What we test
Every tool is checked against normal inputs and against the cases most likely to disappoint someone. The matrix below is the coverage target for the full library, not a claim that every possible file or language has been tested:
| Category | What we exercise |
|---|---|
| Writing & Text | Empty input, over-limit input, names/numbers/dates preserved, meaning drift, local text utilities with Unicode and punctuation. |
| PDF & Documents | Large files, corrupt files, password-protected files, page ranges, rotations, image-only pages, and before/after page counts. |
| Image Tools | PNG/JPG/WebP/BMP/GIF inputs, oversized images, transparency, orientation, output dimensions, and visual quality after conversion. |
| AI Productivity | Missing tasks, multiple owners, absent deadlines, decisions mistaken for actions, and false owner/date inference. |
| Web & SEO | Invalid JSON/CSV/URLs, boundary values, quotes and special characters, malformed timestamps, and exact output format. |
- PDF to Markdown: text-layer PDFs, multi-page documents, headings, bullet lists, simple tables, and scanned PDFs with no text layer.
- Image to Text: clean screenshots, receipts, low-contrast captures, small type, rotated images, and images with no readable text.
- PDF Summarizer: short briefs, long reports, number-heavy pages, and documents with conflicting conclusions. The summary is compared with the source for missing qualifiers and invented figures.
- Email, Tone, Resume, Checklist, Meeting, Transcript, and Alt Text: normal inputs plus empty text, over-limit text, ambiguous instructions, and outputs that could change the meaning of the source.
What a pass means
A pass means the tool produced a usable result for that test case, not that it is correct for every document. For AI tools, we check that the output preserves names, dates, amounts, refusals, and stated uncertainty. A polished sentence that changes the request is a failure, even if it reads well.
Known limits
- OCR is probabilistic. Numbers, dates, order codes, and punctuation need to be checked against the image. Tesseract is designed for printed text, not handwriting.
- PDF tables are visual, not structural. Merged cells, wrapped text, scans, and multi-column layouts can lose their rows and columns.
- AI summaries can omit context. A summary is a reading aid, not a substitute for the source. Check important numbers and caveats in the original.
- AI rewrites can drift. Tone changes are checked for meaning, but a model can still soften a refusal or invent a detail. Review anything that commits you to a date, price, or promise.
- Length limits are real. The current limits are shown on each tool page; long documents may need to be split by section.
What we do not claim
We do not publish an accuracy percentage for OCR or an AI benchmark. Those numbers depend on the document, the language, the scan quality, and the model version. We also do not claim that a tool is a substitute for legal, medical, financial, or accessibility review.
If a result is important, keep the source and verify the output. That is true for every AI product, including this one.
How we update
When a limitation changes, we update the tool page and this page together. If you find a reproducible failure, send the input type, the browser, and what you expected to hello@aibynow.com. Do not send confidential documents; describe the pattern instead.
See also how AIbyNow handles your data and the editorial policy.