How to Summarise a Long PDF Without Reading It

Here is the uncomfortable thing about AI summarisation: the one-click summary is the least useful thing these tools do with a document, and it is the only thing most people ask for.

You upload a 90-page report, click summarise, and get four paragraphs that are technically accurate and tell you nothing you could act on. The problem is not the tool. It is that summarising optimises for what is representative, not what is important — and the single sentence that actually mattered gets averaged away.

This guide covers what works instead.

Why the one-click summary disappoints

A summary compresses a document toward its centre of gravity. If a 90-page contract has one unusual clause on page 62, that clause is 1% of the text and contributes roughly 1% to the summary. It vanishes.

But that clause is the only reason you opened the document.

This is why people try AI summarisation, find it underwhelming, and conclude the technology is overrated. They asked the wrong question. The tools are genuinely excellent at finding things in documents. They are mediocre at deciding what matters, because they do not know why you are reading.

You know why you are reading. Use that.

Pick the right tool first

The three we would reach for, and when.

NotebookLM — for study material and multiple sources. It only answers from documents you give it and cites the passage it used, which makes it the most trustworthy of the three for this job. It also handles several sources at once, so you can load a whole module or case file. Free, with limits most people never reach.

Claude — for long or high-stakes single documents. It holds long material in context more reliably than the alternatives and tends to quote accurately rather than paraphrase loosely. For a contract, a policy document or a long report, this is the one.

ChatGPT — for a quick answer on a short document. Convenient when you are already in a conversation and the PDF is 15 pages, not 150. File analysis requires a paid plan.

For a longer look at what each free tier actually allows, see what you actually get on free AI plans.

The interrogation method

This is the part that replaces summarising. Five steps, and it takes about ten minutes on a long document.

1. Ask what kind of document it is

Before anything else:

What kind of document is this, how is it structured, and what are its main sections?

You are building a map. This takes seconds and makes every later question better, because now you know where to point.

2. Ask your actual question

The one you opened the document to answer. Not “summarise this” — something like:

What does this say about termination and notice periods?

What are the stated risks, and which are described as most likely?

What does the author claim that they do not support with evidence?

That last one is unreasonably good and almost nobody asks it.

3. Ask what you did not think to ask

What is in this document that someone in my position would want to know but probably would not think to ask about?

This is the highest-value prompt on the page. It catches the unusual clause on page 62.

4. Verify before you rely on it

Quote the exact passage that supports that answer.

If it cannot produce the quote, treat the answer as unreliable. NotebookLM does this by default with clickable citations, which is why it is the safer tool when accuracy matters.

5. Only now, ask for a summary

By this point you know what the document contains, so you can direct the summary:

Summarise this for someone who needs to decide whether to sign it. Two hundred words. Lead with the risks.

A directed summary is genuinely useful. An undirected one is wallpaper.

A worked example

Say you have a 60-page tenancy agreement.

The bad approach: “Summarise this.” You get: the document is a tenancy agreement covering rent, deposit, maintenance responsibilities and termination. Thank you, yes.

The interrogation approach:

1. “What are the main sections and how is this structured?” — you learn there is a schedule of conditions at the back you had not noticed.
2. “What am I responsible for repairing, and what is the landlord responsible for?” — the actual question.
3. “What in this agreement is unusual compared with a standard tenancy?” — this is where the surprises surface.
4. “Quote the exact clause about ending the tenancy early.” — now you can read the words yourself.
5. “Summarise the three things I should negotiate before signing.”

Same document, same ten minutes, completely different value.

Prompts by document type

The interrogation method adapts. Here are the questions worth asking for the document types people most often bring to it.

Research papers. Ask “what did they actually measure, and what did they claim that the measurement does not support?” — the gap between those two is where research papers are weakest and where a summary hides the problem. Then ask about sample size and limitations specifically, because those sections get compressed away first. For literature at scale, Elicit is built for this.

Contracts and agreements. Ask “what obligations does this create for me that a standard agreement of this type would not?” Then, separately, “what happens if I want to leave early?” and “what is not covered here that I would expect to be?” Always demand the exact clause text before relying on an answer.

Financial reports. Be careful — this is where table misreading bites hardest. Ask for the narrative first (“what story do these numbers tell?”), then verify every individual figure you plan to use against the document yourself.

Meeting transcripts. Ask “what was decided, what was deferred, and who committed to what?” This three-part question produces far better results than “summarise”. If you are generating the transcripts too, Otter handles the recording side.

Policy and compliance documents. Ask “what would put me in breach of this?” — inverting the question surfaces the operative content far faster than reading forward.

Academic reading you have been set. Load it into NotebookLM alongside your lecture notes and ask what the reading adds that the lectures did not cover. That comparison is the actual assignment, most of the time.

Where all of them still fail

Being straight about the limits.

Scanned PDFs. If the file is images of text rather than real text, quality falls off sharply. Test it: open the PDF and try to select a sentence. If you cannot, expect problems, and expect tables to break entirely.

Complex tables. Every one of these tools reads tables less reliably than prose. A table that spans pages, or uses merged cells, is frequently misread — and the answer will still sound confident. Verify any number you plan to act on.

Diagrams and charts. Broadly ignored or badly described. If the meaning of the document lives in a figure, you are still reading that figure yourself.

Very long documents on free tiers. Upload limits and context windows bite here. NotebookLM handles multiple sources best; Claude handles single long documents best.

Documents in more than one language. Mixed-language documents produce noticeably worse results across all three.

The confidentiality question

Worth stating plainly because most guides skip it.

Uploading a document sends it to a third party. On consumer plans, your content may be used for training unless you turn that off in settings. Business, team and enterprise plans exclude it by default.

If the document is a client contract, patient information, unpublished research or anything covered by an NDA, that is a decision to make deliberately — and often a decision your employer has already made for you. Check the policy before, not after.

The short version

Stop asking for summaries. Ask the document questions, ask what you did not think to ask, demand a quote before you rely on an answer, and save the summary for last when you can tell it what the summary is for.

Use NotebookLM for study and multiple sources, Claude for long or high-stakes documents, and ChatGPT when speed matters more than rigour.

If you are studying rather than working, our NotebookLM study guide covers the same tool from a revision angle, and how to use Perplexity for research covers the case where the source is the open web rather than a file on your desk.

Comparing tools for chat, extract and Q&A: best AI PDF tools.