You asked a question, it went and looked it up, and now every second sentence has a little bracket hanging off it. Paste the answer and get the prose without the footnotes.
Ask a chatbot something it has to look up and it comes back having read six pages, with a marker after every claim saying which one. On screen those markers are neat little chips you can click. Copy the answer out and they turn back into what they always were: raw characters wedged into the middle of your sentences.
It is the same problem as the em dash, one step up. The chat window was rendering something, you took the text out from under the rendering, and now you can see the scaffolding.
ChatGPT Search writes its sources as 【1†source】 — two CJK brackets around a dagger. No language, no style guide and no programming syntax uses that combination for anything, so there is nothing to weigh up and nothing to ask you about. It simply goes, every time, and you will not find a switch for it.
A numbered reference is the opposite problem. [1] in a paper or a technical note is something the author wrote on purpose, and it is also how Perplexity and Gemini mark their sources. Nothing in the text tells those apart, so that one is a question rather than a setting, and the answer starts at leave them.
Deleting a citation throws away the one genuinely useful thing about it. So there is a third option: read the list of sources at the bottom, match each number to its address, and put the address inline where the marker was.
A claim that read revenue rose 12%[1] comes back as revenue rose 12% (https://…), and the list at the bottom goes. You keep the receipts and lose the footnotes, which is usually what you wanted when you started deleting them by hand.
Because nothing in the text tells a citation apart from a reference somebody meant to write. In a paper or a technical note, [1] is the author's own footnote. It is also how Perplexity and Gemini mark their sources. The bracket form ChatGPT Search uses is different - CJK brackets around a dagger, which no language or syntax uses for anything else - so that one is removed without asking and this one is left to you.
It reads the list of sources at the bottom, matches each number to its URL, and puts the URL inline where the marker was. A claim that read "revenue rose 12%[1]" comes back as "revenue rose 12% (https://...)", and the list at the bottom goes with it. A number with no address listed is simply removed, and a run like [1][2] is removed too, because there is no single address for it to become.
Only what looks like a machine wrote it. The source list is removed when a Sources or References heading has something list-shaped underneath it; a heading followed by ordinary prose keeps the prose and only loses nothing at all. And with numbered references left alone, which is where the tool starts, a hand-written bibliography is untouched.
That is what removing numbered references does, and it is why the tool does not start there. An array index and a citation are the same three characters, and no tool can tell them apart by looking. If your text is part prose and part code, clean the prose.
You can, and for two or three it is faster. The reason this exists is the researched answer with thirty of them, where a find-and-replace leaves double spaces in front of every full stop and you spend longer fixing that than you saved.