How to Edit Text in a Scanned PDF Document (2026 Guide)

Before the how-to, a question that saves people a lot of wasted effort: do you want to change the words, or do you want to search them? Those are different jobs with different outputs, and the method everyone recommends first solves only one of them. Choosing wrong is why so many people end up with a mangled document that no longer resembles the page they scanned.

Two different outputs, and which one you want

Character recognition can produce two quite different results from the same scan.

A searchable PDF. The original page image stays exactly as it is, and an invisible text layer is written behind it, positioned word by word. The page looks completely unchanged. You can now select text, copy it, and find it with Ctrl+F. You cannot reflow or rewrite it, because the visible page is still a picture. This is what you want for a contract, an invoice, a signed form, a land registry document, or anything where the original appearance is part of the value.

An editable document. The recognised text is poured into a word processor document, where paragraphs, tables and images are reconstructed as best the software can manage. You can now rewrite anything. What you have is no longer the scanned page; it is a reconstruction of it, and on a document with columns, tables, headers or a stamp, the reconstruction is usually visibly wrong. This is what you want when you need the content and intend to produce a new document from it.

The universally recommended trick, opening a scanned PDF with Google Docs, produces the second kind. If what you actually wanted was a searchable contract that still looks like the contract, that method has thrown away the thing you were trying to keep. Decide first.

Why a scanned page has no text

A scanner produces a photograph. The PDF wraps that photograph in a page, and nothing in the file records that the dark shapes are letters. Character recognition looks at the pixels, finds the shapes, and guesses which characters they represent. It is a statistical process, it is never perfect, and its accuracy is set almost entirely by the quality of the image you hand it rather than by which software you use.

You can tell a scan from a text PDF in one second: try to select a line. If nothing highlights, or a box covers the whole page, it is an image. A useful second check is file size per page. A text page is typically tens of kilobytes; a scanned page is usually hundreds, often more than a megabyte.

Free methods, with their real failure modes

Google Drive and Google Docs

Microsoft Word

Open the scanned PDF directly from Word's File, Open dialog. Word converts it, runs recognition, and gives you an editable document. Same output category as Google Docs and generally similar quality, with the advantage that nothing leaves your machine. Word's layout reconstruction on tables is somewhat better in practice than Docs'.

On a Mac, try selecting the text first

Recent macOS versions include Live Text, which recognises characters in images on the fly. Open the scanned PDF in Preview and try dragging across a line. Quite often you can select and copy it directly without running any conversion at all. This does not modify the file or give you a searchable PDF, but if your actual need was "get this paragraph into an email," it takes three seconds and is entirely local. Worth trying before anything else if you are on a Mac.

What our own tool does, and where it falls short

Our PDF to Word converter runs Tesseract inside your browser, so the scan is never uploaded. It is the right tool for a couple of pages you would rather not send to Google. It is also weaker than the alternatives above in three specific ways, and you should know them before choosing it:

• It renders each page at roughly 144 dpi before recognising it. Recognition engines want about 300 dpi. If your scan was made at 300 or 600 dpi, we are throwing away resolution before the engine ever sees it, and accuracy suffers accordingly. A tool that reads the embedded image at its original resolution will do better.
• Each page gets ninety seconds. Past that it gives up on the page. Long documents are not a realistic target.
• English and Turkish only, and Word output only. There is no language picker, and we cannot produce a searchable PDF at all, which is the output most people with a scanned contract actually need.

When to pay, and what for

The thing paid software actually buys you is not a higher accuracy percentage on a clean page, where everything is roughly equivalent. It buys three specific capabilities:

• Searchable-PDF output. Writing an invisible text layer behind the original image, positioned correctly, while leaving the page untouched. Acrobat Pro does this properly, and it is the single most useful function in this whole category.
• Layout and table reconstruction. Dedicated recognition software such as ABBYY FineReader is noticeably better at columns, tables and mixed layouts than the general-purpose options.
• Language coverage. Recognition depends on a trained language model. Accented Latin scripts, Cyrillic, Greek, Arabic and CJK all need the right model, and a tool with a narrow language set will produce confident nonsense rather than an error.

Pricing on all of these changes, so check the vendor's page rather than a figure in an article. A practical note: several of these are sold as a subscription, and if this is a one-off job for one document, a month of a subscription you then cancel is usually cheaper than the time you will spend fixing free output by hand.

The case for not doing this at all

A large share of people who arrive at this problem should not solve it. If your scanned document is a signed contract, a completed form, a receipt or an official certificate, converting it to editable text and re-exporting produces a document that no longer looks like the one that was signed, no longer carries the scan's own evidentiary character, and now contains whatever errors the recognition made. On a legal or financial document that is a worse outcome than the problem you started with.

What you usually want instead is to add on top: initials in a box, a date, a note explaining a discrepancy, a signature on a line. That needs no recognition at all, and works on any scan however poor. Our annotation tool and signature tool place text, shapes and signatures onto the existing page image and flatten them into the file on download. Two limits worth knowing: the markup is flattened rather than stored as PDF annotation objects, so nobody can reply to it or remove it afterwards, and added text uses the standard Helvetica fonts, which cover Latin-1 only, so accented or non-Latin characters will not render.

Why scans fail recognition, mechanically

Advice about scanning at 300 dpi is everywhere and the reasoning behind it almost never is. The engine does not care about dots per inch; it cares about how many pixels tall a character is. A 10-point character is about 0.14 inches tall, which is roughly 40 pixels at 300 dpi and about 14 at 100 dpi. Accuracy holds up well at 40 and collapses somewhere below 20, which is why 300 dpi is the conventional floor and why a page of 7-point small print may need more even at 300.

The other common failure has a specific cause worth understanding. Tesseract, and most engines, begin by converting the image to pure black and white using a single global brightness threshold. That works perfectly on a flatbed scan, which is evenly lit. A phone photograph is not evenly lit: there is usually a shadow across one side or a bright reflection from a window. A single threshold cannot serve both halves, so one side comes out solid black or washed out, and the recognition on that side produces garbage while the other side reads fine. If your output is good at the top and nonsense at the bottom, this is almost always why, and the fix is flat even light rather than better software.

Practical checklist, in order of how much difference it makes:

• Even lighting beats everything else. No shadow across the page, no flash, no window behind you.
• 300 dpi or better, and higher for small print.
• Square to the page. Engines correct small skew and struggle with perspective distortion from a photograph taken at an angle.
• The right language model for the document's language, not for yours.
• Proofread the digits. The reliable errors are l against I against 1, O against 0, and rn read as m. On an invoice or a reference number, those are the characters that matter and the ones nobody checks.

Frequently Asked Questions

Related Resources

• PDF Annotator — Add text, highlights, and notes to any PDF
• Convert PDF to Word — Extract text from PDFs
• PDF Editors With OCR
• Complete PDF Editing Guide
• PDF Annotation Tools Comparison