How to Summarize PDFs with AI

Different models have different context windows:

• Up to 50 pages: any model works. ChatGPT 4o is fastest.
• 50-200 pages: use Claude Sonnet 4 (200K context).
• 200-1500 pages: Gemini 2.5 Pro (2M context - Only one that fits a whole novel).

Unsure how big your document is in model terms? Paste a typical page into the token counter and multiply by page count. When a document sits near a boundary, go up a tier - A truncated summary fails silently, and that's worse than a slower model. For the long-document case specifically, Claude vs Gemini is the comparison that matters: Claude is more accurate within its window; Gemini's window is 10x larger.

All three are available on AskAI.free Pro.

On AskAI.free, drag-and-drop the PDF into the chat composer (Pro/Max feature). The AI extracts the text automatically - No need to convert to text manually.

If your PDF is image-only (scanned), the model will run OCR - Expect lower accuracy on poorly-scanned documents, and treat any numbers it quotes from a scan as unverified until you check them.

Common mistakes at this step: uploading a password-protected PDF (remove the password first; the extractor can't open it), and uploading 5 documents at once then asking questions that don't name which document you mean. One document per conversation keeps answers clean unless you're explicitly comparing them.

Before asking for any summary, run this check:

"Before summarising: what is this document's title, how many pages or sections does it have, and what is the final section about?"

The last part is the tripwire. If the model describes the ending correctly, the full document is in context. If it gets vague ("the document concludes with closing remarks") or describes a chapter that's actually in the middle, the file was truncated - Switch to a bigger-context model or split the PDF before continuing.

This 20-second check is the single highest-value habit in this guide. Skipping it is how people end up presenting a summary of two-thirds of a contract.

The all-in-one prompt that gives you a usable summary at three levels of detail:

"Summarise this document in three layers:
1. TL;DR - One sentence.
2. Key points - Bullet list, max 10.
3. Action items - What should I actually do based on this?"

You get a one-glance answer plus the depth if you want it. Adapt layer 3 to the document type: for a contract, "obligations, deadlines and penalties that apply to me"; for a research paper, "findings, limitations the authors admit, and what they'd study next"; for board minutes, "decisions made and who owns each follow-up."

Expected output: a tight executive summary in 15-30 seconds. If layer 2 comes back as ten vague bullets ("the document discusses various risks"), the document is probably long and diverse - Ask for a section-by-section summary instead, then have it roll those up.

The AI now has the whole document in context. Ask anything:

• "What does this contract say about termination?"
• "Are there any red flags I should be worried about?"
• "Compare what this paper says about X to the standard view."
• "Show me the relevant quote and its page number."

That last pattern deserves to be a habit, not an option. Appending "quote the exact sentence and page number" to any factual question forces the model to retrieve rather than improvise, and it turns verification into a 10-second page-flip. If the AI can't produce a quote for a claim, treat the claim as a guess.

Even Gemini caps at ~1500 pages of dense text. For longer documents:

1. Split into 100-page chunks (any free PDF splitter works).
2. Summarise each chunk with the tiered prompt above, in separate conversations.
3. Paste all the summaries into a final "summarise the summaries" prompt.

This loses some nuance - Cross-references between chunk 2 and chunk 9 disappear - So tell the final prompt what to look for: "These are summaries of one document in sequence. Note any contradictions or themes that recur across chunks." Keep chunk summaries detailed (300+ words each); over-compressed chunks give the final pass nothing to work with.

Sam is signing a commercial lease, 90 pages including riders. The flow: upload to Claude Sonnet 4 (90 pages fits 200K easily). Verify: "What's the final section about?" - Claude correctly describes the personal-guarantee rider. Tiered summary: the TL;DR flags a 5-year term with a personal guarantee; action items include a rent-escalation clause worth negotiating.

Then targeted follow-ups: "Quote every clause about early termination, with page numbers" (three clauses, pages 34, 41, 88 - The page 88 one contradicts page 34). "What's the total cost if I exit in year 2?" Sam checks the three quoted pages himself, confirms the contradiction is real, and walks into the lawyer meeting with two precise questions instead of paying for a full read. Time spent: 20 minutes.