OCR a PDF: Make a Scan Searchable Without Trusting Every Character

A scanned invoice is a picture of an invoice. You can read it, but you cannot search for an invoice number, copy a client name or select a paragraph. OCR, or optical character recognition, adds a text layer by guessing which letters are in the image.

The word “guessing” matters. OCR is not a scanner that knows what the page says; it is a model that compares shapes to characters. It is very good on clean, straight, high-contrast text. It struggles with handwriting, shadows, stamps, unusual fonts and tables. The useful workflow is not “run OCR and trust it”; it is “run OCR, then check the parts that matter”.

What OCR can produce

OpenToola offers two outputs. The text output gives you a .txt file with the recognised text, page by page. The searchable PDF output keeps the look of the scanned pages and adds an invisible text layer on top, so you can search and select text in a PDF reader.

The searchable PDF is usually what people want for archives. The pages look the same, but the file becomes searchable. The text layer is not visible; it sits behind the image. If the OCR misreads a word, the visible page is still correct, but the search result may be wrong.

The text output is better when you need to copy the content into a document, a database or a translation tool. It is also easier to proofread because you can compare the text with the image side by side.

Choose the right language

The OCR engine needs to know which alphabet and word patterns to expect. Choose Arabic for Arabic documents, Indonesian for Indonesian, English for English. Choosing the wrong language produces strange results: English text recognised as Arabic, or Arabic diacritics turned into punctuation.

If a document mixes languages, OCR will favour the selected one. A contract with English headings and Arabic body text may need two passes, or a manual correction of the parts in the other language. Check the result before assuming it is complete.

The language setting is not the same as the interface language. If your browser is in French but the document is in English, choose English.

Prepare the scan before OCR

OCR works best at around 300 DPI. A 150 DPI scan can be readable to a human but too soft for the engine. A 600 DPI scan is not automatically better; it can be slower and may include more noise.

Straighten the page. A tilt of a few degrees reduces accuracy, especially at the ends of lines. Crop the desk and the shadows. Increase contrast if the text is grey, but avoid crushing thin strokes into black blobs.

Remove stamps and handwriting from the area you need to recognise if possible. A signature over a line of text will confuse the model. If the stamp is part of the document, OCR may output garbage around it; ignore that part when proofreading.

If the PDF is a photo rather than a scan, consider converting it to an image and cleaning it first. A phone photo with perspective distortion is harder than a flat scan.

Run OCR on the right pages

A 200-page PDF can take a long time to OCR. If you only need one contract or one invoice, split the PDF first with Split PDF and run OCR on the relevant pages.

If the PDF already has a text layer, OCR is unnecessary and may replace good text with worse text. Try selecting a sentence in the PDF. If it selects cleanly, the document is already searchable. Use the text output only if you need to export the text.

For a mixed document, OCR the scanned pages and leave the digital pages alone. Combining the results manually is more work, but it avoids degrading pages that were already perfect.

Proofread the result

OCR errors cluster around numbers, names, dates, reference codes and similar-looking letters. 0 and O, 1 and l, 5 and S, Arabic letter forms, and punctuation are common. A wrong digit in an invoice number can break a payment or a filing system.

Open the searchable PDF and search for a word you know is on the page. If the search finds it, the text layer is working. Then search for a name or a number. If it fails, the OCR misread it.

For the text output, compare the first page, a page in the middle and the last page with the images. Do not proofread every word of a 100-page document unless the text is legally important. Focus on the fields that will be used.

What OCR does not do

OCR does not understand the document. It does not know that a column of numbers is a table, that a signature is a signature, or that a heading belongs to the next paragraph. The text output may preserve line breaks but not the logical structure.

OCR does not verify facts. A misread date is still a date in the output. If the document is a contract, a medical record or a financial statement, a human must check the critical values.

OCR does not work well on handwriting. A clear handwritten form may produce a few words; a cursive note may produce nonsense. If handwriting matters, use a specialised handwriting model or transcribe it manually.

OCR does not make a scanned PDF accessible in the full sense. A screen reader can use the text layer, but the reading order may be wrong, especially in multi-column layouts. Accessibility requires structural tagging, not just a text layer.

Searchable PDF versus text file

A searchable PDF is larger than the original scan because it contains both the page image and the text layer. The increase is usually modest, but a 100-page scan can grow noticeably.

A text file is much smaller and easier to search with command-line tools or a database. It loses the layout, the images and the page design. If you need to quote a paragraph, text is convenient; if you need to prove what the page looked like, keep the PDF.

Keep the original scan. The searchable PDF is a derived file; if the OCR is bad, you can run it again with better settings. Do not delete the original after OCR.

OCR on iPhone and Android

On iPhone, save the PDF to Files, open the OCR PDF tool in Safari, choose the file, select the language and output, then run. OCR is CPU-intensive; a long document may take several minutes and drain the battery. Keep the tab open and the screen on if the browser suspends background tabs.

On Android, the browser can read from Downloads or Drive. A phone with less memory may struggle with a 100-page scan; split the PDF first. Check the result in a PDF reader after the searchable output downloads.

Tables, columns and forms

OCR reads text in a visual order. A two-column page may be read across the columns instead of down each column, producing sentences that jump from one topic to another. A table may lose its rows and columns, even if every number is recognised correctly.

If the table is the important part, do not rely on OCR alone. Extract the numbers into a spreadsheet manually or use a table-specific tool. For a form, the labels and values may end up in separate lines; check that the value is next to the right label.

How to test accuracy

Pick three passages: the first paragraph, a passage in the middle and a passage with numbers or names. Compare the OCR text with the image word by word. Count the errors.

If the error rate is low and the critical fields are correct, the OCR is good enough for searching. If names and numbers are wrong, improve the scan or try a higher resolution. Running OCR again on the same image rarely fixes systematic errors; the input needs to be better.

Search for a phrase that spans a line break. OCR may insert a line break in the middle of a sentence; a search for the full phrase may fail even though the words are present. Search for a single distinctive word instead.

When not to OCR

Do not OCR a PDF that already has selectable text. You may replace clean text with an imperfect layer. Do not OCR a page that is only a logo or a photo with no text; it wastes time and produces noise.

Do not OCR a document you cannot legally process or share. The tool runs locally, but the file still contains sensitive information. Do not OCR handwriting and assume it will be accurate; check it or transcribe it manually.

Do not use OCR as a substitute for reading a contract. A searchable PDF is easier to navigate, but the visible page remains the authoritative version. If the OCR text and the image disagree, the image wins.

Privacy

OpenToola runs OCR in the browser. The PDF is rendered to a canvas and processed locally; it is not uploaded to a server. The language data is loaded from the site’s own files.

That matters for identity documents, medical records and contracts. It does not make a shared computer safe. Delete the text file and the searchable PDF after sending them if they contain sensitive information, and avoid OCR on a public machine.

Tools mentioned

• OCR PDF
• Image to Text
• Split PDF
• OpenToola Guides