How to make a scanned document searchable

A scan is a photograph of a page. You can read it, but your computer cannot: searching finds nothing, and there is no text to copy. Text recognition fixes that — and produces a file with a peculiar property, a picture of a page with invisible words laid exactly over it. Here is how it works, what makes it go wrong, and what “searchable” does and does not promise.

What a searchable scan actually is

The finished file has two layers. The visible one is your scan, unchanged. The invisible one is the recognised text, positioned word by word over the words in the picture, and written in a mode that draws nothing at all. When you search, the reader searches that hidden layer and highlights the place on the image where the word sits.

Two consequences worth knowing:

• The picture is never corrected. If the recognition misreads a word, the scan still shows what it always showed. Fixing the text changes what is searchable, not what you see — and that is the right way round, because the scan is the evidence.
• A page can only have one text layer. Add recognised text to a page that already has real text and every search finds it twice, while copying produces each sentence doubled. Any tool worth using leaves such pages alone.

Before recognising: get the picture right

Recognition reads pixels, so the quality of the pixels decides almost everything.

Resolution

About 300 dots per inch is the sweet spot for ordinary print. Below roughly 200 the letters have too few pixels to tell apart — an e and a c become the same handful of dots. Above 400 you mostly get a slower job and a larger file. If your scanner offers a dpi setting, that is the one that matters; scanning at 150 and hoping to fix it later does not work.

Straightness

Recognition works line by line, so a page that sits a couple of degrees off square is noticeably harder to read. Automatic straightening measures how well the rows of text line up at different angles and picks the best one — which means it only works when there are rows of text to line up. On a photograph or a mostly blank page it should decline to guess rather than rotate the page for no reason.

Contrast, and the trap in it

Turning a page black-and-white makes clean print sharper and is genuinely the best setting for a crisp photocopy. It is also the fastest way to destroy a document:

• Pencil disappears almost entirely.
• Carbon copies and faxes fade to blank patches.
• Coloured stamps and signatures go solid black or vanish.
• Photographs become unrecognisable blocks.

Always look at the preview before accepting a high-contrast setting, and keep the original appearance for anything faint or coloured. For a page photographed by phone, where one side is brighter than the other, a threshold that follows the light across the page keeps both ends readable where a single fixed threshold loses one of them.

What recognition can and cannot read

Printed text in a Latin alphabet is what these engines are good at. Expect trouble with:

• Handwriting — a different problem entirely, and not one ordinary text recognition solves.
• Equations, tables and forms — the characters may be read, but the structure is not; a recognised table is text in roughly the right places, not a spreadsheet.
• Decorative or very small type, and anything printed over a picture.
• Accents, which are the first thing to go on a poor scan: pequeña becomes pequefia, Café loses its acute. Tell the tool which language the document is in — it changes which words the engine expects and improves the accents noticeably.

Be wary of anything that offers to tidy up the recognition with AI. Correcting text a machine could not read means guessing at it, and a plausible guess in a document you will later rely on is worse than an obvious error.

Reviewing what came back

Good tools mark words the engine was less sure about. That is a useful hint about where to look — and no more than that. It is the engine's opinion of its own work, not a probability that the word is wrong, and words it felt confident about can still be wrong. Correct the ones that matter: names, amounts, reference numbers, dates.

Two words that get confused, and should not be:

• “Text layer verified” means the words are in the file and can be searched and copied. It is a check of the plumbing.
• “The transcription is correct” means the words match the page. No tool can tell you that — only you can, by reading it.

Keeping the original, or rebuilding the page

There are two ways to produce the finished file, and the difference is worth choosing deliberately.

Keeping the original page adds the invisible text to the file you already have, leaving every byte of the scan alone. The page looks exactly as it did, the file grows by a few kilobytes, and nothing can be lost. This is possible only when you have not changed the page.

Rebuilding the page makes a new page out of the cleaned image. Necessary the moment you crop, straighten or adjust anything — and it is also what stops an old hidden text layer surviving underneath a new one. The cost is that a rebuilt page is flattened: form fields, links, attachments, layers and tags do not survive, and a digital signature certainly does not.

Check the output, not just that it downloaded

Open the finished PDF and try three things: search for a word you know is on page one; select a line and copy it, then paste it somewhere and see whether the spaces and line breaks survived; and look at the page to confirm it appears exactly as it did before. A text layer can be present and still be positioned wrongly, in which case searching finds the word and highlights the wrong part of the page.

A note on what this is not

• Cropping is not redaction. It changes the edge of what is drawn. Never use it to remove something confidential.
• Searchable is not accessible. A proper tagged PDF has a reading structure for screen readers; a text layer is not that, though it is a great deal better than nothing.
• Searchable is not archival. Formats like PDF/A have specific requirements that a recognised scan does not automatically meet.

Doing it without uploading anything

Old paperwork, medical letters, contracts and bank statements are exactly the documents that should not be sent to a website to be read by somebody else's computer. Searchable Scan does all of the above in your browser: the recognition engine and its language packs are served from the same address and run on your device, the finished PDF is reopened and its text read back before you are offered it, and nothing is uploaded or stored.

Related: PDF Compare for finding what changed between two versions, and PDF Form Builder for adding fillable fields to a document.