Don't stop here
Hand-picked guides our readers explore right after this one.
Master ChatGPT with advanced prompting techniques, mega-prompts, and proven frameworks
Read the guideUnlock Google's Gemini with multimodal prompting strategies
Read the guidePrompts for building and using autonomous AI agents for research, coding, data analysis, and automation
Read the guideThe upload succeeds, the file appears in the chat, and then ChatGPT either says it cannot access the document or confidently answers about something that is not in it. Almost always the cause is one of three things. The PDF is a scan, meaning it contains page images with no text layer, and ChatGPT does not run full OCR on uploaded PDFs by default. The PDF is far too long, so only the portion that fits the context window is actually read and the rest is silently ignored. Or the file is protected, corrupted, or built by an exporter that produces a broken text layer. The fix is almost never to re-upload the same file harder. It is to check whether the PDF has selectable text, convert it if it does not, and cut it down to the pages that matter. This page covers how to tell which case you are in within about ten seconds, and what to do about each. If your uploads fail before they finish at all, see the ChatGPT file upload guide instead, since that is a different failure.
ChatGPT says it cannot access or read the uploaded document
Answers are generic and clearly not based on the file contents
Only the first few pages are summarized and later sections are ignored
Details are invented, with page numbers or quotes that do not exist
The same PDF works after you paste the text manually
Tables and multi-column layouts come back scrambled or out of order
A scanned or photographed document is a stack of images inside a PDF wrapper. There is no text to extract, so the tool that reads the file returns nothing. This is by far the most common cause and it affects most documents produced by office scanners, fax gateways, and phone scanning apps.
A long PDF may upload fine but only the part that fits into the model's context is available. Content past that point is not read at all, and the model does not reliably tell you where it stopped, which is what produces confident answers about pages it never saw.
Encrypted files and documents with content-extraction permissions disabled cannot be parsed. The upload itself succeeds because the container is valid, but the text extraction step is blocked.
Some exporters (older LaTeX pipelines, certain CAD and accounting tools, badly subset fonts) produce a text layer that extracts as garbage characters or in the wrong reading order. Multi-column academic papers and dense tables commonly scramble even when extraction technically works.
Free accounts get a small number of file uploads per day. Once you exhaust it, further uploads fail or the file is not analyzed, which can look identical to a parsing failure.
When to try: First, before trying anything else
Open the PDF in any viewer and try to select a sentence with your cursor. If nothing highlights, there is no text layer and ChatGPT has nothing to read. Ctrl+F (Cmd+F) for a word you can see on the page is the same test: no match means no text.
When to try: When the selection test shows no text layer
In Adobe Acrobat use Tools, then Scan and OCR, then Recognize Text, and save a new copy. Free alternatives: upload the PDF to Google Drive, right click, Open with Google Docs (Drive runs OCR automatically), then download as PDF or copy the text. Upload the OCR'd version, not the original.
When to try: For anything over roughly 50 pages
Extract the chapter, section, or page range you actually need into a separate PDF and upload that. A focused 20 page extract gets read completely and answered accurately, while a 400 page upload gets read partially and answered confidently regardless.
When to try: Every time, as a verification habit
Start with: 'Before answering, quote the first sentence on page 1 and the first sentence of the final page, verbatim.' If it cannot, it did not read the file, and you have caught the failure immediately instead of three answers later.
When to try: When the file came from a bank, government portal, or legal service
Open the file with the password, then print to PDF or export a new copy without encryption. Check the document properties for security settings that disable content copying, since those block extraction even when you can read the file yourself.
When to try: When answers show jumbled or interleaved sentences
For multi-column papers and heavy tables, export or copy the content into a .txt or .md file and upload that instead. You lose the visual layout, which is exactly the thing that was confusing the extraction, and you keep the words in the right order.
When to try: When the content is short and accuracy matters most
For a few pages, skip the upload entirely and paste the text into the message. This bypasses every parsing step and is the most reliable path for short but important content such as a contract clause or a results table.
When to try: For one or two pages you need read right now
Vision handles page images directly, so a screenshot of a single scanned page often works when the PDF does not. This does not scale to long documents, but it is a fast workaround for one chart, one form, or one page of a scan.
Save documents as text-based PDFs rather than scanning printed copies whenever you have the choice
Keep an OCR step in your workflow for anything that arrives from a scanner or fax
Split reference documents into sections before uploading rather than after a bad answer
Always make the model quote from the file before you trust a summary of it
Contact OpenAI support through help.openai.com if a normal text-based PDF (one where you can select and copy text) consistently fails to be read across multiple new chats, browsers, and devices, or if uploads fail while you are well inside your plan's allowance. Include the file size, page count, how the PDF was produced, and a screenshot of the response. Note that a scanned document being unreadable is expected behavior rather than a bug, so run the selection test before reporting.
Uploading and reading are separate steps. The upload only checks that the file is a valid PDF. Reading requires an extractable text layer, and a scanned document has none, so extraction returns nothing even though the file itself is perfectly fine. Try selecting text in the PDF: if you cannot highlight a sentence, ChatGPT cannot read it either.
Not reliably as part of normal PDF handling. Image-only PDFs generally come back empty rather than being transcribed. Run OCR yourself first (Adobe Acrobat's Recognize Text, or open the file in Google Docs via Google Drive, which OCRs automatically), then upload the searchable copy. Uploading a single page as an image instead often works, because images go through vision.
Only the portion that fits the context window is read. Everything past that point is not available to the model, and it will not necessarily tell you where it stopped, which is why answers about later chapters can be entirely invented. Split the document and upload the specific section you need.
Ask it to quote verbatim from specific locations: the first sentence of the document, a line from the middle, and the last sentence. Then check those quotes against the file yourself. If the quotes are wrong or it deflects, it did not read the content and any summary you get is unreliable.
Plain text or Markdown, because there is no layout to misread and nothing to extract incorrectly. For structured data use CSV rather than a PDF table. DOCX also extracts cleanly. Reserve PDF for cases where the visual layout genuinely matters, and make sure it has a real text layer.
Product behavior and limits can change. These primary sources were used to verify this guide.