Best UK bank statement converters: a reproducible test methodology

"Best converter" lists are usually paraphrased marketing pages. This page is the opposite: a test protocol specific enough that anyone can rerun it, and the standard we hold ourselves to before making any claim about another product. Named-competitor comparisons appear on this site only after a product has actually been through this protocol, with the date of testing shown.

Test corpus

1. Three digital statements per bank from the banks under test: a short single-page statement, a multi-page statement of at least 60 transactions, and one containing wrapped multi-line descriptions and a year boundary.
2. One scanned statement (print and re-scan one of the above).
3. One password-protected PDF.
4. One non-statement PDF (any letter) to test honesty about unsupported input.

For every digital statement, record ground truth first: transaction count, opening and closing balances, and the full values of five sampled rows including the first and last.

Measurements

For each tool and each document:

• Row recall: does the output contain exactly the true number of transactions?
• Sampled accuracy: are all five sampled rows exactly right (date, description, amount, direction, balance where present)?
• Reconciliation: does opening plus credits minus debits equal closing in the output? Does the tool itself check this, and how does it label a failure?
• Honesty on bad input: what happens with the scan, the encrypted file, and the letter? A guessed answer scores worse than a refusal.
• Data handling: with the browser's network panel open, does the document leave the machine? What does the tool's own documentation claim, and do the two match?
• Price for the workload: cost to export 10 and 50 statements, at prices checked on a stated date.

Scoring

No weighted composite score; they hide failures. Report the raw table and highlight disqualifiers: any silently missing row, any un-flagged reconciliation failure, and any document upload that contradicts the tool's privacy claims are each disqualifying for professional use, whatever else the tool does well.

Our own results

This site's converter runs a stricter automated version of this protocol continuously; the current numbers, including deliberately corrupted fixtures that must never verify, are on the accuracy page. When we publish comparisons against named products, each will link back here, state its test date, and acknowledge where the other product is stronger, because some will be.

Protocol version 1.0, published 15 July 2026.