OCR vs Text-Based PDFs: Why Some Bank Statements Convert Cleanly and Others Don't

Two PDFs that look identical on screen can convert with wildly different accuracy. The difference is whether the file contains real text or just pictures of text — and which one you have determines whether conversion takes 10 seconds or 10 minutes.

The 5-Second Test

Open the PDF in any viewer. Try to click and drag across a transaction row to highlight it.

• If individual characters highlight cleanly — it's a text-based PDF. Convert it directly, accuracy will be very high.
• If the whole row highlights as a single block, or nothing highlights — it's image-based. OCR is required, and quality matters.

Why Text-Based PDFs Are Easy

A text-based PDF embeds every character with its exact (x, y) coordinate on the page. A converter can reconstruct the original table by clustering characters into columns and rows. There's no character recognition needed — the text is already known.

ClearlyLedger uses geometry-based table detection on text PDFs (3-pixel Y-tolerance, 50-pixel column gaps) and reaches arithmetic correctness on essentially every text-based bank statement. Balance verification — covered in the accuracy post — confirms this on every file.

Why Image-Based PDFs Are Hard

When a PDF is image-based, every page is a picture. To extract numbers you have to (a) recognise characters from pixels using OCR, and (b) figure out which characters belong to which row and column using only visual gaps. Both steps add error.

The two biggest quality killers:

• DPI under 300 — characters are too few pixels for OCR to disambiguate 0 from O, 1 from l, 5 from 6, 8 from B.
• Skew over 7 degrees — a tilted scan breaks row alignment; numbers from one row "fall into" the row above or below.

ClearlyLedger rejects scans below these thresholds rather than producing a confident-looking but wrong Excel. That's a deliberate choice; quiet failure is worse than a fix-me message.

How to Get a Better PDF

1. Re-download from your bank's portal — the original PDF is almost always text-based. Statements forwarded by email are sometimes re-rendered as images.
2. If you only have paper, rescan at 300 DPI or higher, with the page laid flat against the bed (no phone photos held at an angle).
3. If you have a phone photo, use a document-scan app (CamScanner, Adobe Scan) that de-skews and binarizes — don't upload the raw photo.

What Conversion Tools Should Do With Each Type

A good converter routes the file based on text density: high-text PDFs go straight to the geometry parser; low-text PDFs go through OCR with a quality gate. A bad converter runs OCR on everything (slow and lossy) or refuses scans entirely.

The split matters for credit card statements too — covered in credit card statement PDF to Excel, since issuer PDFs are nearly always text-based but the statements forwarded inside expense-report apps often aren't.

Related Reading

• Bank statement conversion: the complete 2026 guide
• Balance verification: the only accuracy test
• Bank statement PDF passwords