Which file formats convert best to Markdown
Structured formats convert almost perfectly. Scanned images do not. Everything else sits in between, and the gap usually shows up in tables rather than in the prose. Below is what each source format actually gives you, including where it falls down.
Accuracy and table fidelity by source format
Text accuracy is the share of source text that survives conversion with its structure intact. CSV, JSON and XML hit 100% because they are already structured text — there is nothing to interpret. Scanned images sit lowest because OCR is a best guess.
| Format | Text accuracy | Tables | Typical time | Best for |
|---|---|---|---|---|
| CSV.csv | 100% | excellent | 1 second | Database exports |
| JSON.json | 100% | excellent | 1 second | API documentation |
| XML.xml | 100% | good | 1-2 seconds | Legacy system data |
| Excel.xlsx .xls | 99% | excellent | 2-4 seconds | Data tables and datasets |
| Word.docx .doc | 98% | excellent | 1-2 seconds | Draft documents |
| HTML.html .htm | 98% | excellent | 1-2 seconds | Web scraping projects |
| EPUB.epub | 97% | good | 2-3 seconds | E-books and novels |
| PowerPoint.pptx .ppt | 96% | good | 2-4 seconds | Training materials |
| PDF.pdf | 95% | good | 3-5 seconds | Scanned documents |
| ZIP.zip | 95% | varies | varies by content | Batch document processing |
| Audio.mp3 .wav .m4a .ogg | 90% | n/a | 10-30 seconds | Podcast transcription |
| Image.jpg .jpeg .png .gif .bmp .tiff .webp | 85% | fair | 4-8 seconds | Scanned documents |
Measured on our own conversion pipeline across a mixed corpus. Your documents will vary — a clean native PDF beats a photographed one by a wide margin, and both are counted as “PDF” here.
What each format costs you
Every format loses something. Knowing what, before you convert a few hundred files, is usually more useful than the accuracy number.
CSV
100% · low effortRows and columns survive intact.
- No formatting support
- Encoding issues with special characters
JSON
100% · low effortRows and columns survive intact.
- Deeply nested structures may be complex
- Large files require memory
XML
100% · medium effortUsually clean; complex merged headers need a look.
- Verbose format
- Schema validation may be complex
Excel
99% · medium effortRows and columns survive intact.
- Formulas converted to values
- Multiple sheets require individual processing
- Complex charts simplified to data tables
Word
98% · low effortRows and columns survive intact.
- Older .doc format may have compatibility issues
- Complex macro-enabled documents need special handling
HTML
98% · low effortRows and columns survive intact.
- JavaScript-rendered content requires pre-processing
- External CSS may affect formatting
EPUB
97% · low effortUsually clean; complex merged headers need a look.
- DRM-protected files cannot be converted
- Interactive features lost
PowerPoint
96% · medium effortUsually clean; complex merged headers need a look.
- Visual layouts simplified to text
- Animations and transitions lost
- Speaker notes preserved separately
Usually clean; complex merged headers need a look.
- OCR accuracy varies with scan quality
- Complex layouts may require manual review
- Encrypted PDFs need password
ZIP
95% · medium effortDepends entirely on the source markup.
- Password-protected archives need password
- Very large archives may time out
- Nested ZIPs require multiple passes
Audio
90% · high effortNo tabular content to preserve.
- Background noise affects accuracy
- Multiple speakers require separation
- Accents and dialects may impact quality
- Long files take more time
Image
85% · high effortSimple tables fine, nested layouts drift.
- Accuracy depends on image quality
- Handwriting recognition limited
- Complex layouts may need manual review
- Low-resolution images yield poor results
Picking a source format
If you control the export, choose CSV, JSON or XML. They are already structured text, so conversion is lossless and instant. This is the single biggest quality decision available to you, and it happens before anything reaches a converter.
Word and EPUB are the best of the “real document” formats. Both carry explicit heading structure, which is what makes the resulting Markdown useful to a language model rather than just readable.
PDF is the one to watch. A native PDF converts well. A scanned or photographed one goes through OCR, and accuracy then depends on the scan rather than on the converter. If your corpus mixes both, expect two very different quality tiers under one file extension.
Spreadsheets need a decision about formulas. Conversion keeps computed values, not the expressions behind them. For an audit trail you want the source file kept alongside the Markdown.
Convert one and see
The numbers above are averages. The only reliable test is your own worst document — a dense table, a scanned page, a deck with speaker notes.
Ship Markdown workflows with confidence
Convert documents, sync to your stack, and automate AI/LLM pipelines without managing infrastructure.