Face Search Accuracy: How to Read Similarity Scores and Results

A face search result is a candidate, not an identification. The similarity score tells you how closely the engine thinks two faces match compared with the rest of its index; it is not a percentage chance that both photos show the same person. You confirm a match by opening the source page and finding independent details that agree, not by trusting the number.

This guide explains how to read similarity scores, why lookalikes and background matches happen, what to make of the same photo appearing under different names, why engines disagree, and what an empty or broken result means. It ends with a checklist you can run on any result.

What a similarity score means

Face search engines turn each face into a numeric description (an embedding) and measure how close two descriptions are. The score you see is that distance, rescaled into a number that is easier to read. It is useful for ranking: higher-scoring results deserve a look first. It is less useful as a verdict.

Keep these points in mind:

• Scores are engine-specific. A 90 on one service is not a 90 on another. Each provider uses its own model, scale and thresholds. Never average or compare scores across engines.
• A score is not a probability. "92" does not mean "92% likely to be the same person". Converting a score into odds needs evaluation data about how often that engine is right at that score, under conditions like yours.
• Input quality moves scores. A blurry, side-on or heavily filtered photo produces lower scores for true matches and can produce higher scores for lookalikes.
• Scores rank within one search. The most useful reading is relative: a result scoring far above everything else in the same search deserves more attention than a cluster of similar middling scores.

Each engine documents its own score ranges, so read scores against the tool that produced them. Facehunch shows every result with its source link and score, so you can always check the page behind the number.

Real accuracy claims need a test population, image conditions, a definition of a correct result and a false-positive rate. Independent evaluations of face recognition algorithms (the US National Institute of Standards and Technology runs the best known one) show that performance varies widely between algorithms and image conditions. A single impressive demo search tells you very little.

A rough guide to reading score levels

Confirm a match with the source, not the face

The face tells you two people look alike. The source page tells you who published the photo, what name is attached and in what context. Confirmation comes from the second, so work through it deliberately:

1. Open the page, not just the thumbnail. Check that the image actually appears there and that it is the main subject, not an advert or a "related posts" widget.
2. Identify who the page says is pictured. A caption may name the subject, the photographer, the author or the account owner. Those are different roles.
3. Look for independent agreement. Matching details that are not facial: the same tattoo, the same distinctive jewellery, the same pet, the same named workplace, the same set of photos posted in the same order.
4. Check the date. An older page that contains the photo is a lead on where it came from, not automatic proof of who took it. Our guide to finding an image's original source covers dating and tracing.
5. Write down what each source actually says. If you combine a name from one page, a city from another and a face from a third, you have built a profile that none of the sources made.

A useful stopping rule: decide before you search what would answer your question. If you are checking whether a match's photos are stolen, finding the same photos on another person's account answers it. If the only support for a connection is the score, mark the result unconfirmed and stop.

Lookalikes and false matches

Every face search returns some people who simply look similar. Large indexes contain many faces, and people with similar features, age, hair and expression will sit close together.

Lookalikes are more likely when:

• The input photo is small, blurry, side-on or partly covered (sunglasses, masks, hands).
• Lighting, pose and hairstyle match closely, which makes two different faces feel identical at a glance.
• The person has fairly common features for their demographic, or the photo is heavily filtered or beautified.
• You are looking at middling scores rather than one clear top result.

How to tell a lookalike from a real match:

• Compare stable features: ear shape, eyebrow shape, moles and scars, tooth alignment, hairline. Hair colour, makeup and beards change; ears rarely do.
• Look at several photos of the candidate, not one. A real match stays consistent across angles; a lookalike drifts.
• Check whether the context fits at all: language, location, age over time.

If the evidence is resemblance alone, leave the candidate unconfirmed. Don't contact someone because they look like the person in your photo; they may have no connection to it.

When the background is matching, not the face

Sometimes a result scores high because of what surrounds the face. This happens more with general reverse image search than with dedicated face search, but it can affect both.

Common causes:

• A famous landmark, mural or hotel lobby in the background.
• A recognisable logo, watermark or app overlay (a dating app frame, a TikTok username).
• The same studio backdrop used for many people, such as graduation or corporate headshots.

What to do:

1. Search the untouched image first and keep those results.
2. Search a tight crop around the face, removing logos and text overlays.
3. Label results by which version produced them.
4. If the crop changes the results completely, the first set was probably matching the background.

The background can still be evidence. If you are tracing where a whole photo was first published, the wider scene helps, so don't discard it entirely.

