You can handle this directly in your browser with our free pdf to word tool — nothing gets uploaded.
Most OCR applications and scanner apps will never publish hard numbers. You will not see a label that says “this app achieves a 98.6 percent word accuracy on the standard benchmarks.” The marketing pages talk about “high accuracy,” “AI-powered recognition,” and “best-in-class language support,” but they rarely show you the kind of test data you could use to make a decision. That gap is exactly why comparing OCR accuracy by yourself, on the documents you actually care about, is the only reliable way to choose a tool.
This guide walks through how to build a small, fast benchmark that answers the question that matters: *which scanner app actually gets my documents right?* You do not need a laboratory. You need a handful of test samples, a spot-check method that takes twenty minutes, and a short list of features to weight when the numbers are close.
The reason OCR vendors shy away from publishing accuracy figures is that accuracy is not a single number. It is a curve that drops off sharply as the input gets harder. A scanner that reports 99 percent accuracy on clean laser-printed text can fall to 70 percent on a coffee-stained photocopy, and to 40 percent on a 1970s mimeograph with two-column layout. The vendor would rather quote the clean-number so the headline looks good, and the user ends up with no way to know how the tool behaves on the documents they actually scan.
There is also a second reason: the test set matters. Accuracy on a set of large-print documents from a modern printer tells you almost nothing about how the tool handles the small eight-point text in a medical consent form, or the handwritten notes at the edge of a page, or the low-contrast pages that come off an old photocopier. The only test set that predicts real-world performance is one built from the documents you will actually scan.
A good test set is small but varied. You do not need thousands of pages. You need five to eight pages that cover the failure modes you care about. Here is a practical layout.
| Sample | Description | Why it matters |
|---|---|---|
| Clean print, modern laser | 12-point serif or sans-serif, high contrast | Baseline. A tool that fails here has no business scanning anything else. |
| Small print, 8-point text | Dense paragraphs, e.g. a legal schedule or a data sheet | Tests how well the engine handles close-spaced characters at small size. This is the most common real-world failure in business documents. |
| Handwriting, cursive | A page of notes in cursive script | Tests whether the tool is print-only or genuinely multigraphic. Most consumer tools handle handwriting badly. |
| Low contrast / aged paper | Photocopied or scanned page, gray-on-gray ink on off-white stock | Tests contrast tolerance. Many tools binarize aggressively and lose faint text. |
| Rotated or skewed page | A page photographed at a 5 to 10 degree angle, slightly out of square | Tests page detection and deskew. If the tool does not find the text block, all downstream recognition is unreliable. |
| Two languages on one page | For example, an English body with Japanese footnotes | Tests language fallback and font coverage. Many tools assume a single language and drop the second one silently. |
| Table page | A grid or ledger with light rules | Tests whether the tool preserves structure or just emits a flat stream of words in the wrong order. |
Print or source these pages so they represent the quality you actually have. If you only ever scan a clean phone photo of a modern document, you do not need the handwriting row. If you digitize a box of 1980s files, you need the low-contrast and small-print rows badly.
The formal metric for OCR accuracy is the word error rate (WER) on text-heavy documents or the character error rate (CER) when you care about exact characters. Computing WER and CER automatically requires aligning the predicted text against a gold transcript with a dynamic-programming edit distance. That is the right thing to do when you have a corpus, but it is too heavy for a ten-minute comparison between two phone apps.
The spot-check method is the practical version. You do not need to type the full transcript. You need to sample words from the output and check them against the source.
The procedure:
This gives you a number you can put side by side. If Tool A lands at 44 out of 50 and Tool B lands at 49 out of 50 on the same small-print page, the difference is meaningful even though the absolute sample is small. Run the spot-check on each sample in your set and you get a per-sample accuracy table, which is far more informative than a single blended number, because it tells you *where* each tool wins.
You can also do a quick CER-style check by picking ten random phrases of about ten characters each and counting character-level mistakes. That is useful when you care about exact numbers, codes, or names, where a single character swap changes the meaning. Keep it to a handful of phrases or it stops being fast.
When the accuracy numbers are within a few points, the tiebreakers are usually about document handling rather than raw character recognition. These are the features to check before you commit to a scanner app.
Small-text handling. Ask the tool what happens on your 8-point sample. Does it detect the text block at all? Does it split the columns of a two-column layout, or does it read down the page in a jumbled order? A tool that gets the big headings right but scrambles the body text is a tool that will fail on dense pages, which is where most real documents are dense.
Table structure preservation. If your documents include tables, check whether the output is a flat stream of cells or a structured table. Some tools export to a format that keeps the grid; others just paste the words in reading order and lose the row-and-column relationship. For anything you plan to retype or copy into a spreadsheet, the structured export is the difference between usable and unusable.
Two-language and mixed-script support. If your pages mix two scripts, confirm the tool can do both on the same page without you switching a language setting between them. Many tools pick one language per page and silently drop the other. The mixed-script sample in the test set will expose this quickly.
Page detection and deskew. On the rotated sample, check whether the tool finds the page, corrects the angle, and crops out the background. A tool that needs you to rotate and crop the photo by hand every time is going to be slow on batch work, no matter how good its character engine is. Good page detection is a quiet but massive time saver.
Output format and editability. Accuracy means little if you cannot get the text out in a form you can use. Check whether the tool exports to a selectable text PDF, DOCX, plain text, or all of them, and whether the layout in the export matches the source. A tool that gives you clean text in a format you have to reformat by hand is less accurate in practice than a tool that gives you slightly messier text in a usable layout.
