We tested 14 online PDF tools. Which ones actually upload your file?

Every online PDF tool tells you your files are safe. Read closely and the claims split into two kinds. One kind is a promise: we encrypt the upload, we delete it after an hour. The other kind is a physical statement: the file never leaves your computer, so there is nothing to promise. Both are easy to write on a landing page. Only one is easy to check, and we checked it.

We took the same two synthetic PDFs, merged them on 14 popular online tools plus our own as a control, and logged every request the browser sent, with its size. Then we did the whole thing again with the network cut. The numbers are below, per site, with the date each row was measured.

The results

7 of the 14 tools sent our files to their servers. 5 besides our own completed the merge without sending them anywhere, and 1 of those still worked with the wifi off. 2 could not be completed by our automation and are marked as such rather than guessed at; the notes say why. Every tool that claims local processing was found to do it.

The two test files total 268 KB. A tool that shows roughly that many kilobytes sent the whole documents; a few extra kilobytes is multipart overhead. Sites that could not be driven by the harness are listed as not testable, not as a pass.

What the numbers mean

When a row says a file was sent, the entire document left the machine and arrived at the host named. From that moment its safety rests on the promise next to it: TLS on the way in, deletion after an hour or two, staff who don't look. Those promises may well be kept; the point is that you cannot verify any of them, and the deletion clock starts after your contract, statement or medical letter is already on someone else's disk.

When a row says 0 KB and the wifi-off column says yes, nothing was promised, because nothing was sent. The merging ran inside the browser tab, and cutting the network made no difference. That is a different category of safe, and it is the reason this site exists.

How we tested

1. 1Two synthetic PDFs (268 KB together, a test pattern image and a line of text, no real data) were generated fresh for the run.
2. 2For each site, a fresh Chromium profile driven by Playwright opened the merge page, dismissed cookie banners, put both files into the file input and clicked the merge control. Every non-GET request the page made was recorded with its host and the size of its body, taken from the browser's own network layer.
3. 3A tool counts as uploading when the bodies it sent add up to at least half of the test files. In practice the uploaders sent all of it.
4. 4The offline pass loaded the page again, waited up to 30 seconds for any service worker to take control, cut the network, and repeated the job. A result counted only if a real download started or a download control appeared that wasn't there before.
5. 5Everything was run on 2026-09-15. Screenshots and the raw request logs were kept for every row.

The recording is the whole trick, and it is short:

page.on("requestfinished", async (req) => {
  if (req.method() === "GET") return;
  const headers = await req.allHeaders();
  log.push({
    host: new URL(req.url()).host,
    bytes: Number(headers["content-length"] ?? 0),
  });
});

Do it yourself in a minute

1. 1Open the tool's page and your browser's developer tools (F12, or Cmd-Option-I on a Mac). Choose the Network tab.
2. 2Add a PDF and, if needed, click the tool's go button. Look for a request whose size is about the size of your file. Click it: the host it went to is right there.
3. 3Now load the page fresh, turn off your wifi, and try again. A tool that never uploads finishes the job anyway; an upload tool stops at the first step.

Limitations

• One tool per site, on one day. Sites change; the dates are in the table and the harness is easy to re-run.
• The automation is generic. Where it could not find a file input or a go button, the site is marked not testable, which says nothing about its privacy either way.
• We measured what left the browser, not what happened afterwards. Whether an uploading site keeps its deletion promise is exactly the thing this test cannot see, which is rather the point.
• We make one of the tools in the table. It was run under the same script as the others, and the method above is the argument, not our word.