Uploading a PDF to ChatGPT is easy; getting a complete analysis takes more care. Important pages, footnotes, scanned tables or diagrams can receive little attention. The safest workflow is to map the document first, analyse it in sections and verify where each conclusion came from.
Before uploading: Identify what kind of PDF you have
Try selecting a sentence in the PDF. Selectable words indicate digital text; a page that behaves like one photograph is probably a scan or image-only PDF.
| PDF type | What ChatGPT can usually access | Recommended preparation |
|---|---|---|
| Text-based PDF | Headings, paragraphs and other extracted text | Upload it directly, then verify the document map |
| Scanned PDF | Text extraction may be incomplete or unreliable | Run OCR first or upload important pages as separate images |
| PDF with charts or diagrams | Embedded visuals depend on the plan and upload context | Upload crucial charts separately as images unless visual retrieval is confirmed |
| Complex table-heavy PDF | Layout and exact values may be misread | Use the original spreadsheet when available and check totals manually |
| Very long PDF | Relevant sections may be retrieved instead of every line being considered together | Split by chapter or ask focused, section-by-section questions |
How to upload a PDF to ChatGPT
Open a chat or Project, select the attachment button, choose the PDF from your device, Library or a connected source, and wait for its filename to appear. OpenAI currently specifies a 512MB hard limit per file and a two-million-token limit for text and document files. A PDF can still be too complex for reliable one-pass analysis below those limits.
Step 1: Ask ChatGPT to map the document
First ask what ChatGPT can see: the title, page range, contents, headings, appendices and gaps in extracted text. It should flag pages that appear scanned, complex or unreadable.
Before analyzing this PDF, create a document map. List its title, apparent page range, major sections, appendices, tables and figures. Identify any pages or elements you cannot reliably read. Do not summarise yet, and do not claim that the whole file is readable unless you have verified its structure.
Compare the map with the PDF. If a 120-page document produces only three recognized chapters, address the gap before trusting a summary.
Step 2: Analyze one section at a time
Instead of “Summarise this PDF,” divide the work by chapter, heading or page range. For each section, request claims, evidence, important numbers, assumptions and limitations.
Use a prompt such as:
Analyze pages [X–Y] or the section titled “[name]”. Produce five parts: key claims, evidence, important numbers, assumptions, and unanswered questions. Give a page or section reference for every factual point. If a reference cannot be verified, label it clearly instead of guessing.
After covering every section, request a combined synthesis. This prevents one retrieved passage from representing the whole document.
Step 3: Build an evidence table
For research papers, contracts, policies and reports, request an evidence table with the claim, supporting passage, page, section, confidence and conflicting evidence.
Open several cited pages yourself. References can be wrong, especially when printed page numbers differ from the viewer counter. Our guide to making ChatGPT fact-check answers and provide reliable sources provides a verification workflow.
What happens to charts, images and scanned pages?
OpenAI says most plans use text-based document retrieval: digital text is extracted while embedded images are discarded. Visual retrieval for images, graphs and diagrams inside PDFs is currently documented for ChatGPT Enterprise.
For Enterprise, a PDF uploaded during a conversation can use visual retrieval, including in a Project conversation. PDFs stored as Project Files or GPT Knowledge use text-only retrieval.
If an essential chart is missed, export its page as a clear image and attach it separately. Ask about the axes, units, legend and values. For scans, run OCR first and compare the extracted text with the original.
How to analyse tables without losing exact values
OpenAI warns that exact values may be misread in scanned tables, image-based tables or complex layouts. Upload the original Excel or CSV file when available.
Otherwise, request a row-by-row transcription before calculations. Check sample values, units, signs, percentages and totals, then ask for the formula or code and assumptions.
Why ChatGPT may miss information
- The prompt is too broad: A general summary encourages prioritisation rather than complete coverage.
- The file is image-heavy: Important content may exist only inside scans, charts or diagrams.
- The document is too long or complex: Only the most relevant retrieved sections may shape an answer.
- Several files are mixed together: The model may not give every document equal attention.
- Page numbering is inconsistent: Printed page numbers and viewer page numbers may not match.
- The source itself is ambiguous: Conflicting wording, poor OCR or missing appendices cannot be fixed by a confident prompt.
A reusable prompt for complete PDF analysis
Analyze this PDF for [purpose and audience]. First map the full document and identify unreadable pages, missing text, scans, tables and figures. Then work section by section. For each section, list its claims, evidence, numbers, assumptions and limitations with page references. Separate direct evidence from your interpretation. Finish with an evidence table, contradictions, unanswered questions and a checklist showing which sections were covered. Do not invent missing content or claim complete coverage unless the document map supports it.
Privacy and file storage
Do not upload sensitive documents without authorization, and remove unnecessary personal data. OpenAI says uploaded files are saved to Library and that deleting the original chat does not necessarily delete the Library file. Temporary Chat uploads are not saved to Library. On individual services, content may improve models when “Improve the model for everyone” is enabled; business and institutional policies differ.
Final checklist
- Confirm whether the PDF contains selectable text or scanned images.
- Ask for a document map before requesting conclusions.
- Analyze long documents section by section.
- Require page or section references for factual claims.
- Upload important embedded visuals separately when necessary.
- Prefer spreadsheets for exact table calculations.
- Spot-check quotations, figures and page numbers against the original.
- Remove sensitive information and review file-retention settings.
ChatGPT is most useful as a reading and analysis assistant, not as proof that every page has been processed perfectly. A structured workflow—map, divide, cite, verify and synthesize—produces a much safer result than a one-line request for a complete summary.
