Clioos › Research
How accurate is our PDF-to-Excel conversion?
We tested PDF Probe on 20 synthetic bank statements before launch. This page publishes the result, the method, and the limits — so the number can be checked instead of trusted.
Read this first: this is a synthetic-sample test, not real-user accuracy. The 100% below was measured on 20 clean, synthetic, single-page statements in 4 known layouts (Chase / Bank of America / Wells Fargo / Amex style). No real user files were used. Real uploads — phone photos, multi-page statements, unknown bank layouts — have not been measured yet, and the accuracy there is unknown. Do not quote this page as "100% accurate on real statements".
The result
Test date: 27 October 2026. Tool: PDF Probe, same extraction code as the live product.
| Group | Statements | Transactions | Cell accuracy | Row accuracy | Balance check passed |
|---|---|---|---|---|---|
| Digital PDFs | 10 | 104 | 100% | 100% | 10 / 10 |
| Scanned PDFs | 10 | 107 | 100% | 100% | 10 / 10 |
| Total | 20 | 211 | 100% | 100% | 20 / 20 |
Opening and closing balances matched the answer key on all 20 statements. Extraction took about 19 seconds for all 20 files (roughly 1 second per file, including OCR on the scanned group).
The final round was not the first round. Accuracy across the three tuning rounds was 67.77% → 93.84% → 100% (digital 81.73% / scanned 54.21% in round one). The fixes between rounds are listed under "Method" — they were tuned against this same sample, which is one of the limits below.
Method (so you can reproduce it)
- Sample. 20 synthetic statements: 10 digital PDFs + 10 scanned (rendered at 300 dpi, rotated ±0.7°, light noise and blur, saved as image-only PDFs with zero text-layer characters). Layouts model the common structures of Chase, Bank of America, Wells Fargo and Amex statements — structure only; no real bank data, and all names, accounts and amounts are fictional. 5 statements per bank layout.
- Answer key. Every file ships with a full transaction list (date / description / amount) plus opening and closing balances, where closing = opening + transaction total to the cent.
- Extraction. Text-layer check first (pdfplumber); digital files use two extraction strategies (ruling lines vs. text positions, keep whichever fills more cells); scanned files are rendered at 300 dpi and read with Tesseract 5.3.4, rows grouped by vertical position, amounts parsed by regex (transaction amount = second-to-last amount token, last = balance; two-column layouts use column position for withdrawal/deposit sign). Balances are reconciled afterwards: |opening + transactions − closing| ≤ $0.01 passes.
- Scoring. Transactions match on (date, amount ±$0.01); descriptions compared after normalisation (lowercase, punctuation and extra whitespace removed). Each row counts as 3 cells (date / description / amount); a missed row counts as 3 wrong cells, an extra row as 3 wrong cells.
- What the three rounds fixed. Round 1 → 2: Tesseract page-segmentation mode (psm 6 dropped whole tables; psm 4 reads them, psm 11 as fallback for very short tables), row-grouping tolerance widened for slight scan tilt, and a two-column sign-convention fix in the sample generator. Round 2 → 3: amount parsing moved from x-position splitting to regex on the row text (right-aligned amount columns were being split on the wrong side of the column boundary).
Limits (what 100% does not mean)
- Clean synthetic single pages only. Clear fonts, one page, no folds, no shadows, no perspective distortion, no small-print footnotes. Phone photos of statements, multi-page statements with continued tables, and faint print were not tested.
- Only 4 known layouts. The extractor recognises the table headers and column structure of these 4 layouts. There are hundreds of banks and credit unions; accuracy on an unknown layout is unknown.
- Tuned on this sample. Every fix above was triggered by a problem this sample exposed, so the result may be partly fitted to this sample. A different set of real statements could score lower.
- Description comparison ignores punctuation. Normalisation removes punctuation before comparing, so a misread "&" vs ";" (seen once: "AT&T Wireless" read as "AT&T; Wireless") did not count as an error. Strict character-for-character comparison would be slightly below 100%.
The honest reading: the minimum pipeline passed its go/no-go line on clean samples, so the probe was worth building. Real-world accuracy can only be measured on real uploads, which is what the live product now collects (files are deleted within 1 hour; nothing is used for training or reviewed by a person).
Why we built this: the demand behind it
Before building, we checked whether people actually search for this. Monthly US search volumes below were measured with the Google Ads API on 27 October 2026 (US, English, 12-month average); "convert pdf to excel" was cross-checked in Keyword Planner the same day, since the API did not return it.
| Keyword (US) | Monthly searches | Competition | CPC range |
|---|---|---|---|
| pdf to word | 201,000 | LOW | $0.22–$1.85 |
| convert pdf to word | 90,500 | MEDIUM | $0.36–$2.75 |
| pdf to excel | 27,100 | LOW | $0.29–$2.85 |
| convert pdf to excel * | 18,100 | MEDIUM | $0.44–$4.31 |
| pdf to excel converter | 9,900 | MEDIUM | $0.38–$3.23 |
| pdf to csv | 2,900 | LOW | $0.38–$2.16 |
| bank statement converter | 1,300 | MEDIUM | $2.48–$11.95 |
| extract table from pdf | 390 | LOW | $1.47–$7.90 |
| bank statement to excel | 260 | MEDIUM | $2.00–$11.69 |
| scanned pdf to excel | 50 | MEDIUM | $0.98–$8.24 |
* "convert pdf to excel" was verified separately in Keyword Planner on 27 October 2026; the other figures are Google Ads API measurements from the same day. Volumes are US monthly averages; the bank-statement niche is small in volume but has the highest advertiser bids, which is why the tool leads with statements rather than generic tables.
Try it, and read the guides
- PDF Probe — convert a PDF table or bank statement to Excel. First 10 pages free, no account.
- How to Convert PDF to Excel
- How to Convert a Bank Statement PDF to Excel
- How to Convert a Scanned PDF to Excel
This page is a living document: the URL stays the same and the numbers above are updated only when a new measured test replaces them. Questions: contact@clioos.com.