How to Prepare PDFs for ChatGPT, Claude, and Gemini

PDFs were designed for printers, not language models. When you drop a 200-page report into ChatGPT, Claude, or Gemini, the parsed text usually carries a long tail of headers, footers, page numbers, footnote markers, and broken hyphenated words. The model has to spend tokens — and attention — on all of it.

This guide walks through how to turn any PDF into the kind of clean, paragraph-structured text an AI can actually reason over. The steps work the same whether you're feeding Claude's 200K context window, GPT-4's 128K, or Gemini's 2M.

Why raw PDF text fails

Most extraction libraries pull text in reading order, but they don't know which lines repeat on every page. A typical academic PDF carries the journal name in the header, the page number in the footer, and the article title every few pages. Across a long document those repeated strings can total thousands of tokens.

Hyphenation is the second silent killer. PDFs break long words across line endings ("informa-tion") and most parsers preserve the hyphen. Suddenly your document has hundreds of words the model treats as misspellings.

The clean-text checklist

• Detect and remove repeated headers and footers across pages
• Strip standalone page-number lines
• Re-join hyphenated line breaks ("informa-\ntion" → "information")
• Collapse extra whitespace and double-spaces
• Rebuild paragraphs based on line spacing, not raw line breaks
• Run OCR if no embedded text is found (scanned PDFs)

Born-digital vs scanned: which path?

A born-digital PDF (one exported from Word, Google Docs, or LaTeX) has embedded text — extraction is fast and lossless. A scanned PDF is just images of pages; you need OCR. PDF to Clean Text automatically detects which kind you have and falls back to in-browser Tesseract OCR when needed. Nothing uploads.

How much does cleaning actually save?

On a typical 80-page consulting report we tested:

• Raw extraction: 64,200 tokens
• Header/footer stripped: 58,100 tokens (−9.5%)
• Page numbers + hyphenation fixed: 56,400 tokens (−12.2%)
• Paragraphs rebuilt: 54,900 tokens (−14.5%) and noticeably better summaries

Fourteen percent doesn't sound dramatic until you multiply it by every API call across a year. For RAG pipelines, cleaner chunks also mean better embeddings — your retriever finds the right paragraph more often.

Handling whole folders

If you're preparing a knowledge base or a multi-document brief, run the whole batch through Batch Document Extractor. It handles PDFs, Word docs, images, and plain text together and returns a ZIP with one clean file per input — perfect for feeding into Claude Projects, ChatGPT Custom GPTs, or any RAG ingestion pipeline.

Putting it in your prompt

Once you have clean text, structure your prompt clearly:

1. Lead with the question or task
2. Add the document under a clear delimiter ("---DOCUMENT START---")
3. Repeat the question after the document — models give noticeably better answers when the ask brackets the context

For very long inputs, also consider compressing first with Context Compressor before sending. The combination of clean + compressed often cuts your prompt in half without losing information.