Claude stopped citing us in late September. Four other engines kept citing the same pages.

The twelfth GEO Glossary dispatch.

Since late September, one of the five AI engines we test every week has not cited us once; the other four kept citing the same, unchanged pages. Here is how we worked out where the loss happens, with checks that mostly cost nothing and work for any site.

What we measure

Every week we ask five AI assistants (ChatGPT, Perplexity, Claude, Gemini and Copilot) the same frozen question, "What is X? Cite the web sources you used with their URLs.", in their normal consumer apps, for terms from a fixed panel, and record whether our page is among the sources. Nine core terms are asked every week; the rest rotate. The wording never changes, so a change in the answers is a change in the engine or the web, not in our prompt. The methodology page has the details; Dispatch #4 explains why the five engines cite different sets.

The drop

Claude was a steady citer. Two terms, authoritative statement strength and citation match rate, were cited by Claude in every round from the first probe on June 3 to September 21.

Then it stopped.

The rounds mix different terms, so the steadier measure is the nine core terms: Claude cited at least two of them in every one of the 18 rounds from June 3 to September 21, and none in October (0 of 36).

The other four engines kept citing the same pages: authoritative statement strength, citation match rate and cite-ability were cited by other engines in every October round, and on October 19 by three or four engines each. The body text of these pages has not changed since August 26.

Our server logs show the same break from the other side. Claude has a fetcher, Claude-User, that opens pages while it writes an answer, in the Claude app and in Anthropic's coding tool alike. Through the summer it visited us 20 to 40 times a week, and on the September 21 probe day it opened exactly the pages we were asking about. Leaving out the coding tool's requests, it came on 58 of 90 summer days and on 7 of the 30 days from September 27. In none of the four October rounds did it open a page we were asking about. Anthropic's search crawler, Claude-SearchBot, which our robots.txt allows, has not visited us since our logs began in June, before the drop or after it.

Ruling ourselves out

The first suspicion when citations disappear is your own site. We tested that directly.

Then three checks that go further.

Anthropic's own search still finds us. On October 20 we ran the five key questions through the web search built into Claude Code, Anthropic's coding assistant, both as plain questions and as short keyword searches. Our page came back in 8 of the 10 result lists, in the top three on six. Claude's actual October sources for those terms showed up in one of the ten.

The Claude app's own search does not. Claude shows the searches it runs, and you can open them. On October 20, on three terms, it searched for the term as a quoted phrase and listed nine or ten results each time; our page was in none of the three lists. The same query strings, typed into the Claude Code search the same day, put our page in all three, and the two surfaces shared almost no results: 0, 0 and 1 URLs in common. Region does not explain it: asked from Australia, like the app, Brave's AI answer cited our page for all three on October 28. Brave is one of the search providers Anthropic lists; which one serves a given Claude answer is not public.

The words Claude searches with matter too. On October 26 Claude searched without quotes and in broader words. For cite-ability it searched "citability AI search what makes content citable", respelling the term and turning a question about a definition into a how-to question. For that exact string, Anthropic's developer search did not return our page either; how-to guides from SEO vendors and agencies filled both lists. Asked the same thing on October 28, Brave's AI answer searched "cite-ability citation metric AI search optimization", and for that wording Anthropic's developer search lists our entry sixth. On that service our place depends on the exact words, and the words Claude chose leave us out. (Claude's other two searches on October 26 were on topics led by official specifications, where no engine cited us.)

So the page is in the index, ranks well on other search surfaces, Anthropic's included, and opens on request. What changed is the Claude app's search step, in two places: the words it chooses to search with, and the results that come back for them. Why the app's lists differ from those of Anthropic's developer search, we cannot see from outside.

What Claude does now

On generic definitional questions, Claude now answers from its search results without opening pages, so only what is in that result list can be cited. Before late September it opened our pages.

The result lists look different from a mainstream engine's: of the 28 results in the three we opened on October 20, about a fifth sat on staging, test or odd subdomains, and several matched the quoted words in an unrelated sense, such as thesis-statement strength. Claude's finished answers cited none of those hosts; the gap is in what reaches the last step.

Some answers now treat our terms as undefined. In one October test Claude wrote "I couldn't find an established definition of 'authoritative statement strength'". The name is our entry's label for a method from a research paper, as the entry says, so "not established" is fair; what changed is that Claude used to find and cite our definition.

Some replacement sources cannot be opened by a reader. Five of the pages Claude cited on our terms in October, seven times across three rounds, are programmatic pages from one marketing vendor. We requested each page under the published user-agent names of a browser and of each AI crawler, which shows what the site serves to each name, not what the real crawlers receive. On October 20 the pages returned the article only to requests identifying as OpenAI's training crawler or Common Crawl's, and "Page not found" to the browser and to every answer-time fetcher we tried, Anthropic's included. On October 27, the day after Claude cited two of them again, all five returned "Page not found" to every identity we tried except Anthropic's training crawler, which received the vendor's homepage. These are the destination failures in Dispatch #9, turning up inside working answers.

What it means if this happens to you

• Track engines separately. A blended "AI visibility" score would have shown a dip. Per engine, it was one engine at zero and four unchanged.
• Split your logs by fetcher. AI companies run separate crawlers for training and for answering. If the answering fetcher (Claude-User, ChatGPT-User, Perplexity-User) stops visiting while the others continue, look at the engine, not at your site. Split by product too: Claude-User also covers Anthropic's coding tool.
• Run the three cheap tests. Ask with your brand in the question (are you in the index?), paste your URL (are you blocked?), and ask another engine (is it your content?). If all three pass, the cause is not on your side.
• Open the engine's search step, and read the query. If the engine reworded your term, that is where you were lost. If your page is missing from the result list for a faithful query, the loss happened before any choice among sources, and rewriting will not reach it. If it is listed but not cited, that is a different problem.
• Do not rewrite for one engine. Four engines kept citing our unchanged pages. Changing them to chase one engine's search would put the other four at risk.

What we do not know

We do not know why this changed, or whether it was meant to. We see answers, result lists and our own logs; Anthropic's search pipeline is not visible from outside, and nothing in its public release notes for the period mentions a change to how search finds or ranks sources. We do not know whether other sites lost Claude citations at the same time. Our data covers our pages on our questions.

Two other things changed for our account around September 27 and 28: a new model, which we tested, and Anthropic's merger of Claude's chat and Cowork modes. Anthropic's help article on the merger is dated September 16 and describes a rollout by plan. The merger removed the web-search switch (Claude now searches "when it helps"), so it touches the search step, where we lose the citation. The switch no longer exists and we have no account without the merged experience, so its part cannot be tested; the dates line up, which does not show that the merger caused the drop.

We also do not know whether it is permanent. On the day the fetches collapsed, Claude's training crawler swept about 45 of our pages, which fits a rebuild that might bring us back. Since then it has come and gone (absent from October 13 to 21, back for three days, absent again since October 24), and we attach no reading to that. A month after the drop, Claude still cites none of our pages. The weekly panel will show recovery if it comes, and we will update this page if it does.

One disclosure about our own site. The citation badges on our term pages only ever upgrade: a single "not cited" result is too noisy to remove a citation we have recorded. So 20 terms still show a Claude citation, while Claude has cited nothing of ours since September 21. The badges record that an engine has cited a page at least once, not that it still does.

Data: weekly probe rounds June 3 to October 26, 2026, server logs to October 27, and side checks run October 10, 20, 26, 27 and 28. Every figure traces to a dated round file or screenshot in our measurement archive.