Treat OCR text as a first draft. The page image is what you check against. The text layer helps you search, but you must check it wherever a mistake would matter. This matters most for contracts, invoices, tax forms, IDs, and medical instructions.
Decide what to check first
Start with the words where a mistake would change the meaning. You do not need to read a 20-page document letter by letter if a quick first read can find the fields that matter.
- Names and organizations
- Dates, deadlines, and effective periods
- Prices, balances, percentages, and decimal values
- Account, policy, invoice, and reference numbers
- Addresses, email addresses, and URLs
- Negations and legal qualifiers such as “not,” “unless,” or “before”
Compare text with the visual page
Highlighted low-confidence words show you where to start, but a high score does not mean a word is right. A common word can score well even when the page shows a different name or a number. Open the original beside the recognized text, zoom in on the line, and read the whole line, not just one character.
- Find the uncertain word. Use the review list or search to jump to a low-confidence word.
- Read the line around it. The rest of the line tells you whether a number is a total, a date, or an ID.
- Compare the characters. Zoom the original and look at punctuation, decimals, and similar characters.
- Correct the text layer. Leave the visible scan as it is and fix only the recognized text, if the app lets you.
- Re-test search. Search for the corrected word and a second word from the same page.
Check numbers separately
Check numbers on their own, because one wrong character can change the result. Check separators, currency symbols, minus signs, fractions, dates, and leading zeroes. If the document has a table, compare each value with its row label instead of reading down a column on its own.
Keep the original scan unchanged until the checked, searchable PDF has been opened in the place where it will be used. If a correction turns out to be wrong, you can go back to the original.
Export only after checking
When the file is ready, give it a name that tells the checked copy apart from the original scan. For example, “vendor-agreement-2026-08-reviewed.pdf” tells you more than “scan_004.pdf.” Open the exported file and test the exact words someone will need to find later.
Example
Imagine OCR changes a client name or drops a decimal point, and the sentence still looks normal. Mistakes that look right need more checking than obvious nonsense, because nobody notices them when the text is copied into another system.
Compare uncertain text with the page, correct only what you can confirm on the page, and export a new version. Keep the original text layer or the original scan if someone else may need to check what you changed.
Questions people ask
Can I trust a high OCR confidence score? +
Use it to decide what to check first, but it does not mean the text is right. Wrong names and numbers can still get a high score, so check important fields whatever their score.
Should I correct the image or the OCR text? +
Correct the image when the visible page is wrong, such as a bad crop or rotation. Correct the text layer when the page image is accurate but recognition misread a character.
How do I know the correction was saved? +
Save or export a new copy, reopen it in a second viewer if possible, search for the corrected term, and compare the visual page.