Everyone in this category publishes a number like “99.2% accurate”. We don't, because that number cannot be honestly produced — and the fact that so many sites have one should tell you something.
Accuracy for a search engine is not one number. It depends on your photo, on whether the person is in the index at all, and on which threshold you count as a “match”. Move the threshold and you can produce any accuracy figure you like. Publishing one means picking the flattering threshold and hoping nobody asks which.
A confidence score on every single result, so you can apply your own threshold, and the measured performance of the engine. These figures are read live from our own systems as you load this page.
Cumulative totals count real searches only. Demo searches scan a fraction of the index and are excluded — including them would inflate the numbers with work we never actually did.
Above 90, the engine is very rarely wrong about two images showing the same person. Between 83 and 89, treat it as a lead worth corroborating. Between 70 and 82, a similar face — which may be the same person in poor conditions, or a stranger. Below 70 is noise, and we show it anyway rather than hiding our own error distribution.
The full explanation of what the score measures →In our experience these are the causes, roughly in order of how often they come up.
The most common failure by a wide margin. Sunglasses, a mask, a hand across the face, a heavy profile angle, or a face that occupies a tiny fraction of a large photo. We check for this before you pay, and nothing is charged if it fails.
Anything under 1 KB is rejected outright. In practice that means a thumbnail or a badly compressed screenshot — there is not enough image left to build a signature from.
Frames pulled from video are the usual culprit. The frame that looks expressive to a human is often the one that is smeared and turned away.
Colour filters are harmless. Beauty filters that slim the jaw or enlarge the eyes change the geometry the signature is built from, and they measurably lower scores.
Somebody with no public photographs cannot be found by any face search. An empty result is often the correct answer rather than a malfunction.
These are properties of the technology, not bugs we intend to fix.
The engine measures how close two facial signatures are. It has no knowledge of who anyone is, and a high score is a statement about two images, not about a person. Everything downstream of that — whether an account belongs to them, whether the page is current — is a human judgement.
This is a well-documented property of face recognition as a technology, established repeatedly in NIST's evaluations: false match rates vary by demographic group, and systems generally perform worse on some groups than others. We have not independently measured this engine by demographic, so we will not claim it is exempt. Read a borderline score with that in mind.
A strong match against a photo from fifteen years ago scores exactly like one from last week. The number cannot tell you when the image was taken, and the pages we link to often carry no reliable date either.
Siblings, and identical twins especially, can score in the range normally reserved for the same person. If a result matters and the subject has a twin, the score alone will not separate them.
The index holds public images that have been crawled. Private accounts are absent entirely, some platforms are represented far more thinly than others, and anything published since the last refresh is missing. Nobody indexes the whole web, including us.
The most useful test is one you can verify: search your own face. You already know the right answer, which makes you the only person able to grade the result honestly. If it returns pages you recognise, the engine works. If it returns nothing and you have a public footprint, that is worth knowing too — and it is the test we would run first if we were you.
You already know the right answer — so you can grade us honestly.
Search a face →