Searchable PDF vs image-only PDF

Open two scanned invoices side by side and you cannot tell them apart. Try to select a line of text in each, and one gives you a paragraph while the other gives you a blue rectangle over a picture. That is the entire difference, and it decides whether an archive is an asset or a very tidy pile.

What an image-only PDF actually contains

A scanner produces a raster image: a grid of pixels. An image-only PDF takes that raster, embeds it as a single object per page, and adds the page dimensions. There are no characters in the file. A search for a supplier name finds nothing, not because the OCR failed but because no OCR was ever run. Copying text is impossible, a screen reader has nothing to announce, and any downstream system that expects to read content sees an opaque blob.

This is what you get by default from most multifunction office devices and from cheap bulk scanning, which is why so many organisations discover the problem years later when they try to search an archive they already paid to create.

How the text layer works

OCR reads the image, recognises characters, and records both the characters and the coordinates where each one sits on the page. Those characters are written into the PDF as real text, drawn in an invisible rendering mode so they never appear on screen. The visible page is still the scan. Underneath it, aligned word for word, is machine-readable text.

That alignment is not cosmetic. It is why highlighting a search hit lands a coloured box on the right word instead of three centimetres to the left, and why extraction tools can tell which value sits in which column of a table. A text layer with sloppy positioning will still return search hits, but everything built on top of it degrades.

The important consequence: OCR does not change the image. If the scan is skewed, faint or cropped, the text layer inherits every one of those problems. Capture quality is upstream of everything in this article.

Where PDF/A fits

PDF/A is a constrained profile of PDF designed for long-term preservation. The constraints are the point: everything needed to render the file must be inside it, so fonts are embedded, external references and JavaScript are disallowed, encryption is not permitted, and colour is defined in a device-independent way. A PDF/A file opened in twenty years should render as it does now without depending on a font or a plug-in that no longer exists.

• PDF/A comes in several parts and conformance levels. The level matters: the basic level guarantees visual reproduction, while the accessible level additionally requires logical structure and tagging.
• A PDF/A file can be image-only and still be valid. Conformance is about long-term rendering, not about searchability. The two decisions are independent.
• Later parts of the standard relax some restrictions and allow embedded attachments, which is how source files or extracted data can travel inside the archival document.
• Validate conformance rather than trusting the export setting. A file labelled PDF/A that fails validation is a normal PDF with a misleading claim in its metadata.

For most records programmes the sensible combination is PDF/A with an OCR text layer: preserved rendering, searchable content, one file.

File size is decided by the image, not the text

A common objection is that OCR bloats the archive. It does not. The text layer is plain characters, and characters are tiny next to pixels. What genuinely drives size is the capture decision: resolution, colour mode and compression.

• Bitonal capture with a compression scheme built for black-and-white documents produces very small files, and destroys faint pencil, pale stamps and low-contrast text in the process.
• Greyscale preserves faint content at moderate size and is the safe default for degraded originals.
• Colour is necessary whenever ink colour carries meaning, such as a signature, a stamp, or a red-line annotation. It is also the largest.
• Mixed raster compression separates text from background inside a colour page and compresses each appropriately, which keeps colour pages far smaller. It is lossy on the background, so it belongs on access copies rather than preservation masters.

If storage is a real constraint, the answer is a two-tier delivery: a preservation master captured generously, and a lighter access copy generated from it. Reducing quality at capture cannot be undone, and re-scanning is the most expensive way to fix anything.

Accessibility needs more than a text layer

A text layer makes content available to a screen reader. It does not make the document accessible on its own. Accessibility also requires a tagged structure so that headings, lists and table cells are identified, a defined reading order so that a two-column page is not read straight across, alternative text for meaningful images, and a declared document language. Scanned pages have none of this by default because a scan carries no structure at all.

For a bulk archive, full tagging of every page is rarely proportionate. The practical approach is to tag the documents that are published or shared externally, and to make sure the rest are at least searchable and language-declared. If accessibility obligations apply to your organisation, decide which document classes are in scope before capture, because retro-fitting structure is manual work.

When image-only is the right answer

Not every page benefits from OCR, and pretending otherwise wastes money.

• Documents that are almost entirely handwritten, where recognition would produce a text layer unreliable enough to be worse than none, because it creates false confidence in search results.
• Photographs, plates, seals and pure graphics with no meaningful text.
• Material where the legal or evidential value is the image itself and any interpretation is unwelcome.
• Very short retention material scheduled for destruction, where nobody will search it.

In every one of those cases the retrieval burden shifts to the index fields, which then have to be good enough to find the document without any full-text help. That is a legitimate design, as long as it is a decision rather than an accident.