The classic scenario: you try to copy a paragraph from a PDF and discover you can't select any text — or extraction returns garbled characters and boxes. In almost every case that's a scanned PDF: it's really just photographs of pages, with no actual text inside.

Knowing the type matters because the two demand completely different handling:

Type	What it is	Correct handling
TextBased	Real encoded characters in the content stream	Extract directly: millisecond-fast, free, 100% faithful
Scanned	Each page is just an image	OCR required: slow, error-prone, often paid

Running OCR on a text-based PDF is pure waste — yet about 54% of PDFs are text-based and never needed it.

Method 1: Free online detection (fastest)

Open the home-page demo and drop in your PDF:

• A verdict in about 10–50ms: TextBased / Scanned / ImageBased / Mixed;
• A 0–1 confidence score included;
• For mixed documents, the exact list of which pages need OCR;
• Fully local in your browser — nothing is uploaded.

Why machine detection beats eyeballing

Some PDFs look like they have text but carry "fake" text — broken font encodings (CID/Type0 without ToUnicode maps) that copy out as garbage. pdf-inspector walks the font-decoding chain; such pages are flagged for OCR. You can't spot this by eye.

Method 2: One command from the terminal

After installing the CLI (install page), detect-pdf does exactly this:

detect-pdf report.pdf

It prints the classification, confidence, and the pages needing OCR — ideal for routing in batch scripts: text-based files take the fast path, the rest go to OCR.

Method 3: From code (Node.js / Python)

Node.js

const { detectPdf } = require('@firecrawl/pdf-inspector')

const result = await detectPdf('report.pdf')
console.log(result.type)        // "TextBased" / "Scanned" / ...
console.log(result.confidence)  // 0–1

Python

from pdf_inspector import detect_pdf

result = detect_pdf("report.pdf")
print(result.type)         # classification verdict
print(result.needs_ocr)    # list of pages needing OCR

More detail in the API quick reference.

Method 4: Manual checks (no tools)

Three rules of thumb, right maybe eight or nine times out of ten:

1. Try selecting text: if a cursor can highlight and copy words, it's probably text-based; if one drag grabs the whole page like an image, it's scanned.
2. Zoom way in: at 400% scans usually show paper noise, gray backgrounds, binding shadows, or slight skew.
3. Check file size: page-for-page, scans are much larger (an image per page). A suspiciously tiny "100-page document" may be an empty shell or a low-quality scan.

Where the eyeball fails

None of these catch pseudo-text: selectable, copyable — but garbage when pasted. That's a text-based PDF with broken font mappings, and only a proper detector (method 1) can identify it.

What to do after detection

A PDF arrives
   │
   ▼ Detect (~10–50ms)
   │
   ├─ TextBased ──────────► Extract / convert locally (free, milliseconds)
   │
   ├─ Scanned / ImageBased ──► Send to OCR
   │
   └─ Mixed ──► OCR only the flagged pages; extract the rest locally

That's selective-OCR routing, built into the native pdf-inspector packages, with per-page provenance (native / ocr / fused). The full strategy is covered in the errors & limits guide.

How do I convert a scanned PDF to text?

The browser demo includes no OCR. If your file truly is scanned:

• Use a desktop tool with OCR (e.g. Adobe Acrobat) or a cloud service;
• Or use the selective OCR built into pdf-inspector's native packages (Rust / Node.js / Python), which runs OCR only on pages rejected by quality checks.

FAQ

Text selects fine but copies out as garbage — why?

Usually missing or damaged ToUnicode CMaps on CID/Type0 embedded fonts: glyphs render correctly but the character-to-Unicode mapping is broken. Confirm with online detection; pdf-inspector handles most common CMaps and explicitly flags pages it cannot decode.

What does Mixed mean?

Part of the document is native text and part is images — say, typeset body chapters plus scanned stamped appendices. Per-page routing is the cheapest path: extract text pages directly, OCR only the image ones.

How should I read the confidence score?

A 0–1 float expressing how sure the classifier is. Below-threshold results deserve human review, or conservative treatment (treat as scanned). Engineering guidance lives in the troubleshooting guide.

Related pages

• Online parser: detection + Markdown conversion in one place;
• PDF to Markdown guide: your next step once it's confirmed text-based;
• CLI guide: full flags for detect-pdf.