← Troubleshooting

The extracted text is wrong

OCR (optical character recognition) reads text from images the way a person squints at a blurry photo — usually right, sometimes not. When extraction comes back with errors or missing fields, here's what's going on and how to get a better result.

Why errors happen

Accuracy depends almost entirely on image quality. The most common culprits: blurry or shaky photos, receipts photographed at an angle, shadows or glare across the text, crumpled thermal paper, tiny fonts, and low-contrast printing (gray text on white). Handwriting is harder than print, and decorative or unusual fonts confuse the reader too.

How to get a cleaner extraction

1. Retake the photo. Lay the document flat, fill the frame with it, and shoot straight down in bright, even light. No flash — it creates hotspots. This single step fixes most bad results.
2. Flatten the paper. Smooth out curls and creases, especially on receipts. A second sheet of paper on top for a minute helps.
3. Crop tightly. If the tool lets you, crop out the table, hands, and background before processing. Less noise means fewer misreads.
4. Scan instead of photographing when accuracy matters (tax documents, contracts). A $0 phone scanner app beats the best camera technique.

What to do with the result you got

Always skim the output before using it — check totals, dates, and names especially. CSV and Excel outputs open in any spreadsheet app, where fixing a misread cell takes seconds. For W-2s and W-9s, verify every number against the original; masked SSNs/TINs are shown partially by design.

Good to know

• Stylized logos and decorative headers are the most commonly misread elements — and usually the least important ones.
• Tables with no visible gridlines (common in invoices) are harder to parse than ruled tables. If columns merge, that's why.
• Handwriting recognition works best with neat, separated letters. Connected cursive is its weak spot.
• Reprocessing the same image gives the same result — the fix is a better image, not a second run.
• If a clean, straight scan still extracts badly, send us the file through the report form. Real examples are how the models improve.