How does face search work?

It searches geometry, not photos

The system first locates the face in the photo. Then it places dozens of reference points on it: eye corners, the centre of each pupil, nose tip, nostrils, mouth corners, chin, brow ends.

It then measures the ratios between those points — the distance between the eyes relative to the width of the face, the length of the nose relative to the eye-to-mouth gap, the angle of the jaw, dozens of them. On their own the ratios mean nothing; together they form a list of numbers, called a face signature or embedding.

What gets searched is that list of numbers, not your photo. Because the ratios are relative distances rather than absolute pixels, they survive a great deal. Shrink the photo, crop it, convert it to black and white, put a filter on it — the ratios between the points stay largely the same. A frame taken five years ago can still match, because bone structure barely changes.

The limit is where that geometry breaks. If the face is turned far from the camera, half-hidden by hair or a hand, or if sunglasses hide the eye corners, the system cannot build a reliable signature. When it cannot read the ratios, it does not guess — it skips that face.

How it differs from reverse image search

This is the point people confuse most. Reverse image tools like Google Images, TinEye and Bing search for the same file — copies and near-exact duplicates of the frame you upload. To find where a photo came from, where it first appeared, or whether it is being used without permission, they are fast and free. Try them first.

But their limits are just as clear. Crop the photo, re-save it or filter it, and reverse search usually returns nothing. And it cannot find another photo of the same person, because that is a different file. Google restricts matching people by their faces not as a shortcoming but as a deliberate choice; on privacy grounds it is intentionally weak at it.

Face search looks for the person, not the file. It can find the same face in a completely different photo — different clothes, different year, different background. That is what makes it powerful — and what makes it unsettling. Among free tools, Yandex leans closest to faces; you can try it from the camera icon in the search box, uploading a photo and cropping to the face. Our own face search works the same way, but over an index built only of faces.

How much the index sees

The comparison runs against an index: a collection of faces already found on public pages and turned into signatures ahead of time. The sources are open profiles, forums, news sites, blogs and public directories.

What sits outside the index matters as much as what is in it. Private accounts, pages behind a login, direct messages and sites closed to search engines are all out of reach. No legitimate face search breaks into an account to look behind it; it only sees what is already public.

This is also why the answer comes back in seconds. The slow work — turning pages into signatures — was done in advance. Comparing your photo's short list of numbers against millions of stored lists is cheap, fast arithmetic.

And there is a time dimension. The index is a snapshot of the internet at one moment. Pages come down, profiles close, new content appears. A result that shows today can be gone tomorrow, and a search that comes back empty today may find something months from now.

What the score says and does not say

Results come with a similarity score from 0 to 100. Roughly: above 85 is very strong, 70 to 85 likely, below that weak. The score tells you how close two signatures are.

What it does not tell you is identity. It is a measure of probability, not proof. Siblings, relatives and sometimes complete strangers with similar faces score highly. With identical twins the system cannot meaningfully tell them apart. A high score means 'a candidate worth checking', not 'this is definitely the person'.

How to confirm a match

Treat a match as a lead, not a verdict. Look past the score to the whole page: do the username, location, date, other photos and the story line up? If you see the same face on several independent pages with details that agree, your confidence grows.

Do not decide anything final on a single result — especially a decision you cannot take back, like accusing someone or ending a relationship. The system shows you where to look; what you found is yours to judge.

Why it sometimes finds nothing

A search can come back empty, and there are many innocent reasons. On the photo side, the most common are:

Sometimes the reason is simpler: the person has no photo on any public page. Someone who keeps their social media private is largely invisible to the indexed internet.

Either way, no result does not mean the person is not online. It only means they were not found in indexed public sources. Absence is not evidence.

• There is no clear face; it is too small, too dark, or shot from the side.
• Sunglasses, a mask or hair cover the key points of the face.
• A heavy filter or AI retouching has distorted the facial ratios.
• The image is heavily compressed or blurred.

Face data and ethics

A face is biometric data, protected far more tightly than ordinary personal data. In Turkey the KVKK treats it as special category data; in Europe GDPR Article 9 and in the US the Illinois BIPA law give it similar protection.

In practice the line is in use, not intent. Searching your own face, searching with someone's permission, or verifying who you are talking to on a dating app are legitimate. Following someone without their consent, harassing them, or using results for hiring, housing or credit decisions is not. These tools are for verifying the person in front of you or finding your own photos — not for tracking someone down.