Old scans and poor PDFs: what text recognition can still make readable today

Old PDFs become searchable when text recognition (OCR) converts the scanned image into text – and this recognition is considerably better today than many people remember from earlier attempts. Modern, AI-assisted text recognition usually reads skewed, stained and moderately resolved scans reliably; its limits lie with faded thermal paper, coarse fax copies and handwriting. So if you want to unlock a legacy collection, you don’t need to rescan the entire archive – you need an honest sorting: what the machine reads, what it reads with gaps, and what it cannot read.

In brief

• A scanned PDF is invisible to the search because it is technically a photo – only text recognition makes the content searchable.
• Modern, AI-assisted text recognition reads even skewed, stained and multi-column documents surprisingly reliably – considerably better than the software of ten years ago.
• Hard limits remain: heavily faded thermal paper, coarse fax copies, hasty handwriting and resolutions where even the eye has to guess.
• As a rule of thumb: what a person can read comfortably on screen, the machine can usually read too.
• Instead of processing the whole archive, prioritise by usefulness – manuals, reports and contracts first, dead stock last or not at all.

Why are old scans invisible to the search?

A scan is a photo of paper. To the human eye, it shows a test report with measurements – to the computer, it is an area of light and dark pixels, with no letters, no words and no meaning. That is why no search in the world can find the content of such a document, however good it is: there is simply no text for it to search.

In companies that have grown over time, this problem often affects the most valuable part of the collection. The documents from the 2000s and 2010s – maintenance reports, signed contracts, manufacturer documentation, handover reports – sit as scans on the server or on DVDs from earlier archiving drives. Text recognition, technically known as OCR (optical character recognition), closes this gap: it recognises the letters in the image and places the recognised text invisibly behind the document. The appearance stays the same, but the content becomes searchable.

What can modern text recognition really do today?

If you last tried text recognition ten years ago, you probably remember a jumble of letters – and therefore underestimate what is possible today. The current generation is AI-assisted: it no longer just recognises individual characters, but understands context, and so corrects precisely the errors that earlier software failed on. ‘Maintenance report’ no longer becomes ‘Maintenance rep0rt’, because the system knows which word makes sense at this point.

Specifically, modern recognition copes with: pages fed in at an angle, stamps and signatures across the text, punch holes and fold creases, multi-column layouts, tables with measurements and mixed documents in which printed text and form fields run into each other. Even a report that was copied on the dusty site photocopier one January and then scanned can often still be unlocked well. The result is rarely one hundred per cent error-free – but for the search it is enough if the key terms are correct, and with usable documents they almost always are.

Where are the honest limits?

No reputable provider should promise that every document will become readable. The limits can be classified well by document type:

Document	Realistic expectation
Clean scan of a printed document	Very reliably searchable, even at moderate resolution
Skewed, stained scan with stamps and punch holes	Usually easily searchable, occasional recognition errors possible
Fax copy or copy of a copy	Usable to patchy, depending on graininess and contrast
Faded thermal paper (old faxes, till receipts)	Often only fragments are legible – what has faded is gone
Clearly legible block capitals, short handwritten notes	Often usably recognisable, but needs checking
Hasty cursive, Kurrent script, scribbled notes	Cannot be reliably unlocked

A simple rule of thumb for your own assessment: open the scan on screen. What you can read comfortably yourself, the machine can generally read too. Where you have to guess, text recognition guesses too – and guessed text is more dangerous than missing text, because it creates a false sense of certainty. Reputable systems therefore flag uncertain recognitions instead of silently treating them as fact. Here, too, the basic principle of every good AI answer applies: the source is supplied with it, and if in doubt, a look at the original document decides, not the recognised text.

How do you approach a legacy collection sensibly?

The most common mistake in thinking is the demand for completeness: all or nothing. In practice, the opposite has proven its worth – prioritising by everyday usefulness. First come the collections that people actually search: manuals, setting and maintenance reports, contracts, project documentation, correspondence with current customers. Then the rest, and some of it deliberately not at all.

For old receipts – invoices, delivery notes, accounting folders – a sober classification pays off: their leading home remains accounting, with its statutory retention obligations; the ERP and the accounting system know what happened. Legacy documents belong in the knowledge base where they contain knowledge that is missing in everyday work: the how and why. A scanned exchange of letters about a complaint from 2016 is knowledge – the related copy of the invoice is merely a receipt that is better kept in accounting.

Processing is followed by tagging, and that is split in two: the system reads attributes such as customer, project, installation, date and topic from the recognised text and enters them itself. It only suggests the document type; this is confirmed by a person in the review area, because it determines who may see the document later. With clear-cut legacy collections, this is done quickly. After that, the legacy collection is not only searchable but also classified. In Goma-it’s AI knowledge base, text recognition is part of the capture process for every document – scanned legacy documents, fresh scans and mobile phone photos go through the same process, with data stored on EU servers and without your documents being used to train third-party AI models.

What do you do with documents that can no longer be read?

For the cases beyond the limit, there are three honest options. First: rescan – if the paper original still exists, a fresh scan at a good resolution often achieves more than any software acrobatics on the poor old scan. Second: add the core information by hand – for an illegible but important contract, it is often enough to record the key points as a short note on the document; then the search will at least find the matter together with where it is filed. Third: accept it – a faded fax from 1998 that nobody has needed for twenty years can stay illegible.

All that matters is that the decision is made consciously and documented, rather than illegible documents silently counting as ‘captured’. A legacy collection that is ninety per cent easily searchable, with the remaining ten per cent honestly marked as a gap, is worth more in everyday work than a supposedly complete archive that nobody trusts.

What comes after the legacy collection?

Processing old scans is typically one building block of a larger project: the legacy collection is made readable, while ongoing sources – email mailboxes, network folders, new scans – are captured automatically straight away. How these sources work together and the order in which connecting them has proven its worth is described in the overview Connecting documents and systems.

If you want a realistic assessment of your own collection – what will become easily readable, what with limitations, what not – this can be clarified quickly using a few sample documents. How such a start works, from the first review to day-to-day operation, is shown on the implementation process page.