Guide

Why ChatGPT Invents Page Numbers From Your PDFs (and How to Stop It)

A 4-minute read

You ask ChatGPT or Claude to back up a claim with a citation, and it hands you a confident-looking "p. 142." You go to check it — and page 142 says something else entirely, or the number doesn't exist. If you have to stand behind your references, that's not a quirk. It's a dealbreaker.

Here's the reassuring part: the model usually isn't lying on purpose. It just can't see a real page number to give you — so it fills the gap with a plausible one. Fix that, and the made-up citations mostly stop. Here's why it happens and exactly how to fix it.

Why it makes them up

Two things are usually going on:

• It only knows the file's position, not the book's page. When you upload a raw PDF, the model often just knows "this is the 34th page of the file." The moment you trim or combine sources, that number stops matching the printed page — so "page 34" is meaningless.
• You stripped the real numbers out. If you converted the PDF to plain text or Markdown first, the page numbers printed on each page are gone. Asked to cite anyway, the model invents one. That's the hallucination.

The key insight: the real number is right there on the page

Claude and ChatGPT are multimodal — they read the page image, not just extracted text. And on almost every book, paper, or scan, the real page number is printed in the header or footer. If you keep the document as a PDF instead of converting it, the model can literally see "88" sitting in the corner and cite that. You just have to tell it to look.

The instruction that fixes it

Paste something like this alongside your PDF:

Use the attached PDF as the source. Cite by the page number printed on each page (header or footer), not the page's position in the file. If a page has no printed number, cite the section and say so. Never guess a page number, and if a claim isn't supported by the source, say "not in the source."

Each line is doing a job:

• Cite the printed number, not the file position — kills the useless "page 34 of the file."
• No printed number? Cite the section and say so — gives the model an honest fallback instead of a reason to invent.
• Never guess; say when it's not in the source — this is the line that actually stops the hallucination. You're giving it permission to say "I don't know," which models rarely do unless asked.

Now you can verify in ten seconds

Because the number is printed on the page, every citation becomes checkable. The model cites p. 88? Open page 88 and look at the corner — it should say 88, and it should say what the model claims. A made-up citation falls apart the instant you look. That quick check is the whole point: not "trust the AI," but "make the AI cite something you can trust."

In one line

Keep your sources as PDFs so the printed page numbers survive, and tell the model to read them and never guess. Do that, and "the AI made up a citation" turns into "the AI cited page 88 — and page 88 checks out."