Identifying a person from a photo means narrowing the field: a reverse image search finds where a picture has already appeared online, a face-recognition tool compares one face to another, and a metadata check reads what the camera recorded. Each approach returns a candidate, never a certainty, and the gap between a candidate and a confirmed identity is where most people go wrong.
That gap matters most in the ordinary cases β a stranger in a dating profile, a face attached to a threatening message, a person in a family photo nobody labeled. What follows is the working order professionals use, the exact on-screen options in the tools that still allow face comparison, and the failure cases that produce a wrong name on a real person.
Why a reverse image search is the only safe first step
Before any face-matching, run the photo backward through a search engine. Google Lens, accessible at lens.google.com or through the camera icon in the Google app, accepts an upload or a pasted image URL and returns pages where a visually similar image appears. Bing's Visual Search does the same from bing.com/visualsearch. Neither one matches faces β they match the whole picture, which is exactly why they carry the least risk of a false accusation.
The results split roughly into three outcomes. A match on a stock photo site means the image is a stock model or a stolen asset. A match on a social profile means you've found where the picture circulates. No match means the image is likely original or cropped beyond recognition β and that a reverse search alone won't answer your question.
Run a second pass after cropping. Upload the head-and-shoulders portion alone, then the full frame, because Google Lens weighs background and clothing heavily and a cropped face can surface a completely different result set. Save the output URLs; they become the input for the cross-reference step later.
One limitation catches most people: reverse image search indexes the open web, and a photo sent in a private message, stored in a closed group, or posted with strict privacy settings will not appear. Silence from Google Lens is not evidence of anything. Treat it as a filter that removes the easy cases and forces you into a slower, more deliberate comparison.
What facial recognition actually compares when you feed it two photos
A face-recognition engine converts a face into a numeric template β a vector of measurements describing distances between the eyes, nose, jaw, and mouth β and then computes a similarity score against another template. The score is a number, not a name. Everything a tool tells you afterward is a threshold someone chose for that score.
The thresholds differ by tool, and that difference is the single most useful thing to know. Amazon's Rekognition documentation describes a default similarity threshold of 80 percent for its CompareFaces operation, adjustable between 0 and 100. PimEyes, a public face-search engine, does not publish a comparable single figure and returns ranked results instead. A score of 85 means "above whatever line this vendor drew," not "85 percent certain this is the person."
Practical consequences follow. Raise a threshold and you get fewer results with fewer false positives; lower it and you flood the output with lookalikes. For a match you intend to act on β reporting a profile, contacting someone, telling a family member you found a relative β use the highest threshold the tool allows and require a second source before you name anyone.
Face comparison fails predictably in specific conditions: heavy occlusion (masks, hair across the face, sunglasses), profile angles beyond roughly 30 degrees from frontal, low resolution where the inter-eye distance falls below a few dozen pixels, and children, whose facial geometry changes fast enough that templates from different years diverge. Any of these should push you back to the slower human comparison instead of accepting a score.
Running a face comparison in Amazon Rekognition and reading the output
Rekognition is the most documented option because AWS publishes its parameters. The console route: open the Rekognition console, choose Compare faces under the Face analysis section, upload a source image and a target image, set Similarity threshold (default 80), and read the Similarity score per matched face in the JSON response. Through the API, the same call is CompareFaces with a SimilarityThreshold parameter.
The response also returns a bounding box for each matched face. Check it. A high score on a face occupying 4 percent of a crowded photo is far weaker evidence than the same score on a portrait, because the engine had fewer pixels to measure. If the bounding box lands on the wrong person in a group shot, the score is meaningless regardless of its value.
Watch the cost side. AWS bills Rekognition per image processed, with a free tier for new accounts that covers a limited monthly volume, then per-image charges that vary by region β confirm the current rate on the AWS Rekognition pricing page, because the figure changes and I won't quote a number that may have moved. A hobbyist running fifty comparisons spends cents. A researcher running a thousand hits a budget question.
Failure handling: if CompareFaces returns no match above your threshold, do not lower the threshold until something appears. Instead test the same pair at 90 and at 70 and compare the two result sets. If a different face wins at 90 and the same face wins at 70, the evidence is unstable and you should treat the comparison as inconclusive.
The free and public face-search engines, and what they give back
PimEyes is the best-known public face-search engine: you upload a face, it returns web pages showing visually similar faces, and it displays a percentage confidence with each. Its index is built from publicly accessible images, and its own documentation states that results are candidates to verify, not confirmed identities. The percentage is not a legal or scientific identification.
