How to make a PDF searchable

A searchable scanned PDF holds two things per page. The visible content is the original scanned image, exactly as it arrived, drawn in exactly the same place. Behind it sit the recognised words, each positioned at the coordinates where that word appears in the image, drawn in an invisible text rendering mode — a PDF text mode that draws no marks at all.

So the words are present as real text with real positions, and they display nothing. Visually the output is indistinguishable from the input. Functionally the document changes character completely: Ctrl+F finds terms, a text selection highlights the right region of the image, copy yields words, and desktop search indexers can finally see inside the file.

The reason this matters for records is that nothing is retyped and nothing is re-laid-out. A recognition error cannot alter what the page shows, because what the page shows is still the photograph of the original. For archives where the appearance of the page is itself the record — signed documents, historical material, anything that might be examined — that property is the difference between an acceptable process and an unacceptable one.

Adding a text layer only adds. The images that made the file large are all still there, and you now also have per-word text and positioning data for every page. The output is therefore always larger than the input, which surprises people who expect a document containing text to be more efficient than one containing pictures.

The growth from the text itself is modest — text is compact next to scanned images. If the increase is a problem, compress the searchable output rather than compressing before recognising: the text layer is untouched by image compression, so you keep full searchability while the images shrink.

That ordering is worth stating plainly. Recognise first, compress second. Compressing first degrades exactly the pixels the recogniser needs, particularly by introducing artefacts around character edges, and you cannot recover accuracy afterwards.

The hidden layer inherits every recognition error. The consequence is narrow but specific: a misread word will not be found by a search, even though the page reads perfectly to a human. You see the right word, the index holds the wrong one, and the document appears simply not to exist for that query.

This is a different failure from a visibly wrong conversion, and it is harder to notice because there is nothing on the page to alert you. An archive can look completely successful and have gaps that only surface when someone fails to find a document they know is there.

Practical mitigations: when a search comes up empty for a document you believe exists, search for a shorter fragment or a different distinctive term from the same page, since errors cluster on particular words rather than affecting everything. If specific fields must be reliably findable — a contract number, a client name — check those on a sample of documents rather than assuming. And where a batch of scans is poor quality, the honest conclusion is sometimes that the important fields should be catalogued manually and the text layer treated as a bonus.

Testing one document with Ctrl+F confirms the tool worked. Making an archive genuinely searchable needs a little more: check that whatever indexes your files can see the new text. Desktop search on both Windows and macOS reads text layers in PDFs, but an index has to be refreshed before newly processed files appear, and replacing a file in place may not trigger a re-index immediately.

Keep the naming and folder structure you already have. A text layer makes full-text search possible; it does not replace the metadata you get from a sensible filename, and the two together are what makes an archive usable.

Finally, keep the originals if the documents have any long-term significance. The searchable version is derived and can always be regenerated, possibly better, with a future recogniser.