> Reverse image search looks for the file; reverse face search looks for the face. The technical difference, why Google suppresses face results, how to read a similarity score, and which one to use for which question. **EXPLAINER · 10 SEPTEMBER 2026** # WHY GOOGLE IMAGES CAN'T FIND PEOPLE. You uploaded the photo to Google Lens and it said "no results found". That does not mean the person you are looking for is not online. You asked Google **a different question** — and it answered that question correctly. ## SHORT ANSWER **Reverse image search looks for the file; reverse face search looks for the face.** When you hand Google Lens a photo, the question you ask is "where do these pixels appear online?". The question you usually mean to ask is "where does this human appear online?" — and those are not the same question. If the person never posted that *one* photo, but did post *other* photos of themselves, image search comes back empty and face search finds them. The second reason is policy: **Google deliberately does not surface results linking a face to other photos of the same person.** So "nothing found" usually means "I did not look for that". ## IN THIS GUIDE 1. Two different questions 2. What image search actually does 3. What face search actually does 4. Google's policy wall 5. What a similarity score is and isn't 6. Side by side 7. Which one for which question 8. What neither can find 9. FAQ ## TWO DIFFERENT QUESTIONS Start with an example, because the whole issue lives inside it. You have a profile picture from a dating app. Two scenarios are possible: - **Scenario A — the photo was stolen.** A scammer is using a frame downloaded from a model's Instagram. The same file is still online. *Image search finds this, and is ideal for exactly this job.* - **Scenario B — the person is real but the frame is new.** The human across from you genuinely exists; they uploaded this photo only to that app, but they have *other* photos on Instagram, LinkedIn and a news site. *Image search finds nothing. Face search may find all three accounts.* The two scenarios look identical — a photo in hand, no results — but one is "there is nothing here" and the other is "you asked the wrong question". Without making that distinction you cannot pick a tool. ## WHAT IMAGE SEARCH ACTUALLY DOES Reverse image search engines extract a **fingerprint** from the image as a whole: colour distribution, edge and texture patterns, composition, plus in modern systems the embedding vectors produced by a network that understands image content. The strengths and weaknesses follow directly from that: - **What it is very good at:** finding copies of the same file. It catches resized, mildly compressed and slightly cropped versions too. For finding a stock photo, a news image or a stolen portfolio frame it is unmatched. - **What it cannot do:** find a *different* photo of the same person. Two photos of you taken yesterday and today are strangers at the pixel level. Their fingerprints do not match, so to the engine they are two unrelated images. - **Where it loses:** heavy filters, conversion to black and white, mirroring, different crops and low-resolution screenshots all corrupt the fingerprint. One exception is worth naming: **TinEye** can sort results by publication date. It does not search faces, but for "where was this image published first" it is sharper than any of the others — which carries direct evidentiary weight in an unauthorised-use claim. ## WHAT FACE SEARCH ACTUALLY DOES Face search has two stages, and both ignore the rest of the image. **Stage 1 — detection and alignment.** A detection model locates the faces in a photo, marks eye, nose and mouth landmarks, and rotates the face into a standard position. This is what makes faces photographed from different angles comparable at all. **Stage 2 — measurement.** A recognition model turns the aligned face into a series of numbers — typically a 128- or 512-dimensional vector. That vector is a learned compression of measurements such as eye spacing, nose-to-chin ratio and cheekbone geometry. Two vectors from two different photos of the same person land close together; vectors from different people land far apart. Comparison is done via **cosine similarity** between the two vectors, scaled into a score. The practical consequences: - **Barely matters:** clothing, background, lighting, colour, image size, filters. - **Matters a lot:** face angle (a profile shot loses half the measurement), resolution (below about 200 pixels the measurement weakens), occlusion (glasses, masks, hair), and the years in between. Trace's engine runs these two stages with two separate open models: YuNet for detection, SFace for recognition. We name them because it explains where the threshold comes from — the subject of the next section. ## GOOGLE'S POLICY WALL It is not that Google lacks face recognition; Google Photos has been grouping faces in your own archive for years. But on the *search* side it does not expose this to users: results connecting a face to other photos on the open web are systematically not surfaced. That is a choice, and an understandable one: a search engine where anyone can find anyone from a photograph becomes infrastructure for stalking and harassment. But the choice has a side effect people misread: **an empty Google Lens result does not mean "this person is not online", it means "I do not answer that question".** For the same reason you see oddities in the results: hand Google a portrait and it points you at the T-shirt on a shopping page or a set of "similar style" images. The engine is ignoring the face and interpreting the rest of the picture. ## WHAT A SIMILARITY SCORE IS AND ISN'T This section exists to correct the single most common mistake about face search. **A score is not an accuracy percentage.** "92" means neither "92% confident" nor "92% the same person". The score is a scaled distance between two face measurements — an expression of *how much these two images resemble each other*. In Trace the floor is 70, and that number is not arbitrary: it corresponds to a cosine similarity of 0.363, the threshold published by the team behind the recognition model we use. Below that value, *by the model's own definition*, the answer is "not the same person". The practical meaning of the bands comes from our own live measurement. Across sixteen real searches we found: **genuine matches appear in the 90–100 band**, while searches containing no genuine match top out at **86**. Therefore: - **90–100** — the band where genuine matches live. - **80–89** — strong resemblance. Do not decide without confirming on a second frame. - **70–79** — possible, but in practice mostly lookalike strangers. - **Below 70** — never shown. The floor used to be 50, and the result was users seeing dozens of irrelevant people's cards on a paid search. A label does not fix a list that is itself wrong. And most importantly: **no score is proof of identity.