Why you cannot search a scanned PDF

There are no words in the file

A PDF can hold two completely different things that look identical on screen. A document exported from Word holds text — actual characters, each with a position and a font — which is why you can select a sentence, search for a name and copy a paragraph. A scanned document holds a photograph of a page. There is nothing in it but pixels, in exactly the way there is nothing but pixels in a photograph of a road sign.

Every symptom follows from that one fact. Search finds nothing because there is nothing to match. Selecting text does not work because there is no text to select — the cursor draws a rectangle over a picture instead. Copying gives you nothing. The file is large for its length, because images are heavy and words are not. And converting it to Word produces either an empty document or a picture pasted into one.

The quick test: try to select a single word with your mouse. If you get a text cursor and a highlighted word, there is real text. If you get a box, it is a picture.

What OCR actually adds

Optical character recognition reads the shapes in the picture and works out which letters they are. The useful part is what happens next: the recognised words are written back into the PDF as invisible text, positioned over the words in the image. The page looks exactly as it did, because visually nothing changed — but now it can be searched, selected, copied and indexed.

That is a deliberately conservative design and it is the right one for documents. Nothing is re-typeset, so a contract still looks like the contract that was signed, and the original scan remains the visual record. Making a scanned PDF searchable does this in one pass, and usually straightens and rotates the pages as well, because the reading needed that anyway.

If what you want is an editable document rather than a searchable one, that is a different and much harder job: the text has to be extracted, the layout inferred, and the result rebuilt as paragraphs and tables. Converting to Word does that, and the result always needs checking in a way a text layer does not.

The setting that decides everything: 300 DPI

Recognition accuracy depends far more on the scan than on the software, and the single most important number is resolution. 300 DPI is the standard, and it is not arbitrary — it puts roughly 30 pixels of height into ordinary body text, which is comfortably more than the minimum needed to distinguish similar letter shapes.

At 150 DPI accuracy drops noticeably, and the errors are the insidious kind: a 1 read as an l, a 5 as an S, rn as m. Below that it degrades fast. Going above 300 rarely helps and makes files much larger — 600 DPI is worth it only for very small print or poor originals.

Three other capture settings matter almost as much. Greyscale rather than black-and-white, because a hard threshold destroys the thin parts of letters. Flat and square to the sensor, because a page photographed at an angle has letters that change shape across the page. And evenly lit — the classic failure is your own shadow falling across a page you are photographing from above.

What no OCR will read

Being honest about this saves a great deal of time. Joined-up handwriting is not recognised by general-purpose OCR, and the reason is structural rather than a matter of effort: reading print works by finding the boundaries between letters, and in cursive there are none. Hand-printed letters, block capitals and form boxes often work well, because those are discrete shapes with gaps between them.

Text that is part of a photograph — words on a sign, on a product, over a busy background — is much harder than text on paper, and heavily stylised or decorative lettering is harder still. Photocopies of photocopies lose the thin strokes that distinguish letters. And a document that has been stamped, annotated or highlighted reads worse wherever the marks overlap the text.

A good tool tells you how confident it was, and that number is there to be used. A low confidence in the middle of a sentence you can read from context matters far less than a low confidence on a digit in an account number. The dangerous errors are the confident ones, because nothing flags them.

Why this matters beyond convenience

A scanned archive with no text layer is, for practical purposes, lost. Nothing can search it, nothing can index it, no compliance system can find a name in it, and no screen reader can read it aloud — which in many jurisdictions makes publishing it a legal problem rather than an inconvenience.

It is also why scanning at the right settings the first time is worth insisting on. Re-scanning a box of documents costs far more than scanning it properly once, and no amount of processing recovers detail that was never captured.