Most “chat with your PDF” demos use clean, digital PDFs — files that already contain a perfect text layer. Real life is messier: photographed exam papers, marked homework, archive scans. So we ran a simple test: five pages that exist only as images, six questions each, scored against the answers that are visibly on the page (each question worth 2 points, 12 per page).
| Test page | Assistant A | Assistant B | Information on the page |
|---|---|---|---|
| Exam paper with handwritten answers | 0 — file refused | 4 / 12 | 12 / 12 |
| Vocabulary test with teacher's marks | 0 — file refused | 6 / 12 | 12 / 12 |
| Two-column journal article (clean print) | 0 — file refused | 10 / 12 | 12 / 12 |
| Contract with a fee table (clean print) | 0 — file refused | 12 / 12 | 12 / 12 |
| Same exam paper, tilted & blurred | 0 — file refused | 0 — processing failed | 8 / 12 |
Assistant A rejected every file with the same message: the text is too short to read; scanned and image-only PDFs are not supported. Five pages, zero attempts.
Assistant B was genuinely good at clean print — near-perfect on the journal article and the contract, including a multi-step fee calculation. But every question whose answer was handwritten came back “Not in document”, even when the handwriting was large and legible. It couldn't see the teacher's marks either. And when we degraded the scan — tilt, shadow, blur, the way real phone photos look — the file failed to process at all.
The information was there. On each page, everything asked for could be read directly from the image, including the degraded one (8 of 12 points were still recoverable). The tools just never got to it: one reads only the text layer, the other's OCR throws the handwriting away before the AI ever sees the page.
VersoDoc is an AI document companion built for scanned and image-only documents. It restores the page first — deskew, denoise, layout recovery — then reads print, handwriting, and marks together. The documents other tools can't open are the ones we optimize for.
It's not launched yet. Join the waitlist and you'll be in the first group invited. Pro will be $6.99/month (or $59/year); files are auto-deleted after 30 days and never used for training.