Other engines occupy narrower niches. Yandex's image search has historically been strong on faces in Eastern European and Russian-language sources, and TinEye specializes in exact or near-exact copies rather than faces β useful for spotting a stolen profile picture, useless for finding a person. Social platforms run their own recognition internally but do not expose it to the public; Facebook's face-recognition feature for suggesting tags was discontinued in 2021, and Meta has since restricted related features in several jurisdictions.
Access varies by country and changes without notice. Several US states and the EU's GDPR framework impose conditions on processing biometric data, and public face search sits in an unsettled area in multiple jurisdictions. Check the rules where you live before uploading anyone's face to a third-party index β including your own.
The practical read: use a public face engine to generate candidates, then verify every candidate through a separate, human-checked path such as a named social profile with consistent history, a workplace page, or a news photo with a caption. A single engine result naming a stranger is a lead, and treating it as an answer is how people end up accusing the wrong person.
Reading EXIF and metadata before you trust the pixels
Photos taken on phones and cameras often carry EXIF data: capture date and time, device model, GPS coordinates where location tagging was enabled, and sometimes software edits. On Windows, right-click the file, choose Properties, open the Details tab, and scroll to the Origin and GPS sections. On macOS, open the image in Preview, press Command+I, and expand the Exif and GPS tabs. On Android and iOS, the stock gallery apps show a map pin and capture details under an info button.
What EXIF gives you is context, not identity. A GPS fix places the camera, a timestamp places the moment, and a device model can corroborate a story β "this was shot on the phone he said he lost" β but none of it names a face. It's most valuable as a consistency check: if a profile picture's EXIF says it was captured in 2016 on a camera the person claims never to have owned, that's a discrepancy worth noting.
Two things erase it. Most social platforms strip EXIF on upload, so a photo downloaded from Instagram or WhatsApp usually arrives clean. Screenshots never carry it at all. Messaging apps that offer "send without metadata" remove location before delivery. If the file is stripped, you cannot recover the original from the copy β you'd need the untouched original from the sender.
If metadata exists, treat it as one fact among many. A GPS coordinate matching a neighborhood is not a home address, and a timestamp matching an event is not proof the person attended. Use it to test other evidence, never as the evidence itself.
Cross-referencing a face against public records without crossing a line
Public records β property, court, business registration, professional licensing β are searchable in most US states and many other countries, and some include photographs. A professional license lookup, for example, often shows a headshot alongside a name and license number. That combination is far stronger evidence than a face-search percentage because the photo is attached to a record by the issuing body.
The workflow: take a candidate name from a face engine, search the relevant state or national licensing database, and compare the licensed photo with your image. If the licensed photo matches and the name matches, you have a public record tying a face to a name. If the license photo differs, discard the candidate.
Court records and news archives add corroboration. A local news photo with a caption naming the person at an event, cross-checked against the face you're investigating, is the kind of two-source confirmation that holds up. One caveat: people appear in news photos for reasons unrelated to any story about them, and a caption is not an identification of everyone in the frame.
Boundaries matter here more than anywhere. Do not contact a person you've identified from a photo in a way that implies accusation, do not publish a name you've derived from a face search, and do not use these methods to locate someone who has asked not to be found β in many jurisdictions that crosses into harassment or stalking. If the photo is tied to a threat, a fraud, or a missing person, hand your evidence to the platform or to law enforcement and let them verify it.
Where every method fails and how to notice it early
The most common failure is a lookalike accepted at a low threshold. Face engines rank by similarity, and similarity is not identity β two unrelated people with similar bone structure can score in the 70s. If your only evidence is a single score below the tool's default, you have nothing.
The second failure is a reused image. A profile photo lifted from a stock site or another person's account will reverse-search cleanly to the original owner, and a face engine may match it to someone entirely different. Always run the reverse search first precisely to catch this: a stock-photo match kills the investigation before it misidentifies anyone.
The third failure is age. Templates from a child's photo and the same child ten years later diverge enough that mainstream engines score them low. Old photos of adults are more stable, but weight change, facial hair, and surgery all move the score. Compare like with like where possible.
The fourth is jurisdictional. A tool that works in one country may be blocked, rate-limited, or legally restricted in another, and a result set that's legal to view in one place may be illegal to act on in another. When a result feels like it might cross that line, stop and ask a lawyer rather than guessing.
