“Summarize this PDF” is the prompt most people try first, and it is the one least likely to give you anything useful. You get five paragraphs of plausible paraphrase: no numbers, no citations, no way to tell what the model actually read. This workflow flips the dynamic. You ask narrow questions, demand quotes with page numbers, and turn a long document into a table you can verify line by line. It is aimed at analysts, researchers, and operators reading more than two long files a week — the people who get burned when a fabricated figure slips into a deck.
TL;DR
- Never open with “summarize.” Open with one specific question tied to a section or column.
- Force citations: “Quote the exact text and give the page or row number for every figure.”
- Spreadsheets go through Advanced Data Analysis (real Python on Plus, Pro, Team, and Enterprise), not the chat model’s guesswork.
- Spot-check at least 3 quotes or cells against the original before you trust anything.
- Reuse a file repeatedly? Put it in a Project so you stop re-uploading and burning context.
Where ChatGPT helps and where it does not
This workflow is for extracting structured data, comparing two or three files, or finding one specific answer buried in a long report. It is not the right tool for files holding sensitive personal or proprietary data you are not allowed to send to OpenAI, and it is not a replacement for SQL or a notebook when a dataset runs to hundreds of thousands of rows. For anything bigger than a spreadsheet, ChatGPT is best used to write the query, not run the analysis.
File limits you have to respect (as of June 2026)
Most “ChatGPT lost half my document” complaints trace back to a limit nobody read. The current caps:
| Constraint | Limit (June 2026) | Notes |
|---|---|---|
| Max size per file | 512 MB | Hard ceiling across all file types |
| Text / document token cap | 2M tokens per file | Applies to PDF, Word, slides — not spreadsheets |
| CSV / spreadsheet size | ~50 MB | Token cap does not apply; row width affects it |
| Image upload | 20 MB per image | Used for scanned-page OCR too |
| Uploads per 3 hours (Plus) | 80 files | Free ~3/day; Pro effectively unlimited |
| Files at once in one message | 10 | Number them so the model never guesses which is which |
| Project file storage | Free 5 / Plus 25 | Enterprise and Pro higher; shared 25 GB/user cap |
The in-app context window matters more than people think. On a Plus plan, GPT-5.5 reads roughly 320 pages of text in one conversation; the full 1M-token window is only available on the $200 Pro tier. A 60-page report fits comfortably; a 600-page contract does not, and the model will silently work from whatever fragment it retrieved.
The seven-step workflow
- Skim it yourself first. Note section names and exactly what you want. Two minutes here saves ten chasing hallucinations.
- Upload with one specific question, never “summarize.” Try: “What conversion rate does Section 3 report, and on what sample size?”
- Demand citations on every claim. Use a fixed phrase: “Quote the exact text and give the page number for each figure.” If it cannot cite, treat the number as fiction.
- Make spreadsheets describe themselves first. “List the column names, data types, and row count” before any analysis. This confirms the file parsed fully.
- Number multi-file comparisons. Upload them in one message: “File 1 vs File 2 — show the 5 metrics that differ most, with the figure and page from each.”
- Export a verification table. Ask for Markdown or CSV columns:
claim | source page | my note. This is the artifact you keep, not the chat. - Save the working prompt as a template. Most file work falls into three or four repeat patterns; do not retype them.
Worked example: a 60-page market report
Upload the PDF and ask for the table of contents first — this confirms the whole file parsed. Drill into one section at a time rather than asking everything at once; narrow context means sharper retrieval. For each numeric claim, re-ask: “Quote the sentence and page.” Finish by exporting the claim | source page | my note table and spot-checking three figures against the original PDF. If a “page 41” citation points at the wrong sentence, the model retrieved badly and every other figure is suspect too.
Spreadsheets: use Advanced Data Analysis, not the chat model
For any spreadsheet past a few thousand rows, switch on Advanced Data Analysis (the renamed Code Interpreter, on Plus, Pro, Team, and Enterprise). It writes and runs real Python with pandas in a sandbox instead of reasoning over text. Click View analysis to read the actual code — if the pandas logic is wrong, the answer is wrong, and you can see exactly where. Ask it to output a cleaned CSV or a chart rather than a prose summary; you can verify a file, you cannot verify a vibe.
Common mistakes that produce wrong answers
- Trusting unsourced numbers. Always re-ask for the exact quote and page. An uncited figure is a guess.
- Dumping 10 files with a fuzzy question. The model will mix up which file said what. Number them or compare in pairs.
- Analyzing before the file fully parsed. Long PDFs truncate silently. Confirm the TOC or row count first.
- Treating spreadsheet output as final. Spot-check a few cells; read the Python in View analysis.
- Opening with “summarize.” Summaries average everything; specific questions surface what matters.
- Re-uploading the same file every chat. Put it in a Project once and query it across conversations.
FAQ
- Does ChatGPT actually read the whole file?: Not always in one pass. It chunks long files and retrieves the most relevant parts per question, so sections can be silently skipped — especially past the in-app context limit (~320 pages of text on Plus). Verify with targeted, section-specific queries rather than trusting a single broad answer.
- Free, Plus, or Pro for file work?: Free is too tight — about 3 uploads a day and no Advanced Data Analysis. Plus ($20/mo) is the practical floor: 80 uploads per 3 hours, Advanced Data Analysis, and reliable PDF parsing. Pro ($200/mo) only matters if you need the full 1M-token window for documents over ~320 pages.
- What about scanned PDFs?: Mixed. ChatGPT OCRs them, but accuracy drops sharply on tables and footnotes. For a critical scanned figure, upload the page as an image (20 MB cap) and cross-check against the original.
- Can I analyze a Google Doc by link?: No. The model cannot follow auth-walled URLs. Export to PDF or paste the text. The same applies to most SharePoint and Notion links.
- Why does it say I hit the upload limit?: On Plus you get 80 files per rolling 3 hours; Free is roughly 3 per day. The counter is per-window, so wait it out or consolidate files. Project storage is separate (5 files Free, 25 Plus).