How to summarise a long PDF with AI without missing what matters
Split long PDFs into sections, ask for page references, and check key points against the source, so an AI summary doesn't skip dates, numbers, or exceptions.

Why long-PDF summaries miss things
Give an AI assistant a 200-page report and ask for a summary, and you will get a confident page of bullet points. The problem is what isn't in it. Long summaries tend to over-represent the introduction and conclusion, smooth over exceptions ("except where…", "unless…"), and drop exact numbers in favour of words like "significant".
The fix is not a cleverer prompt. It is giving the AI less to read at once, asking it to show where each point came from, and checking the points that matter to you.
Step 1: Decide what matters before you start
Write down three to five things you need from the document. For a rental agreement that might be rent, deposit, notice period, and penalties. For a research paper: method, sample size, main result, and limitations. For an exam notification: eligibility, dates, fees, and document requirements.
This list becomes part of every prompt. A summary written for "anyone" will rarely match what you needed.
Step 2: Cut the PDF into sections
Use the table of contents to find natural breaks such as chapters, parts, or annexures. Then:
- Use Split PDF to break the file into chapter-sized parts. Roughly 10 to 40 pages per part works well for most assistants.
- Use Remove pages to drop covers, blank pages, and long indexes that add noise.
- If only one part is relevant, use Extract pages and leave the rest out.
All three run in your browser, so you are not uploading the full document just to trim it.
Step 3: Convert to Markdown so page numbers survive
A PDF's printed page numbers don't always match the PDF's own page count, and AI tools may not report either reliably. PDF to Markdown creates one section per page with a page heading, so when the AI cites "page 47" you can find it. It also keeps the file small enough to fit more into one conversation. If the PDF is a scan, run OCR PDF first.
Step 4: Summarise each section, then combine
Use the same prompt for every part so the summaries line up. For example:
"Summarise this section in no more than 10 bullets. For each bullet, give the page number. List every date, amount, deadline, and number exactly as written. List any exceptions or conditions separately. Then answer: [your three to five questions]. If something is not in this section, say 'not in this section'."
The last line matters. It gives the AI a way to say "not here" instead of guessing.
Once every part is done, paste the section summaries into a new chat and ask for an overall summary that keeps the page references. Then ask a second question: "What topics in these summaries might need more detail, and which sections should I reread?"
This two-stage approach takes longer than one big prompt, but each stage is small enough to check.
A worked example: a 90-page insurance policy
You want to know what your health policy covers before a planned surgery. Your list: waiting periods, exclusions, room-rent limits, and how to claim.
- The contents page shows the definitions, coverage, exclusions, claims procedure, and schedule of benefits. Split the PDF into those five parts.
- Skip the definitions for now. Summarise coverage, exclusions, and claims separately with the fixed prompt, adding your four questions.
- The exclusions summary mentions a waiting period for "specified illnesses" on page 38. You open page 38 and find the list, which includes your procedure, and a waiting period that you then compare with your policy start date.
One long prompt might have mentioned "some waiting periods apply". The section-by-section approach gave you the page, the list, and the number.
Step 5: Check what you will act on
You don't need to check every bullet. Check the ones you will act on: the deadline you'll put in your calendar, the fee you'll pay, the clause you'll rely on in an argument. Open the page the AI cited and read the sentence yourself, plus the sentences around it, because exceptions often sit in the next line. With a text export from PDF to text, you can search for a few words of the AI's quote and jump straight to the page.
Common mistakes
- Trusting "the document says" without a page number. No page reference means you haven't verified it yet.
- Summarising a scan without OCR. The assistant may read only part of the image, or misread figures.
- Letting the summary replace the source for legal or money decisions. Use the summary to find the right pages, then read those pages.
- Uploading the whole file when you need one chapter. It is slower, and you share more than you need to.
FAQ
How many pages should I send at once?
There is no single right number, because limits differ by assistant and plan. Chapter-sized parts of roughly 10 to 40 pages are easy for both the AI and you to check.
Does Split PDF upload my file?
No. Split PDF runs entirely in your browser.
Why convert to Markdown instead of uploading the PDF?
Markdown from PDF to Markdown keeps a heading for each page, so page references are easier to check, and the file is smaller.