The same photo under different names

Finding one photo attached to several names is common and has several innocent explanations as well as a suspicious one.

Make a small context table for each result: page URL, exact caption, what role the name seems to play, date, and whether the page presents the photo as a portrait of the named person. Then look for the account with the deepest history: many years of posts, tagged friends, consistent locations. That is usually the real owner; newer accounts with the same photos and few posts are usually the copies.

If the photo is yours and is being used under other names, see what to do when a fake account uses your photos. Don't publish lists of "suspected identities" based on captions.

When search engines disagree

Google Lens, Bing, TinEye, Yandex and face search engines often return completely different results for the same photo. That is expected, not a sign that one is broken.

• Different indexes. Each engine has crawled a different part of the web. Face search engines index faces from pages that general engines treat differently.
• Different methods. Copy search looks for the same picture; face search looks for the same face in any picture. They answer different questions.
• Different inputs. Check that you uploaded the same file to each. A full photo in one and a crop in another will diverge.
• Different freshness and filters. Some engines refresh faster, and some deliberately limit results about people.

How to combine them:

1. Separate exact copies from similar images.
2. Group results that point to the same source page or account.
3. Open each unique source and record what it actually establishes.
4. Write a combined finding: the strongest source-backed conclusion and the gaps that remain.

A large cluster of results may all trace back to one irrelevant source, while a single result on one engine may be the decisive one. Count sources, not results. For choosing which engines to use, see how to choose a reverse image search tool.

Thumbnails that don't match the page

You open a result and the photo isn't there, or a different photo is. Usual reasons:

• The page was edited after it was indexed: a new profile picture, a replaced product photo, an updated article.
• The image sits in a gallery, carousel or "load more" section that hasn't loaded yet.
• The page needs a login to show the image (common for social profiles).
• The thumbnail came from a different page on the same site, such as an index or tag page.

Record the result URL, the live page URL, the date and whether the image is absent, changed or behind a login. A stale thumbnail tells you the photo was there at some point. It does not show the page publishes it now. Don't bypass logins or access controls to chase a cached image.

Results that point to deleted pages

If a result leads to a "page not found", a suspended account or an empty profile, the source has probably been removed since the engine indexed it.

• Check whether it is really gone: a 404 or "account suspended" is different from a login wall or a temporary error.
• Look for a legitimate archived copy. The Internet Archive's Wayback Machine may have a snapshot of public pages.
• Note the date and what you observed. "Search preview still visible" is not the same as "image still hosted on this page".
• If you are asking for removal, direct the request at what remains: the search engine's cached preview and the source page are handled by separate processes.

A deleted fake account can still be useful information. It suggests the photos were reused and the account was reported or abandoned.

When a search returns no results

An empty result means the engine found nothing in its index at that moment. It does not prove the photo is unique, private or genuine.

Common reasons:

• The photo is new. Indexes take time to discover new pages, from days to months.
• The photo lives behind a login or in a private account, which crawlers can't see.
• The face is hard to read: small, side-on, low light, heavy filters, sunglasses.
• The image is AI-generated, so no earlier copy exists. See how to check for AI-generated photos.
• The engine doesn't cover the relevant sites at all.

A bounded follow-up works better than repeated uploads:

1. Search the best-quality version of the photo once.
2. Try a different method (copy search if you used face search, and the reverse).
3. Try a different photo of the same person if you have one.
4. Save the empty result and the date, and stop unless the question still matters.

Uploading the same file again a few minutes later won't change what is in the index. For upload errors and other technical problems, see reverse image search not working.

Checklist for confirming a match

Run through this before you treat any result as confirmed:

• The input photo was clear, front-facing and cropped to one face.
• The result is either the same photo or the same person in clearly different photos.
• Stable features (ears, moles, scars, teeth, hairline) agree.
• I opened the source page and the photo is actually there as the main subject.
• I know what role the name on the page plays (subject, photographer, author, account owner).
• At least one non-facial detail agrees independently.
• I have not combined details from separate sources into a new claim.
• I compared results within one engine, not scores across engines.
• If the only support is the score, I have marked it unconfirmed.

Face search is for checking someone you are already talking to, protecting yourself, or finding your own photos. It is not for identifying strangers or locating people. For decisions with real consequences, such as hiring, renting or legal matters, use a proper verification process. See background checks for the difference.

Written by Facehunch with AI assistance and checked against the cited sources. Our editorial policy.