The audit corpus
People who run the fingerprint test contribute a coarse description of their browser to a shared dataset. This page is that dataset, seen from the outside. It exists to answer one question — how common is a configuration like yours? — and it is deliberately bad at answering any other.
Most of what follows is a description of what we cannot show you yet. That is the honest state of a young dataset, and hiding it behind a fuller-looking chart would make every other number here worth less.
Distinct hardware + browser combinations. The unit everything below counts.
Raw contributions. Higher than the line above because people re-run the test.
Submissions from the most-repeated configuration. The concentration tell.
Probe suite 1.0.0 · collected 2026-07-16 → 2026-07-17 · counts reset whenever a probe changes, because a probe that changes by one instruction is a different experiment and pooling the two would invent a trend.
Three things to know before reading the numbers
A configuration is not a person. The corpus stores no key linking two submissions, on purpose — that is what stops it becoming a tracking dataset. The cost is precise: one machine submitting twenty times and twenty machines submitting once look identical to us. Counting configurations rather than submissions is the closest honest approximation, and it is still an approximation.
Unseen never means fake. The long tail of real hardware is enormous and permanently under-sampled. If your setup is absent here, the only supported conclusion is that nobody with your setup has run the test — which is a fact about us, not about you.
Nothing high-entropy is collected. No IP, no full user agent, no fonts, no timezone, no canvas image. Those are the identifiers the audit exists to study; collecting them to measure them would defeat the point.
Why most cells below are hidden
A cell is only drawn if it clears two independent floors. They catch different failures, and a cell needs both.
- At least 20 submissions — a floor on depth
- Privacy. A cell describing a handful of people describes those people. This is the same threshold the test itself uses before telling anyone how rare their configuration is.
- At least 5 distinct configurations — a floor on breadth
- Anti-duplication, and the more interesting of the two. Without it, one enthusiastic machine re-running the test twenty times clears the first floor by itself and manufactures a cell that looks like a population. It is currently doing exactly that work: our own test browser contributed 21 submissions from a single configuration, and this floor is what keeps it out of the picture.
Right now that hides 28 of 34 cells. As more people contribute, the same floors let the same cells through — nothing here needs rewriting for the picture to fill in.
Browser
Engine family and major version. Coarse by design: the corpus keeps no full version string, because that is the high-entropy identifier it exists to study rather than to collect.
- Google Chrome 1509 configs · 48 subs
12 further values withheld, covering 20 configurations across 47 submissions. They are counted here and named nowhere: a value held by two or three contributors is a description of those contributors.
Operating system
Family and major only. An absent platform means nobody has contributed from it — on a corpus this young that is the expected state for most platforms, not a finding about them.
- Windows 1910 configs · 49 subs
- macOS 135 configs · 28 subs
8 further values withheld, covering 14 configurations across 18 submissions. They are counted here and named nowhere: a value held by two or three contributors is a description of those contributors.
GPU vendor
The WebGL vendor string. The renderer string beneath it is far more specific and is deliberately not given its own dimension: its cells are narrow enough that one could describe a single machine.
- Google Inc. (Apple)9 configs · 33 subs
- Google Inc. (NVIDIA)7 configs · 37 subs
7 further values withheld, covering 16 configurations across 25 submissions. They are counted here and named nowhere: a value held by two or three contributors is a description of those contributors.
Device class
Desktop, mobile or tablet, derived from the user agent and the touch surface. The coarsest dimension here, and so the likeliest to survive the gates while the others are still suppressed.
- desktop26 configs · 89 subs
1 further value withheld, covering 3 configurations across 6 submissions. They are counted here and named nowhere: a value held by two or three contributors is a description of those contributors.
Which combinations occur together
Everything above counts one dimension at a time. This one counts pairs — and pairs are where the interesting question lives, because a fingerprint is rarely implausible in any single field. It is the combination that does or does not exist.
Not yet. This view needs about 200 distinct configurations before it says anything true; we have 29.
The threshold is high because a pair table is only meaningful when an empty cell is informative. At 29 configurations, an empty cell means nobody has visited with that combination — which is a statement about our sample size and nothing else. Drawing the grid now would produce something that looks like knowledge and is not, and the mistake would be hard to see precisely because the grid would look right.
What this page is not
Commercial anti-bot systems build the same kind of dataset and then use it in the opposite direction: cluster millions of real visitors, measure how far a new one sits from every known cluster, and reject the outliers. That works for them because they have the millions, and because being wrong about a rare visitor is a cost they are willing to pass on to that visitor.
This page will not grow into that, however much data arrives. There is no distance score here and there will not be one — not because it is technically out of reach, but because a rarity number presented as a verdict is a claim the data cannot support. Rare configurations are mostly just rare. The audit itself answers the question this dataset cannot: not is your browser unusual, but does your browser contradict itself — which needs no population at all, and is the reason the checks are decided from your browser’s own answers rather than from this table.