How to Summarize a Long PDF With AI — Without Losing the One Number Your Boss Will Ask About
You drop a 60-page report into a chatbot, ask for a summary, and get a clean one-pager back in seconds. Then your boss asks about one number — and it is not there.
That is not a bug in your prompt. Summarization is compression, and without instructions about what must survive, the model compresses away exactly the details you needed most.
This guide shows you the workflow that keeps them: pick the right tool for the document, extract before you summarize, and run a five-minute check before you trust the result. Figures are as of November 2026 unless noted.
Why the details vanish — and why a bigger window does not fix it
A summary is a smaller version of the document, so something has to go. The question is what — and the model decides that for you unless you tell it otherwise.
Researchers at Stanford and Berkeley measured the problem in 2023. In their "Lost in the Middle" experiments, they hid the one document containing the answer among 19 distractors and moved it from the start to the middle to the end. For GPT-3.5-Turbo, accuracy fell from 63.2% when the answer sat at the end to 53.8% when it sat in the middle — worse than answering with no documents at all. The pattern showed up in every model they tested, including ones built for long context.
Two things follow from that. First, the middle of a long document is a partial blind spot: a clause on page 41 or a caveat in section 7 is the most likely thing to disappear. Second, a bigger context window does not solve it. Today's flagship models advertise windows of roughly a million tokens — about 750,000 words, or several novels — yet Anthropic's own documentation warns that accuracy and recall degrade as the context grows, an effect it calls "context rot."
One independent test makes the cliff concrete. A 1,300-word article summarized cleanly, keeping its argument, examples, and counterarguments. The same approach on a 4,000-word article kept the main message but lost most of the supporting evidence, quotes, and statistics. The summary looked complete. The proof behind it was gone.
So the fix is not a longer window or a cleverer one-line prompt. It is a workflow that makes omission visible before you rely on the result.
Pick the right tool before you upload
Two kinds of tools summarize documents, and they fail in different ways.
- General chatbots (ChatGPT, Claude, Gemini): Flexible — they reformat, answer follow-ups, and compare across documents. Their weakness is confidence: when a detail is missing from what they read, they can fill the gap from general knowledge and sound certain doing it.
- Source-grounded notebooks (Gemini Notebook, formerly NotebookLM): They answer only from the files you upload, and every claim links back to the exact passage it came from. If the answer is not in your sources, the tool says so instead of inventing one. Google renamed NotebookLM to Gemini Notebook in July 2026; the free tier allows 50 sources per notebook, each up to 500,000 words.
Use the chatbot when you want to explore a document. Use the notebook when a wrong number would cost you — a contract, a client report, a paper you plan to cite.
The 5-step workflow that keeps the details
The fix is not a better prompt. It is a better order: decide, clean, extract, then summarize.
1. Decide what must survive — before you upload
Write the must-keep list first, while you still know what matters.
Before you paste anything, write down the numbers, names, dates, decisions, and exceptions that cannot drop out. For a quarterly report, that might be revenue, churn, the three risks, and who owns each action item. For a contract, it is dates, amounts, penalties, and renewal terms.
This list does two jobs: it tells the model what to protect, and it gives you something to check against at the end.
2. Clean the file first
Two minutes of prep saves a summary full of holes.
If the PDF is a scan — a photo of pages rather than selectable text — run it through OCR first. AI cannot read text that is not there. Then remove what you do not need: long appendices, reference lists, and boilerplate headers. Every page you cut is attention the model can spend on the pages that matter.
Check tables and charts separately. Models often misread visuals, so ask for tables and figures on their own instead of trusting the prose summary to carry them.
3. Extract before you summarize
Summaries compress, and compression is where details die. Extraction does not compress — it lists.
Ask for a structured extraction first, then summarize from the extraction. The prompt looks like this:
From the document below, extract — do not summarize yet — every key claim with the section it appears in, every number with what it represents, and every decision, deadline, and exception. If the document does not state something, write "NOT STATED." Do not add outside facts and do not calculate new values.
