I hid GPS coordinates in a PDF and ran it through 7 metadata removers

What I hid in the file

Removing metadata sounds like one job but it is at least five. Most tools only do the first. So I built a single PDF carrying one of each, with exact values planted in exact places — that way "did the tool remove it" has an answer rather than an impression, and anyone rerunning this gets the same number I did.

• A photograph with its EXIF intact — camera make and model, a photographer's name, and GPS coordinates at a known latitude.
• A review comment tagged with a reviewer's name and the moment they wrote it.
• An attached spreadsheet, riding along inside the document.
• A script set to run the moment the file is opened.
• Ordinary document properties — the Title, Author and Producer fields that a properties dialog shows you.

Then each returned file was read back and searched for those exact values: in its raw bytes, in its parsed object graph, and in every decoded object including the ones nothing points at. That third check matters more than it sounds, and it is the one nearly everybody skips.

What came back

The pattern is consistent. These tools clean the document layer — the fields a PDF reader shows in its properties dialog — and stop there. That is a defensible reading of the word "metadata". It is not what someone stripping a document before sending it to a stranger believes they are getting.

The two worst results are not in the table

Two of the files hid the script rather than removing it. Every tool left the open-on-load JavaScript in place. In two of them it did not even show up when I searched the file for it, because the object had been packed into a compressed stream. Still there, still runnable, and invisible to anyone checking by eye.

One tool deleted the reviewer's name from the annotation and left it elsewhere in the file. Unlinking something is not removing it. I know that failure well because I shipped it myself, and only found it by writing a third kind of check — ExifTool and qpdf both say the file is clean when it is not.

Do they do what they claim?

A tool that says it clears the author field and clears the author field is working correctly, however narrow that is. So I went back and read what each one actually promises.

IronSoftware comes out of this best. It is the least ambitious tool in the test. It names four fields, removes four fields, and never suggests it does more. That is the whole trick, and it is the only one of the six whose description of itself I could not fault.

Metadata2Go is the one that stings. Its page warns that metadata "can include sensitive details like your location". Location is the harm it picks out by name — and the file it handed back still had the coordinates sitting in it.

So are these tools bad? Mostly, no. The problem is that what they say is far narrower than what "remove PDF metadata" sounds like to someone about to email a document to a stranger. In my opinion, that's a bit misleading. And in the two places the wording reaches further, it reaches for exactly the thing it does not deliver.

"Just use the command line"

This is the reply every thread gets, and it deserves a straight answer rather than a shrug. So I ran the same file through the tools people actually name. These are not competitors to the seven above — they are the advice you get handed, which is a different question and gets its own table.

exiftool -all= does not remove anything. It is the most repeated answer on the internet, and it writes an incremental update: a new revision that hides the old values while leaving them in the file. The file gets bigger. One more command puts the author's name back:

exiftool -Author file.pdf            → Jane Doe
exiftool -all= file.pdf              → (nothing — it looks clean)
exiftool -pdf-update:all= file.pdf   → Jane Doe

ExifTool is not hiding this. It prints a warning on every single run:

Warning: [minor] ExifTool PDF edits are reversible.
         Deleted tags may be recovered!

The advice that gets repeated never carries the warning with it, and -overwrite_original with the output sent to a log is how most people meet it.

Piping the result through qpdf --linearize fixes the reversibility, and that combination is the better answer. It still only reaches the document layer: the coordinates, the reviewer's name, the attachment and the script all came back untouched.

Ghostscript left everything. Rewriting a PDF through it does not strip metadata by default — the Info dictionary is carried straight across — so the invocation people quote as a cleanup step is not one.

The one that beats me, and what it costs

mat2 removed all five. It is built by people who do this seriously, it is what the privacy community recommends, and on removal it is the equal of anything here including my own tool. If your threat model is "this document must not be traceable to me and I accept any cost", it is the right choice and I would use it myself.

The cost is the whole document. mat2 handles a PDF by rendering every page to an image and building a new file around the pictures, so the metadata is gone because the original file is gone. Across the seven corpus files that contain text, the text was destroyed in all seven — unselectable, unsearchable, invisible to a screen reader — and the output averaged 7.6 times the original size.

That is a deliberate trade and a defensible one. It is also the wrong trade for someone who wants to email a signed contract, and it is invisible to a table that only counts leaks: rasterise every page and you score a perfect row while handing back something nobody can read properly. That is why the row above carries what the tool cost as well as what it took.

Which leaves a narrower and more honest claim than "we beat everyone". On removal alone, mat2 matches me. The difference is that the document survives — and against the tools most people are actually pointed at, the coordinates come through every one of them.

Why Adobe Acrobat Pro is not scored here

Because I did not run it, and a results table is no place to guess. Acrobat is the reason this category has a standard at all: Sanitize Document went after image EXIF, annotation authorship, attachments and scripts years before anything free did, and every tool above is measured against a bar Adobe set. Leaving it out entirely would have been a quiet way of pretending otherwise, so it sits at the top of the claims table, marked untested, for now.

I skipped it because it is paid desktop software and I am not trying to contest it. That choice has a cost worth naming: every Adobe cell on this page is Adobe describing their own product, which is precisely the kind of evidence the rest of this report exists to distrust. I am not going to pretend my untested cells are worth more than anyone else's.

Fixing that takes one licence and about ten minutes. The test files are public and the checker is a single command. If you have Acrobat Pro — or you work at Adobe — run it and send me what comes back at hello@lyonite.com. I reply within 48 hours and publish what I get, credited and dated, whichever way it goes — including if it beats me on all five.

So which one should you use?

Cleaning is not the only thing that decides a tool, and one test file cannot tell you about batches, encryption or price. Short version:

The one thing none of the six do, at any price, is check their own work. They clean the file, tell you it is cleaned, and that is the end of it. Ours re-opens the file it just wrote, reads it back with the same inspector the viewer tab uses, and reports from that — including anything it skipped and why. A cleaner that only ever lists its wins is teaching you to trust it further than it has earned.

What this test cannot tell you

One file, one run, one day. A tool may behave differently on a different document, and any of these may have improved since publication — which is exactly why the test files ship alongside the finding instead of being described and kept.

On the two that changed nothing: Metadata2Go and PDF Candy returned files with Title, Author, Subject and Keywords entirely intact. Both use a field-by-field interface where removal is a separate confirmed step, and that step was applied in both cases before downloading. This was not a missed button — I am saying so up front because it is the first thing anyone will ask.

Another thing I wanted to point out is that I did not notify these tools before publishing. Nothing here is an exploit. Every result is a measurement of what a product does against what its own page says it does, taken through the public interface anyone else uses. That is a product review, not a vulnerability report, and reviews do not normally run past the vendor first. If any of them wants to send a correction or a new result, the file and the checker are public and I publish what I get.

And I am not neutral here. I build one of the tools in that table and it sits at the top of it. That is exactly why the file, the method and the checker are all public. I am not asking you to trust this site. I am handing you the file so you can check it yourself.

Questions people ask about this