How to Redact a PDF Properly: Why Black Boxes Are Not Enough

A PDF page is not a picture. It is a set of drawing instructions, executed in order, that happen to produce a picture. A content stream might say: set the font, position the cursor, draw these glyphs, then move, then draw more glyphs, then paint an image into this rectangle.

When you draw a filled rectangle over part of that content, you add one more instruction at the end: paint a black rectangle here. Everything the earlier instructions drew is still there, underneath, in the file. The rendering engine stops showing it because something is now painted on top, but the data has not changed at all.

Anything that reads the file rather than looking at the rendered pixels will find the original text. Selecting with the cursor does this. Copy and paste does this. A search does this. Text extraction utilities do this. Automated indexers do this. This is not a subtle exploit; it is the ordinary behaviour of every PDF tool, and it is why redaction failures keep happening in documents that were carefully reviewed on screen.

The distinction is worth stating plainly: masking changes what a document looks like. Redaction changes what a document contains. Only the second one protects information.

Even teams that understand content streams frequently miss the other locations, because none of them are visible on the page.

Document metadata is the most commonly overlooked. The info dictionary stores author, title, subject, keywords, creator, and producer, and it is often populated automatically from the original application. A redacted personnel file may still name the employee in the title field. A separate XMP metadata block can carry additional structured information.

Bookmarks and outlines often quote the text of the heading they point at, so a redacted section title may survive in the navigation pane. Annotations and comments are separate objects with their own content, and a comment attached to a redacted region may restate the very information that was removed. Optional content groups, used for layers, may contain a complete copy of content that is switched off rather than deleted.

Two more are easy to forget entirely. Embedded files turn a PDF into a container, so an attached spreadsheet is still attached after the page has been redacted. And page thumbnails are separate small images that are sometimes generated at the time of creation and left in the file, showing what the page looked like before it was altered.

Any redaction process that does not address all of these is incomplete, regardless of how good the page looks.

Verification is the step that separates a process from a hope. Run all five checks before the document leaves your control.

There are two approaches that genuinely work, and they trade precision against certainty.

The first is content stream editing. This means identifying the drawing operations that render the sensitive glyphs and removing them from the page content, then removing any images in that region and re-rendering the affected area. Done correctly, the page keeps its text layer and everything else about the document remains intact. It is precise, and it requires a tool that understands the format deeply. Done incorrectly, it can leave fragments behind or corrupt the page.

The second is rasterisation, and it is the approach to choose when certainty matters more than searchability. The pages are rendered to images, the sensitive regions are covered in those images, and a new PDF is built from the resulting bitmaps. Because the original text and image data are never carried into the new file, there is nothing to recover. The cost is that the document is no longer searchable or selectable, the file is often larger, and any accessibility features are lost.

You can do the rasterisation approach with the tools you already have. Export the pages as images with PDF to JPG, cover the sensitive areas in any image editor, then rebuild the document with JPG to PDF. The result is a PDF in which the original content genuinely does not exist. It is slower than clicking a redaction button, and for a document where a mistake would be serious, that trade is worth making.

Whichever route you take, follow up by clearing metadata. The more aggressive compression mode in Compress PDF strips document properties such as author, creator, and subject, which is a useful final step once the page content has been handled.

It is worth being explicit about this, because the honest answer depends on which of the two methods above the tool uses — and ours uses the stronger one.

The Redact PDF tool renders each page to a canvas, paints the redaction boxes onto that raster, and re-embeds the page as an image. The original text and vector objects are never written to the output, so the covered content is genuinely gone. You can target it two ways: by searching for a word or phrase, or by defining a rectangular region as a percentage of the page. The text-search mode places a box over each match; the region mode applies the same rectangle to every page.

That makes it a real redaction rather than a cosmetic mask. It is suitable for removing personal data before disclosure, for journalism where a source must not be identifiable, and for any case where the information has to be unrecoverable rather than merely invisible.

There are two trade-offs you should know about. First, because a redacted page becomes an image, every other word on that page also stops being selectable text — so if you need the document to stay searchable while removing one sentence, rasterising the whole page is the wrong approach. Second, the file gets larger, because images take more bytes than text. And as with any tool, verify the output yourself: the five checks below take two minutes and are the only reliable assurance you will get.

When the consequence of failure is a disclosure, the process matters more than the tool.

Almost every public redaction incident falls into one of a small number of categories.