How to proofread OCR text quickly (a checklist)

OCR on a clean printed page gets nearly every word right. The trouble is that the few mistakes it makes are not spread evenly. They gather in exactly the places that matter: amounts, dates, names, reference numbers and the small marks that change a word's meaning.

Reading every line slowly is tiring and you'll still miss things. This checklist focuses your attention where errors actually happen.

Before you start: know which parts were OCR'd

When Docuflat converts a PDF, each page gets its own heading, and pages that were read with OCR are marked (OCR). Typed pages are copied out exactly and don't need proofreading for recognition errors. Start with the (OCR) pages.

1. Numbers first

Numbers are where one wrong character costs the most, and where context can't help you spot the mistake. For every amount, date, quantity, phone number and reference number:

Compare it with the original, digit by digit.
Watch for the classic confusions: 0/O, 1/l/I, 5/S, 8/B, 6/b, 2/Z.
Check that totals add up. If the document lists items and a total, add the items yourself. A mismatch points straight at a misread figure.
Check separators: 1,250.00 versus 1.250,00, and lakh-style 1,25,000.

Indian-language documents often use their own digits (१२३, ১২৩, ૧૨૩). OCR keeps them as printed, which is correct, but check them just as carefully.

2. Names and codes

Names don't follow dictionary rules, so the OCR can't use context to correct itself. Check people's names, place names, company names, survey and khasra numbers, case numbers, PAN and GSTIN-style codes, and email addresses. A single wrong letter in any of these makes the text useless for searching or matching.

3. Short words that flip meaning

In English, a few short words are easy to swap and change the meaning entirely: not/nor, no/on, shall/shali, in/is, of/or. In contracts and orders, read every clause that contains a negative or a condition.

4. Script-specific marks

Small marks carry meaning in many scripts, and they are the first things lost in a faint scan:

Devanagari (Hindi, Marathi, Nepali, Sanskrit): matras above and below the line, the anusvara dot, nukta dots (क़, ज़), and conjuncts that may split with a halant.
Bengali and Assamese: the hasanta in conjuncts, and য়/য, ড়/ড pairs. In Assamese, check ৰ and ৱ.
Gujarati: the anusvara (ં) and the short and long i signs.
Tamil: the pulli dot (க் versus க) and similar pairs such as ன/ண and ர/ற.
Telugu and Kannada: subscript consonant forms below the main letter.
Arabic and Urdu: dots above and below letters, which distinguish ب, ت, ث, ن and ي.
European languages: accents (é, è), umlauts (ä, ö, ü), ñ, and ß.

Our OCR language pages list the most common errors for each language.

5. Line ends and layout

Hyphenated words split across two lines ("infor-" / "mation"). Rejoin them so search works.
Running headers and page numbers appear in the middle of the text on every page. Delete them.
Two-column pages can interleave lines from both columns. If a paragraph suddenly stops making sense, check the layout of the original.
Stamps, signatures and handwriting produce short lines of nonsense. Delete them, or type in what they say if it matters.

A fast method for long documents

For a document of tens of pages, use an AI assistant as a second pair of eyes, but not as the judge:

Paste the OCR text into ChatGPT, Claude or Gemini.
Ask: "List every number, date, amount and proper name in this text, with the line it appears on. Don't correct anything."
Check that list against the original document. It's much faster than reading every line, because you only look at the items most likely to be wrong.
You can also ask: "Point out words that look like OCR errors, but don't guess the correction." Then look at those spots in the original yourself.

Asking the model not to correct is important. AI tools are good at noticing that something looks wrong, but they will confidently "fix" a name or a figure to something plausible and incorrect.

When to stop proofreading and rescan

If you are correcting more than a word or two in every line, the scan is the problem. Rescanning at 300 dpi, straightening the page or choosing the right language will save more time than correcting by hand. See 10 ways to get more accurate OCR.