** A score tells you how similar two images are; it does not tell you who anyone is. Identical twins, close relatives, and people who simply look alike can all score highly. ## SIDE BY SIDE ## WHICH ONE FOR WHICH QUESTION - **"Is this profile picture stolen?"** → Image search. Google Lens + Yandex + TinEye. Free and better for this job. - **"Who published this image first?"** → TinEye. The only right tool. - **"Where else does this person have an account?"** → Face search. Image search cannot answer this. - **"Where is my own photo being used?"** → Both. Image search finds verbatim copies, face search finds your other photos. Step-by-step in [this guide](https://traceaifacescan.app/blog/where-are-my-photos-used-online/). - **"Is this person a scammer?"** → Image search first (a stolen photo is the most common case), face search if that is empty, then the [behavioural signals](https://traceaifacescan.app/blog/is-this-person-a-scammer/). The right method is always in the same order: **try the free one first.** Image search takes thirty seconds and answers the question outright in a meaningful share of cases. Only move to a paid face search once it comes back empty. ## WHAT NEITHER CAN FIND Being honest here is more useful than selling. There are places no kind of search reaches: - **Private accounts.** Photos on a private profile are not indexed anywhere. Neither can see them. - **Content that stays inside an app.** Dating app profiles and photos in messaging apps are not published on the open web. - **Images with no face.** Face search needs a face to measure; there is nothing to be done with a landscape or a drawing as a profile picture. - **Generated faces.** An AI-generated face has no history online. Both come back empty — but read correctly, that emptiness is information in itself. - **People who genuinely leave no trace.** The simplest and most frequent reason. An empty result may be the answer rather than a failure of the tool. ## FAQ ### What is the difference between reverse image search and reverse face search? Reverse image search looks for **the image itself**: where do these, or very similar, pixels appear online? Reverse face search looks for **the face**: it converts the face into a numeric measurement and compares it against the measurements of faces in other photos. The practical difference is that image search requires the person to have posted *that* photo, while face search only requires them to have posted *any* photo of themselves. ### Does Google Images do face recognition? Not in a way it exposes to users. Google has face recognition technology — Google Photos has grouped faces in your own library for years — but in search it does not surface results that connect a face to other photos of the same person on the open web. That is a deliberate policy choice, not a technical gap, and the reasoning is understandable: such a search engine becomes infrastructure for stalking. ### How does face search work technically? In two stages. First a detection model locates and aligns the face. Then a recognition model turns that aligned face into a vector, typically of 128 or 512 numbers, encoding measurements such as eye spacing, nose-to-chin ratio and cheekbone geometry. The cosine similarity between that vector and the vectors of indexed faces is computed and scaled into a score. Lighting, clothing and background barely affect the measurement; angle, resolution and occlusion affect it a great deal. ### Does a similarity score mean an accuracy percentage? No. The score is a scaled distance between two face measurements. In Trace, 70 corresponds to a cosine similarity of 0.363 — the threshold published by the team behind the recognition model we use; below it, the model's own definition says "not the same person". In our own live measurement genuine matches appeared in the 90–100 band while searches with no genuine match topped out at 86. And no score is proof of identity. ### Which one should I use? If your question is "where did this image come from, is it stolen?", use image search — Google Lens and TinEye are both free and better at that, and TinEye also gives you the first publication date. If your question is "where else does this face have an account?", use face search. The right approach combines them: **free image search first, face search when that comes back empty.** ### Can it tell identical twins apart? Usually not. Identical twins produce very close face measurements and both score highly; the same happens with close relatives. This is the most concrete illustration of why a score is not proof of identity: a result always has to be read alongside other information. ## SOURCES AND NOTES 1. Trace's engine uses YuNet for detection and SFace for recognition; the 70-point floor corresponds to a cosine similarity of 0.363, the threshold published by the team behind SFace. 2. Band meanings come from Trace's own measurement: across 16 live searches, genuine matches appeared in the 90–100 band while searches containing no genuine match topped out at 86 (August 2026). The floor was raised from 50 to 70 as a result. 3. Google's decision not to surface face-matching results across its search surfaces is set out in the company's own facial recognition product policy. Written by Trace Labs · Published 10 September 2026 **Responsible use.** Trace is a verification tool. Only search photographs you have the right to search — your own, or with the subject's consent. Trace returns publicly available web results with a similarity score; it is not a background check, not an identification service, and results are not proof of identity. Google, Yandex, TinEye, PimEyes and FaceCheck.ID are trademarks of their respective owners; Trace is not affiliated with any of them. **THE RIGHT QUESTION** ## SEARCH THE FACE, NOT THE FILE. If Google came back empty, your question is not unanswerable — it just needs asking somewhere else. No account, no subscription, credits that never expire.