What is citeturn0search0? The ChatGPT leftover that shows an answer was copied without its sources

citeturn0search0 is where a link to a source used to be. It is a citation marker from ChatGPT. Inside the app, that spot showed a clickable link to a web page the answer relied on. When someone copied the answer into a document, a LinkedIn post, or a report, the link did not travel and the code did.

So the string tells you one thing for certain: the sentence in front of it came from a ChatGPT answer that searched the web, and its source fell off on the way. It does not tell you which page was cited, whether that page said what the sentence says, or who wrote the rest of the document.

The fix is not to delete the code. The fix is to put the source back.

What do the parts of citeturn0search0 mean?

OpenAI publishes a guide to citation formatting for developers who build products on its models. It describes a marker made of a type name, a source ID, and an optional location inside the source. OpenAI calls its recommended markers “highly recommended because they closely match the markers our models are trained on.”

Read the string with that guide in hand:

Part	What it means
cite	The marker type: this is a citation.
turn0	The first tool call in the answer. A tool call is one action the model takes while answering, such as running a web search. Counting starts at zero.
search0	Source number zero from that search, in other words the first one. The next would be search1.

The guide also wraps each marker in three special characters: U+E200 to open, U+E202 to separate, U+E201 to close. They sit in Unicode’s Private Use Area. Unicode is the standard that gives every letter and symbol a number, and private-use characters are, in its words, code points “whose interpretation is not specified by a character encoding standard.” Readers normally cannot see them, which is why the visible text looks like one glued-together word.

One limit, stated plainly: that guide documents the format for developers. OpenAI does not publicly document how the ChatGPT app stores these markers or why they sometimes survive a copy. The best record of the leak comes from people who catalogue it.

Who has documented it, and where?

Wikipedia’s volunteer editors keep a field guide called Signs of AI writing. It lists citeturn0search0 under internal formatting bugs and explains it this way: “These are places where the chatbot links to an external site, but a human pasting the conversation into Wikipedia has that link converted into placeholder code.” The guide says the pattern was first observed in February 2025.

We opened one of the examples the guide points to: an edit made on 8 February 2025 to Wikipedia’s list of English-medium schools in Bangladesh. One pasted sentence reads “For more information, you can visit their official website:” and then, where the website should be, the marker.

We inspected the raw characters. Each of the four markers in that edit is built exactly as OpenAI’s guide describes: U+E200, then cite, then U+E202, then turn0search0 or turn0search1, then U+E201.

That sentence is the whole problem in miniature. It promises a source and delivers a code.

The complaint is older than this exact marker. In OpenAI’s community forum, a user reported in September 2024 that copying a search answer left placeholders such as 【9†source】 instead of links. No OpenAI staff member replied in that thread.

What does it prove, and what does it not?

What you will hear	What the evidence supports
”It is a watermark OpenAI adds to catch copying.”	OpenAI’s guide describes citation markup that software should parse and strip before a reader sees it. Nothing in it describes a watermark.
”It proves the whole text is AI-written.”	It shows that the passage around it came from a ChatGPT answer that used search. It says nothing about the rest of the document, or how much a person rewrote.
”The code tells you which website was used.”	The ID numbers a result inside one conversation. It is not a web address, and nothing outside that chat turns it back into one.
”Delete it and the text is fine.”	Deleting it hides the symptom. The claim still has no source.

Wikipedia’s guide makes the last point in bold: “Please do not merely treat these signs as the problems to be fixed; that could just make detection harder.”

Why is this a sources problem, not an AI problem?

Using ChatGPT to find information is not the issue. OpenAI’s own help page on web search says “Search results and citations can be incomplete, outdated, or incorrect.” It then says what a reader should do with a citation, which a lost link makes impossible: “Open a cited source to check that it supports the answer, review when it was published or updated, and use an authoritative source when accuracy matters.”

A surviving link would not prove the claim either. In a 2023 Stanford audit of four AI search products, only 74.5% of citations supported the sentence they were attached to. A link is where checking starts. That is the difference between a mention and a citation: a citation is something a reader can open and check.

The Verified Research Loop, one of Len’s books, names the habit behind the leftover. It calls it paste-as-research: “moving generated output into a memo, deck, brief, or conversation while leaving the claims inside it in an unrecorded evidence state.”

What do the other leftovers look like?

