The short answer: extract the PDF's text with PDFEdit's PDF to Text tool — free, in your browser — then paste the relevant sections into ChatGPT, Claude, or any AI as plain text. Clean extracted text beats uploading the raw PDF in almost every case: you control exactly what the model sees, it works on free AI tiers, and the model reasons better over clean text than over a jumble of PDF layout data.
Why text beats uploading the PDF
Three practical reasons:
- Not every AI accepts PDFs. Free tiers and many API workflows are text-only. Plain text works everywhere, every time.
- You control the context. A 200-page PDF upload burns context on boilerplate, headers, and appendices. Extracted text lets you paste the three chapters that matter.
- Cleaner input, better answers. PDF text extraction via upload often mangles reading order, duplicates headers/footers on every page, and chokes on tables. Extracting yourself first — with page-by-page review — lets you clean that up before the AI ever sees it.
The workflow
Step 1: Check the PDF type
Try selecting text in the PDF. If it highlights, it's a text PDF — extraction will be instant and accurate. If not, it's a scan: run it through OCR first (English, Spanish, French, German supported), then continue.
Step 2: Extract the text
Open PDF to Text and drop in the file. You get:
- Page-by-page view — inspect each page's text separately
- Word/character counts — know how much you're working with
- Search — find the sections you actually need
- Copy or TXT download — grab it all or just the parts
The "Page Dividers" option adds --- Page N --- markers between pages, which is genuinely useful for AI work — the model can cite "page 12" and you can verify.
Step 3: Trim ruthlessly
This is where most people go wrong — pasting everything. Before pasting:
- Cut boilerplate. Headers, footers, page numbers, and legalese appendices add noise, not signal.
- Keep structure. Leave headings and section breaks in. A model navigates "## Section 3: Results" far better than a wall of text.
- Paste sections, not the whole thing. For a 50-page report, paste the executive summary + the two sections relevant to your question. You can always paste more in a follow-up message.
- Note what's missing. If you cut tables or figures, tell the model: "Tables were removed; figures not included." It prevents hallucinated confidence about things it never saw.
Step 4: Prompt with context
Don't just paste text and write "summarize." Give the model a job:
- "This is a 40-page market report. I'm deciding whether to enter this market. List the 5 strongest arguments for and against, citing page numbers."
- "Here's a contract, sections 4–7. Flag anything unusual compared to a standard SaaS agreement."
- "This is chapter 3 of a textbook. Explain the key concepts as if I'm preparing for an exam, with examples from the text."
The page markers from Step 2 pay off here — "citing page numbers" actually works when the text includes them.
Scanned PDFs and AI: the OCR detour
If your PDF is scanned, there's no shortcut: OCR it first. PDFEdit's OCR renders each page and recognizes the text in your browser. Proofread the result — OCR errors ("rn" → "m", "0" → "O") become AI errors if you don't catch them, and the model won't know the difference. For important documents, this proofreading pass is non-negotiable.
Privacy: think before you paste
Extracting locally is the safe half — the text never leaves your device during extraction. The risky half is pasting it into an AI service. Rules of thumb:
- Never paste credentials, personal ID numbers, private health or financial details, or unreleased proprietary content into a chatbot unless you've checked the provider's data policy.
- Redact first if the document has sensitive bits mixed with the parts you need — extract, delete the sensitive passages, then paste.
- Prefer the AI provider's documented enterprise/privacy settings for work documents.
Common mistakes
Uploading the PDF and assuming the AI read it well. It probably read something. Clean extracted text you reviewed is verifiable — an upload is a black box.
Pasting 100 pages and asking a vague question. Long context + vague prompt = generic answer. Narrow the text, sharpen the question.
Forgetting tables flatten. "What was Q3 revenue?" works fine from flattened table text. "Describe the trend in the chart on page 8" doesn't — the model never saw the chart.
Skipping the proofread on OCR'd text. Garbage in, garbage out, with extra confidence.
Chunking strategy for long documents
AI context windows are large now, but "fits" isn't the same as "reasons well over." Long pasted text degrades answer quality — the model loses track of structure. Better approach:
- Extract once, paste in logical chunks. Chapters, sections, or ~10-page blocks — each as its own message or conversation turn.
- Summarize-then-synthesize. Ask for a tight summary of each chunk ("bullet the key points, keep page references"), then paste the summaries together and ask your real question over those.
- Keep page markers. The
--- Page N ---dividers from PDFEdit's tool let the model cite pages, and let you verify claims against the original. - Front-load the question. Tell the model what you're after before pasting the chunk — it reads with purpose instead of skimming aimlessly.
This chunked approach consistently outperforms dumping everything at once, even when the whole document technically fits.
Handling tables and figures for AI
Tables flatten into text sequences, which models actually handle reasonably — "Q3 revenue was $4.2M" is extractable from flattened table text. Help the model by:
- Labeling what you removed. "The following is the flattened text of Table 2 (regional sales). Columns were: Region, Q1, Q2, Q3."
- Keeping one row per line when you can — paste table text with line breaks between rows.
- Describing figures briefly yourself. "Figure 3 shows a bar chart of declining sales 2021–2024, steepest drop in 2023." Ten seconds of your description beats the model's confident guess about a chart it never saw.
Citing and verifying AI answers
AI summaries of your PDF are drafts, not sources. When the answer matters:
- Ask for page citations ("cite the page number for each claim") — possible because your text has page markers.
- Spot-check 2–3 citations against the original PDF before quoting the AI anywhere.
- Remember the model only knows the text you pasted. If you trimmed a section, its "complete" answer is complete only within what it saw.
Do it with PDFEdit
- Extract PDF text — page-by-page text, search, stats, TXT download, all in-browser
- OCR scanned PDFs — recognize text in scans before feeding AI
- Convert PDF to Word — when you want an editable document out of the process too
Extract locally, trim hard, paste with a sharp question. That's the whole technique — and it works on every AI chatbot, free or paid.
