What is browser fingerprinting?
Definition
Identifying a visitor from properties a page's JavaScript can read, such as screen, GPU, plugins, languages and timezone, combined into one profile.
No single property gives a visitor away. The combination does, and so do contradictions between properties that a real device would never produce. For automated browsers the second effect is the one that bites: a headless browser rarely fails on being unique, it fails on being inconsistent.
What a page can read
A few lines of JavaScript, with no permission prompt, return most of what fingerprinting scripts start from. This is the probe we ran, exactly as served:
const fp = {
userAgent: navigator.userAgent,
webdriver: navigator.webdriver,
languages: navigator.languages,
timezone: Intl.DateTimeFormat().resolvedOptions().timeZone,
screen: [screen.width, screen.height, devicePixelRatio],
window: [innerWidth, innerHeight],
plugins: navigator.plugins.length,
brands: navigator.userAgentData?.brands.map(b => `${b.brand} ${b.version}`),
};
const gl = document.createElement('canvas').getContext('webgl');
const dbg = gl && gl.getExtension('WEBGL_debug_renderer_info');
fp.gpu = dbg ? gl.getParameter(dbg.UNMASKED_RENDERER_WEBGL) : null;
document.body.textContent = JSON.stringify(fp, null, 1);
Real fingerprinting scripts go much further: canvas and audio rendering, font measurement, media codecs, dozens of smaller properties. The point here is how much a handful already says.
Two headless builds, one machine
We loaded that page in Google Chrome 155 started with --headless, and in the headless shell that Playwright installs, Chrome for Testing 148. Neither was given any other flag.
| Property | Chrome 155, headless mode | Chrome for Testing 148, headless shell | What a site compares it with |
|---|---|---|---|
| User agent | Contains HeadlessChrome/155.0.0.0 | Contains HeadlessChrome/148.0.7778.96 | The client hint brands and the TLS handshake |
| Client hint brands | Google Chrome 155, Chromium 155 | Chromium 148, HeadlessChrome 148 | The user agent string |
| navigator.webdriver | false | true | A person’s browser reports false |
| Languages | en-GB, en-US, en | en-US | Accept-Language and the exit country |
| Screen | 800 × 600 at 1× | 800 × 600 at 1× | Common desktop sizes |
| Window | 756 × 469 | 800 × 600 | The screen: a window as large as the screen has no browser frame |
| Plugins | 5 | 0 | Five PDF viewer entries, per the HTML standard |
| GPU via WebGL | ANGLE (Apple, ANGLE Metal Renderer: Apple M5, Unspecified Version) | None, WebGL did not start | The platform the user agent claims |
| Timezone | The machine’s own zone | The machine’s own zone | Where the exit address geolocates |
Read the first two rows together. Chrome 155 in headless mode says HeadlessChrome in its user agent and Google Chrome in its client hints, inside the same browser. The two claims contradict each other before any clever check runs. The headless shell is at least consistent: it announces itself in both places, sets navigator.webdriver to true without being asked, lists no plugins and has no WebGL. The HTML standard fixes the plugin list at five PDF viewer entries when PDF viewing is supported and an empty list when it is not, so zero is itself a signal.
The timezone row is where proxies come in. The page reads the zone of the machine running the browser, whichever country the exit address is in. A German exit with a browser reporting another zone, or a French residential address with an English-only language list, is a contradiction the address cannot fix. Set the browser’s timezone and locale to match the geo-targeted exit, or do not target a country.
Why randomising makes it worse
Changing a property to a random value does not hide a device. It produces a combination no real device has, and a rare combination is easier to follow than a common one. The W3C’s guidance for specification authors describes the same arithmetic from the other side: every characteristic that differs between users adds to the surface that can be used, in concert, to identify one, and it says plainly that a determined fingerprinter cannot be stopped by technical means alone. The workable aim for an automated browser is to be ordinary and consistent: a real engine, a normal window size, plugins and a GPU present, and locale settings that agree with the exit.
The layers before this one
Fingerprinting in the page comes after two checks that need no JavaScript at all: the TLS handshake and the header set. Both are measured in proxy.wiki studies: the TLS ClientHello fingerprint study and What HTTP clients send before you configure anything. For agents that drive a real browser, Proxies for AI agents that browse covers which identity the agent inherits from its tooling.
How this was measured. 21 October 2026, macOS on Apple silicon. Google Chrome 155.0.8059.85 launched with --headless, and Chrome for Testing 148.0.7778.96 chrome-headless-shell from Playwright’s browser cache, each with a fresh profile and --dump-dom against the probe page served from 127.0.0.1. One load per build, repeated once with an earlier version of the probe that gave the same values.