The same Wikipedia guide documents cousins of the marker. You may meet:

• citeturn0news0 or citeturn1file0: the same format, pointing at a news result or an uploaded file.
• iturn0image0turn0image1: a row of images that did not survive the paste.
• :contentReference[oaicite:0]{index=0} and oai_citation: other ChatGPT placeholders for links.
• ({"attribution":{"attributableIndex":"1009-1"}}): a block of code at the end of a sentence.
• [cite: 1]: the guide attributes this form to text copied from Google Gemini.
• ?utm_source=chatgpt.com at the end of a link. A UTM tag is a label on a link that tells analytics software where a visitor came from, and OpenAI’s publisher FAQ says ChatGPT adds this one automatically. This is the good case: the link survived.

People also search for chatgpt://url-citation?citation-number=0. We found no public documentation for that form. It is not a web address anyone else can open, so it leaves the reader in the same place: no source.

A regular expression is a search pattern that matches text by its shape rather than by exact words. This one finds all of the above. Paste it into any editor with regular-expression search (VS Code, for example) or run it with grep -E:

turn[0-9]+(search|news|file|image)[0-9]+|oaicite|oai_citation|attributableIndex|\[cite: ?[0-9]|chatgpt://|utm_source=chatgpt\.com

Each hit is a claim to check, not a verdict on the author.

What should you do when you find one?

If you wrote it. Do not just delete the code. Go back to the chat and select Sources, which OpenAI’s help page says shows “cited sources and other relevant links.” Open the page, check that it says what your sentence says, and paste the real link. If no page supports the sentence, cut the sentence or mark it as unchecked.

If you are editing, teaching, or reading it. Ask for the source, not a confession. “Which page supports this sentence?” is a fair question for any claim, written by a person or a model. It also gets a better answer than an accusation.

If you run a team. Keep a claim-to-source record for anything that leaves the building: the claim, the link, the passage, and who checked it on which date. This site calls that an evidence ledger. A leftover then becomes a row with an empty source cell, which is easier to fix than a rumour.

How does TRACE turn a pasted answer into a checked one?

TRACE is a five-step routine for turning a research question into a decision record in which every claim that matters can be traced to its source. It comes from The Verified Research Loop, which is available now in English and Dutch. You do not need the book to use it.

1. Target. Fix the question, the scope, and the output before a single search runs.
2. Retrieve. Gather the actual source documents, from more than one place, instead of trusting one confident answer.
3. Assess. Give each source a role: Primary (close to the event or the data), Secondary (reporting on it), Advocacy (made by someone with a stake), or Grey (everything else). Then check what that source can support for this one claim.
4. Cross-check. For each claim that carries the decision, record what every source actually shows about it, one row per source, and keep disagreement and dependence visible, instead of padding the memo with a token second link.
5. Export. Ship the memo with a claim ledger, so someone else can check the work without redoing it.

A hypothetical example. A marketing manager pastes a ChatGPT answer into a board memo: “Dutch e-bike sales grew strongly in 2025 citeturn0search0.”

• Target: the board is deciding whether to open a second store. The growth claim carries that decision; the paragraph on company history does not. Check the first.
• Retrieve: reopen the chat’s sources, or search again, and open the page itself rather than the summary.
• Assess: is it a national statistics office (Primary), a news story about those figures (Secondary), or a trade group that promotes e-bikes (Advocacy)? Does the passage cover the Netherlands, 2025, and sales, or something nearby, such as imports?
• Cross-check: note what each source says about the claim and whether one simply copied the other. If the statistics office is the only real source, record that dependence rather than adding a news story that repeats it.
• Export: the memo carries one line per claim: the claim, the link, the passage, and who checked it when.

The book is careful about its own limits. TRACE makes a research record inspectable; it has not been shown to produce better decisions. That is also the honest promise of this routine. The TRACE framework page has the one-page version.

Try this today (10 minutes)

1. Open the last three documents you shared that drew on an AI answer.
2. Run the pattern above over each one.
3. For every hit, find the page that supports the sentence, or cut the sentence.
4. Write down two numbers: hits found, sources restored. That pair is your team’s baseline.

Cite this:What is citeturn0search0? The ChatGPT leftover that shows an answer was copied without its sources.Len P. van der Hof. https://lenvanderhof.com/en/blog/what-is-citeturn0search0/ · Published 3 October 2026.