A worked example: matching a profile photo to a licensed professional
Suppose a dating profile shows a man in scrubs, and you want to know whether the photo is real. The image has no EXIF because it was downloaded from the app. Reverse search returns nothing on the full frame. Crop the face and search again β now a stock photo site surfaces the same image, credited to a photographer and licensed for commercial use.
That single result settles it: the profile is using a stock model. No face engine is needed, no name is searched, and no real person is misidentified. Time spent: about four minutes.
Now change the input. The same face appears in a local news article about a clinic opening, captioned with a name. Search the state medical board's license lookup for that name, open the record, and compare the licensed headshot. If the headshot matches the news photo and the news photo matches the profile, you have three independent sources converging β reverse search (context), news archive (name), licensing board (official photo).
If the license lookup returns a different face under that name, the news caption was wrong or the name is common. Search the board for other licensees with the same surname and compare photos against your image. The convergence test is the method; any single hit is a hypothesis.
What you may legally do with an identification, by jurisdiction
The law splits roughly into three layers: what you may collect, what you may store, and what you may publish. Collection of publicly available images is broadly permitted in the US, but Illinois's Biometric Information Privacy Act (BIPA) imposes consent and retention rules on biometric identifiers, and a 2023 Illinois Supreme Court interpretation in Cothron v. White Castle held that a claim can accrue each time data is collected. Texas and Washington have their own biometric statutes with different trigger points. In the EU, GDPR Article 9 treats biometric data used for unique identification as a special category requiring explicit consent or another lawful basis.
Storage is the layer people forget. Saving face templates from a public engine onto your own drive can itself trigger obligations under those statutes, depending on where you are and what the data is. If you're doing this for a business β vetting a candidate, checking a claim β involve your legal team before the first upload, not after.
Publication carries the highest exposure. Naming a person you've identified from a face search, especially in a public post, invites defamation and privacy claims if you're wrong, and you will be wrong some percentage of the time. Hand evidence to the platform, the police, or a lawyer. Let a body with authority make the identification.
None of this is legal advice for your situation. Biometric and privacy law is jurisdiction-specific and changes; consult a lawyer licensed where you live before you run a comparison you intend to act on.
FAQ
Does Google's reverse image search identify faces?
No. Google Lens and Bing Visual Search match whole images and visual patterns, not faces, and neither returns a name from a face. They're the right first step because they surface reused or stock images, which is often enough to end the question.
Can I identify someone from a photo with no face visible?
Sometimes, through context rather than biometrics. Clothing, a distinctive vehicle, a location visible in the background, and EXIF GPS can narrow a place and time, and public records or news archives may then connect a person to that place. Without a face or metadata, expect a slow process with a real chance of no answer.
Is a 90 percent face-match score enough to name someone?
Not on its own. Scores are threshold-dependent β Rekognition's default sits at 80 and is adjustable β and a 90 against one engine can fall below another's bar. Require a second independent source, such as a licensing record or a captioned news photo, before attaching a name to a face.
What does a face search cost?
Public engines like PimEyes charge subscription fees for full results, and cloud APIs like Rekognition bill per image with a limited free tier for new accounts. Rates change, so check the vendor's current pricing page. A small number of comparisons is cheap; bulk use is a budget line.
Can someone find me from a photo I posted?
Yes, if the photo is public and unaltered. Public face engines index open web images, so a clear portrait on an open profile is searchable by anyone who uploads it. Reducing exposure means limiting public portraits, stripping metadata before posting, and using platform privacy settings β not assuming a small audience protects you.
Why did a face search return a stranger with a similar name?
Because similarity ranking surfaces visual lookalikes, and common surnames collide. A high score on a person with no other connection to your image is a false positive. Check age, location, and any corroborating record before treating a match as real.
What should I do if the photo is tied to fraud or a threat?
Preserve the original file and its metadata, screenshot the context where you found it, and report it to the platform and to local law enforcement rather than confronting anyone. Give them the evidence and let them run the identification β acting on your own match risks naming an innocent person.
This article is general educational information, not a legal opinion or a diagnostic tool, and it makes no promises about the accuracy of any identification. Biometric and privacy rules vary by jurisdiction and change over time; consult a qualified lawyer where you live before collecting, storing, or acting on an identification. If a photo is connected to an immediate threat or to someone at risk of harm, contact local emergency services or a crisis line rather than relying on these methods.