 404 cost a citation, or that fixing one earns it back.
• The unreachable 18 percent. Sites that block automated reads may differ from the ones that do not (large publishers often block), so the measured share may lean toward smaller sites. The study did not test who the blocked sites are.
• One fetch. A page that was down for maintenance on 2026-11-10 would count as a server error (unmeasured), not gone. A 404 caused by our user agent or fetch location is possible; the second request repeated every result from the same machine.
• The consumer apps and Google AI Overviews. Not included.

Method note #

Source: the Index answer ledger's citation-events table, daily export of 2026-11-10 08:10 UTC, read with DuckDB and joined to the run table. Filters: the buying-question panel, runs not flagged as brand-prompted, successful runs, a cited URL beginning with http, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity, run dates 2026-07-06 to 2026-11-09 (201,266 distinct (run, cited URL) events, 23,729 runs, 79,206 address strings, reproduced exactly before sampling). Twelve social hosts were removed by root domain. Each remaining address received its most recent run date; one address per root domain was drawn at random (DuckDB random(), seed 0.42), then 320 per group, groups defined by that date: before 2026-08-01, August, September, October, and 2026-11-01 to 2026-11-09. Fetches used Python urllib with a Chrome user agent, 24 parallel workers, a 15 second timeout, manual redirect following (up to six requests), and the first 150 KB of each body. Statuses: 404 and 410 are gone; 401, 403, 429, 451, any 5xx, timeouts and connection errors are unmeasured; 200 to 299 with HTML, PDF or text is loaded, with a flag for any redirect, a homepage landing from a deeper address, or an error-like title. Five pages returned 402 or an unfollowed 308 and count as measured and not gone. Intervals are Wilson, 95 percent. Pages cited by multiple engines and by six or more runs are counted from the ledger events for the drawn address. Nothing was written to the pages fetched. Rerun this on the then-current ledger on or after 2026-11-22 and compare the November group.

Frequently asked questions #

Do AI engines cite pages that no longer exist? #

Rarely. In the Machine Relations Index ledger, 1 of 271 measurable pages cited in the first nine days of November returned 404 or 410 on 2026-11-10 (0.4 percent). Of 1,043 measurable pages last cited between July and October, 34 (3.3 percent) did. Pages that blocked automated reads, 17.9 percent of the 1,600 sampled, could not be tested.

Does the share of dead cited pages grow over time? #

Not in this sample. It was 3.8, 2.9, 3.4 and 3.0 percent for pages last cited in July, August, September and October. Because a dead page cannot be cited again, the study does not treat that as a decay rate.

What should I do when a page an AI engine cites is removed? #

Redirect the address to the page that now answers the same question, or return 410 if it is gone for good. Two sampled pages were redirected to their homepage, which leaves a reader or crawler with no answer to the question the page used to hold. The study did not test whether any of this changes citation.

Is a 403 or 429 a dead page? #

No. They are blocks. 174 of the 1,600 sampled pages returned 403 and 42 returned 429, and they are counted as unmeasured, not as gone and not as alive.

Which engine cites the most dead pages? #

The study cannot say. Among pages cited by a single engine from July to October, Google AI Mode had 3 of 43, ChatGPT 9 of 216, Perplexity 16 of 436, Claude 4 of 145 and Gemini 2 of 149. The counts are too small and the intervals overlap.