Turn a scanned PDF into reviewed, searchable text

Check whether OCR is needed

Try selecting and copying one sentence. If you get readable text, direct extraction is usually faster than OCR and avoids recognition mistakes. If you can select only a page image, OCR may help. Some PDFs combine scanned and text-based pages, so a single test on the cover does not describe the entire file. Start with a small page range before processing a large document.

A quick test: is it really a scan?

Open the PDF and drag across a sentence. If the words highlight individually, the file has a text layer and you do not need OCR; use PDF to Text or PDF to Word with the text-layer method for the most accurate result. If the whole page highlights as one block, or nothing highlights, the page is a picture. Mixed documents exist, so test a page in the middle and the last page, not only the cover.

Improve the source before recognition

Correct obvious rotation, choose the language that matches the document and use a clear scan. Shadows, handwriting, faint photocopies and very small type can lower recognition quality. Local OCR downloads an engine and language data, which may take time on the first run. Those network downloads are different from uploading your selected document; the recognition itself runs in the browser.

Getting a better scan

The biggest gains come before OCR. A scan at about 300 dpi, held straight, evenly lit and without shadows near the spine, can be recognized far more accurately than a tilted photo taken in poor light. Black text on a plain background works best. Color backgrounds, watermarks behind the text and very light photocopies reduce accuracy, as do handwriting and stamps over text.

• Use a flatbed scanner or a scanning app with automatic edge detection rather than a plain camera photo.
• Keep pages flat and straight; rotate sideways pages before OCR.
• Choose the language that matches the text. The OCR Correction tool offers English, Spanish, French and German.
• Use the high-detail option for small or faint print, knowing it needs more memory.

Review the content that carries risk

Proofread names, dates, invoice totals, account identifiers and table columns against the scan. Confusion between O and 0, l and 1, or a decimal point and a speck can materially change meaning. Correcting the extracted text does not change what the scan visibly shows. Keep the original scan as evidence of the source, particularly when using the text for records, research or an important submission.

Proofreading efficiently

You will rarely have time to proofread every word, so prioritize by consequence. Check the text a mistake would damage most: names, dates, amounts, account and invoice numbers, addresses and anything in a table. A systematic list of confusions helps: 0 and O, 1 and l and I, 5 and S, rn and m, a comma and a full stop. In the correction tool you edit the recognized text for each page in a box, so fix errors there before exporting rather than in the finished file.

Choose the right export

Use TXT for simple searching or analysis and DOCX when you need an editable text-first document. A searchable PDF keeps the visual page and adds a text layer; that does not restore the original typography or create an accessible tagged document. The correction workflow uses page-level text rather than reconstructing exact word geometry, so selected text can be out of position. Test search and copy behavior in a second viewer.

Choosing between TXT, DOCX and a searchable PDF

• TXT for searching, analysis or pasting into another application. Smallest and simplest.
• DOCX when you need an editable, text-first document. Expect to tidy columns and tables.
• Searchable PDF when you must keep the page looking exactly like the scan but want to find words with Ctrl+F. The corrected text is stored as an invisible layer behind the scan image.

Avoid overpromising accuracy

OCR is a starting point, not a verified transcription. A document can look clean and still contain recognition errors. Test one representative dense page, one table and one poor-quality scan before choosing a workflow for the full file. If precise layout, complex scripts, handwritten content or regulated archival requirements matter, use specialist software and independent review.

What OCR cannot do well

Handwriting, stamps overlapping text, very low-contrast photocopies, and unusual scripts are poor candidates. For a few critical facts, typing them in by hand while looking at the scan is more reliable than trusting recognition. For a large archive or formal records, use specialist software with verification built in, and have a person review a sample.

A step-by-step run-through

Here is a reliable sequence for a typical ten-page scanned document.

• Open the PDF and run the selection test to confirm it is image-only.
• Choose the OCR Correction tool, select the first two pages and the correct language, and run recognition on those pages only.
• Read the recognized text against the scan. If errors are frequent, stop and improve the source: rescan or rotate pages.
• If the sample is good, process the remaining pages, then proofread names, dates and amounts on every page.
• Export a searchable PDF to keep the visual page, or TXT or DOCX if you need the words.
• Open the export in a second viewer, search for a few known phrases and check they are found on the right pages.

Questions people ask

Why is the first run slow? The OCR engine and language data are downloaded from this site the first time and then reused from your browser’s cache, so later runs start faster.

Is my scan uploaded for recognition? No. The recognition runs in your browser tab. The downloads are the software and language files, not your document.

Why does search find a word but highlight the wrong place? A searchable PDF made from corrected page-level text is positioned approximately. It is reliable for finding pages and phrases, less so for exact word geometry.

Can OCR read handwriting? Not reliably. Printed text is the use case. For handwritten notes, plan to type the important parts yourself.

Privacy and handling of scanned records

Scanned documents often contain the most sensitive material people handle: identity documents, medical letters, bank statements and signed contracts. Local recognition means the pages are processed in your browser, not sent to a recognition service, which is a meaningful reduction in exposure. The extracted text, however, is as sensitive as the scan, and it is easy to leave copies in a downloads folder, a cloud-synced directory or an email draft.

Decide where the outputs will live before you start, delete working copies you no longer need, and avoid pasting recognized text into chat tools or documents shared more widely than the original. If your organization classifies documents, apply the same classification to the text export as to the scan.