Once the extraction exists, the summary is a rewrite of a complete inventory instead of a guess about what mattered. The step also makes gaps visible: a "NOT STATED" is honest, while a smooth sentence that quietly invents the answer is not.
4. Chunk long documents with overlap
When a document is too long for one pass, split it by its own headings — and overlap the cuts.
Summarize each section separately, then merge the section summaries into one final summary. Keep about 10% overlap between chunks, so a fact that straddles a boundary appears in both. Repeat your must-keep list at the end of every chunk prompt, because instructions at the start of a long input get the same middle-of-the-context treatment as the document itself.
As a rule of thumb, 1,000 tokens is roughly 750 words. A 60-page report runs to about 30,000 words, or 40,000 tokens — well inside a modern window, and still long enough for the middle to go soft. Chunking is about attention, not capacity.
5. Ask for the numbers separately
Numbers get buried in prose summaries. Pull them out on their own.
Ask for a separate table of figures: each number, what it represents, and the section where it appears. Do the same for action items — task, owner, deadline — and for anything the document explicitly leaves unresolved. If the summary and the numbers table disagree, trust the table and check the original.
What to extract first depends on the document
A contract and a research paper do not lose the same things, so do not ask for the same summary.
- Contract or agreement: Clauses, dates, amounts, penalties, and renewal terms. Ask for each clause separately and never let the model summarize across clauses.
- Research report or white paper: The headline finding, the sample size and method limits, and the numbers behind each claim. A finding without its sample size is a slogan.
- Meeting notes or transcript: Decisions, owners, and deadlines — kept separate from what was merely discussed. "We talked about X" is not "we decided X."
- Financial or quarterly report: The three numbers that move the story, how each changed from last year, and the risks and caveats. Year-over-year change matters more than any single figure.
The 5-minute check before you trust it
Never forward a summary you have not checked. Five minutes catches most failures.
- Count check: Ask how many sections, clauses, or chapters the document has, and how many the summary covered. A mismatch is a dropped section.
- Spot-check three numbers: Pick the three figures that matter most and find each one in the original. If one is wrong, check them all.
- Run the challenge prompt: "Compare this summary with the document. List anything you left out — exceptions, dates, numbers, minority views, unresolved questions. Do not rewrite the summary." Omissions are easier to catch when you ask for them directly.
- Click the citations: In a source-grounded tool, every important claim should open the passage it came from. If you cannot click through, treat the claim as unconfirmed.
- Read the middle yourself: If the section that matters most sits in the middle of the document, read that section in the original. That is where the model is weakest, and where your judgment is cheapest.
Why chatbots invent facts in the first place — and when they can be taught to say "I don't know" — is covered in our earlier guides to how chatbots generate answers and whether they can admit uncertainty.
Before you upload: 30 seconds on privacy
A file-upload button is not a privacy guarantee.
Remove names, account numbers, passwords, health information, and unpublished financial data before you upload — or confirm your workspace plan's data policy first, because consumer and workplace accounts handle uploads differently. Do not upload anything you would not email to a stranger. Our comparison of what happens to prompts in ChatGPT, Claude, and Gemini walks through each provider's controls.
Key Takeaway
A good AI summary is not a shorter document. It is a complete inventory, shortened on purpose, with every number traceable to the page it came from.
Tell the model what must survive, extract before you summarize, and check the middle — the summary is a map, not the territory.
Sources
- Liu, N. F., et al. (2023). "Lost in the Middle: How Language Models Use Long Contexts." arXiv:2307.03172 — arxiv.org/abs/2307.03172
- Google Gemini Notebook help center — source limits and citations: support.google.com/gemininotebook
- Rephrase.world — "ChatGPT Won't Summarize a Long Article? Here's How to Fix It" (1,300-word vs. 4,000-word test): rephrase.world

Comments
Post a Comment