Batch behavior. If you are scanning dozens of pages, per-page friction compounds. Look at how the tool handles a folder of images: does it process them in order, does it keep page names, does it let you fix one bad page without re-running the whole batch? This is the angle the companion guide on batch OCR of scanned PDFs covers in more depth, and it is worth reading before you buy a tool you will run overnight.
When you run the spot-check, the mistakes you see will fall into a handful of patterns. Recognizing them tells you more than the raw error count does.
Missed characters on small text. The engine sees the line but drops individual letters, often the ones that are close together. You will see words that are almost right but missing a stroke. This is the signature of a tool that does not zoom in on small regions before recognizing them.
Wrong word substitution. The output has a full word where the source has a different one, usually a near-homophone or a visually similar pair. This shows up most on low-contrast and aged paper, where the engine is guessing rather than reading. A high rate of substitution errors on your clean-print sample is a red flag that the engine is weak overall, not just weak on hard input.
Order errors on multi-column and table pages. The text is all present but it is in the wrong sequence. The engine read down one column then down the next, or it flattened a table into a single line. This is a layout-analysis failure, not a character-recognition failure, and it makes the output look accurate by word-count even though it is unusable as a document.
Silent language drop. On the mixed-script page, one language is complete and the other is missing or replaced by boxes. The tool is not broken; it simply did not run the second recognizer. The fix is usually a setting you did not notice, but the symptom is a clean-looking output that is half empty.
Deskew and crop failures. The output includes the photo background, or the page is cut off at one edge, or the text is slightly tilted in the result. These are preprocessing failures that make every downstream check harder and sometimes lower the apparent accuracy even when the character engine is fine.
Garbled or mojibake characters. The most alarming failure mode is when the output contains nonsense symbols, boxes, or random characters that do not match the source in any way. This usually means the tool guessed the wrong character encoding or the wrong font and decoded the bytes incorrectly. If you see this on a clean, modern print sample, the tool is not just weak on hard input; it is misconfigured for your document type. The companion guide on garbled OCR output in scanned PDFs goes into what causes this and how to clean it up afterward.
Fifty words gives you a usable estimate when you are comparing two tools on the same page. The margin of error is wide enough that a four-word difference is suggestive rather than decisive, so run the check on more than one sample and look at the pattern rather than a single number. If you have the time, two hundred words per page drops the noise enough that a ten-percent difference between two tools is almost certainly real.
Word error rate counts whole-word mistakes: a substitution, a deletion, or an insertion of a word. Character error rate counts individual character mistakes inside and between words. For general reading and editing, word error rate is the more meaningful number, because a document with ten wrong letters spread across two hundred words is usually still readable. Character error rate matters more when your output is names, numbers, codes, or anything where a single wrong character changes the meaning. Use word error rate as your primary check and character error rate as a secondary check on the fields that matter.
Trust the one that matches the documents you will actually scan. A tool that is excellent on clean print and weak on aged paper is not a bad tool; it is a tool for a different job. If your work is mostly clean digital output, the clean number is the one that predicts your experience. If your work is mostly older files, photocopies, or handwriting, weight the harder samples more heavily and treat the clean-print number as a minimum floor, not a representative score.
You can, but you are trusting a number you did not measure on a tool you did not pick. The online service will be accurate enough on clean, modern, single-language, single-column documents, which is a large slice of the use case, but it will fail quietly on the edges. If the documents have small text, two languages, or a table layout, the spot-check on one sample takes ten minutes and saves you an hour of cleaning up wrong text in the final output. For high-stakes documents, the test is not optional.
First, identify the failure mode from the patterns above. If it is a deskew or crop problem, fix the preprocessing and re-run; the character engine is probably fine. If it is small-text misses, re-scan at a higher resolution rather than a higher-DPI setting that does not actually change the input, and check whether the tool has a zoom-for-small-text option. If it is a language drop, change the language setting and re-run just that page. If it is garbled or mojibake characters, the encoding was guessed wrong; check the tool’s encoding option or the companion guide on garbled scanned PDF output for the cleanup steps. Re-scanning the same image with a different tool is usually the fastest fix when the failure is a layout or order problem, because the new engine will make different mistakes.
Related reading: annotate highlight notes in pdf for study, annual inventory count ocr records retention, attach pdf to text message.
What metrics should I use to evaluate OCR accuracy?
Focus on Character Error Rate (CER) and Word Error Rate (WER). Run the scanner on three test documents: a printed text page, a handwritten note, and a low-resolution scan. Compare the output text against your original manually. A low CER indicates high precision, but ensure the app preserves formatting like line breaks and columns.
How do I test OCR on different document types?
Test at least four categories: clean digital-born PDFs, scanned printed text, handwritten notes, and images with background noise. Use the same lighting and resolution for each test. Check if the app handles complex layouts like tables or multi-column articles correctly. Note which file formats produce the cleanest text without manual correction.
What is the impact of image quality on results?
Low-resolution images under 300 DPI significantly reduce accuracy. Test your app with 150 DPI, 300 DPI, and 600 DPI versions of the same document. Observe how the app handles blurry edges or skewed pages. Ensure it includes preprocessing features like skew correction and contrast adjustment. Poor input quality will always degrade output, regardless of the algorithm used.
How should I verify the output text?
Automate the comparison by converting both the original and OCR outputs to plain text. Use a diff tool to highlight discrepancies. Manually review a random 10% sample to catch context errors that metrics might miss. Pay attention to punctuation and special characters. If the app supports layering, check if the original PDF remains searchable alongside the extracted text layer.