Measurement · 14 September 2026

Which online file tools actually upload your files?

Almost every free online file tool says something reassuring about your privacy. We stopped reading the statements and counted the bytes instead. Across eighteen tests of twelve services, sixteen measurements ended with the entire file on somebody else's server. One tool sent nothing at all. One could not be tested without an account.

What was measured

One number, chosen because no implementation can hide it: how many bytes leave the machine, and to which server, from the moment a file is chosen. A tool that processes your document on its own servers has to send it. A tool that processes it in your browser has nothing to send.

The browser was run through a local proxy that counts the volume of every outbound tunnel. The traffic stays encrypted and is never opened — only its size is recorded, per destination host. That sidesteps every way a page could disguise an upload inside the browser, because the measurement happens below the browser.

The test file was a 3.65 MB JPEG or a 3.65 MB PDF. Each tool was loaded, cookie banners were declined where one appeared, the file was attached through the page's own file input, and the tool's own action button was pressed. A tool was recorded as uploading if a single destination host received more than 70% of the file's size.

The instrument was checked before it was trusted

A measurement nobody has validated is an opinion with numbers on it. Two tools with known answers were included as controls. Squoosh, Google's image compressor, is openly client-side: it should send nothing.TinyPNG is a server-side API with a documented endpoint: it has to send everything.

Squoosh, after the image was chosen: 2 KB — less than a single analytics ping.
TinyPNG, same image, same moment: 3 741 KB to tinypng.com.

An earlier version of this test asked the browser automation library for each request body instead, and reported TinyPNG as processing locally, because that library does not expose large payloads. It was wrong, the control caught it, and the method was rebuilt. That failure is why this section exists.

Results

Adobe Acrobat online is recorded as inconclusive rather than as local: the file could not be attached without signing in, so nothing was measured. An untested tool is not a passing tool.

Three things the numbers show that the pages do not

Every PDF service tested uploaded the document. Eleven measurements across merging, splitting and converting to images, from five different companies, and not one did the work in the browser — even though the libraries to do so have existed for years and this site runs on them.

Your file does not always go to the company whose page you are on. On ImageResizer.com the image left for freeconvert.com, a different brand. On Smallpdf it went to acloudflarestorage.com bucket — third-party object storage. Neither is improper, and both are invisible from the page you are looking at.

Advertising networks are contacted in the same breath. While the document was in flight, several of these pages were also talking to ad and session-recording hosts — doubleclick, adnxs, ad-score, clarity. The file goes to one company; the fact that you are here goes to several.

What this does not say

It does not say these tools are unsafe, dishonest or badly run. Uploading is a legitimate engineering choice, and it buys real things a browser cannot do: optical character recognition, files larger than memory, heavy compression, work that survives closing the tab. Several of these companies delete uploads within an hour and say so plainly.

It says only this: the file left your computer. Whether that matters depends entirely on the file. For a holiday photo, it does not. For a signed contract, a payslip, a medical result or a scan of a passport, it is the whole question — and nothing on the page tells you the answer.

Our own tools, measured the same way

We build browser-based file tools, so we have an obvious interest in this result. That is exactly why the method is published and the raw data is downloadable: so the measurement does not depend on believing us either.

Repeat it yourself in ten seconds

You do not need a proxy to check any of this. Load the tool and wait for the page to finish. Then disconnect — turn off wi-fi, unplug the cable, switch the phone to airplane mode. Now do the job. A tool that still works with no connection had nothing to send and nowhere to send it. One that stalls or errors was using a server. The test takes longer to describe than to perform, and it does not require trusting a privacy policy, including ours.

Limitations

One file per tool, one attempt, on 14 September 2026, from a desktop browser in Portugal, following each site's default path. Services change, and a tool that uploads today may not tomorrow. Where a site could not be exercised without an account, the result is recorded as inconclusive rather than guessed. The measurement covers the free web interface only, not the paid products or desktop applications these companies also sell.

Raw results: upload-test-2026-09.json — every host, every byte count, every note, including the failures.

Related reading: merging PDFs without uploading them ·splitting a PDF that cannot leave your computer ·what happens to files you upload