Reverse Image Search Engines: How They Work

By Bill Hartzer · Updated 27 October 2026

What a reverse image search engine is

A reverse image search engine is a search engine you query with a picture instead of words: you supply an image, and it returns copies of that image, visually similar images, or pages about what the image shows. The technical name is content-based image retrieval (CBIR) — retrieval driven by the pixels rather than by the text around them.

Each service is built for one of two jobs.

Finding copies. The engine looks for the same image — resized, cropped, recolored or lightly edited — wherever it appears. This serves copyright enforcement, provenance checks and fact-checking. TinEye is built only for this.
Understanding content. The engine recognizes what is in the picture — a landmark, a product, a plant, a line of text — and returns information or look-alike images. Google Lens and Bing Visual Search are built primarily for this.

A third category, face search, applies the same machinery to faces. It is run by separate companies, has drawn most of the regulatory action in the field, and is covered below as mechanism and law, not as a use. Which picture search engine suits a reader depends on the job, so engines are described here by design, not ranked.

How the matching works

Reverse image search has gone through three generations of technique, and large engines now combine them.

Perceptual hashing. The image is reduced to a short fingerprint designed to change very little when the image is resized or recompressed, so near-identical images produce near-identical codes. TinEye describes such a compact signature but does not publish its algorithm. Hashing is fast for copies and poor at "similar but different."
Local feature descriptors. Algorithms such as SIFT (Scale-Invariant Feature Transform, which finds distinctive keypoints and describes the pixels around each) match images that share parts, even after cropping. Descriptors are grouped into "visual words" in an inverted index, the structure text search uses.
Learned embeddings. A neural network maps each image to a vector, an embedding, so that similar images sit close together. Search becomes approximate nearest-neighbor lookup — finding the nearest stored vectors without comparing every one — using libraries such as Meta's FAISS, released in 2017.

The most concrete public account of a major engine is Microsoft's 2018 KDD paper, "Web-Scale Responsive Visual Search at Bing." It reports "tens of billions of images in the index," "thousands of features for each image," and responses "in less than 200 ms." A visual-word index first cuts the candidates by about 10,000 times, one neural-network feature cuts them by about 1,000 times more, and the full feature set reranks the rest; product quantization compresses an 8 KB feature vector to about 25 bytes per image. The figures are from 2018, and the 200 ms claim covers images already in the index; the paper's own table shows a median of 1,083 ms for user uploads.

The main services and what each one indexes

No engine indexes the whole visual web, and each decides what to leave out. As of October 2026:

Google Lens launched on 4 October 2017. During 2022 it replaced the older "search by image" in Google Images as the default visual search, first in Chrome and then on the web, and Circle to Search, announced with Samsung on 17 January 2024, extends it to anything on an Android screen. It identifies objects, text, products and places and mixes the answer with ordinary web results.

TinEye, launched in beta on 6 May 2008, finds copies only, performs no facial recognition and states that it does not typically add social-media images. The free search is the shop window; the company earns revenue licensing its matching technology.

Yandex Images uses the same family of techniques; its reputation among open-source investigators for strong matching is anecdotal, and no published measurement tests it. Baidu runs an image search about which nothing was documented for this page.

Commerce visual search — Alibaba's Pailitao (2014), Pinterest, JD.com and Amazon's Shop the Look — applies the same methods to a product catalog. A result is a product, not a source.

Index sizes: what is documented and what is not

Index size is the figure most often asked for and least often given.

Documented image-index figures
Engine	Figure	Date	Nature
TinEye	2,013,406,246 images	23 Sep 2011	Company counter
TinEye	More than 5 billion	Jun 2014	Company counter
TinEye	More than 77.6 billion	Sep 2025	Company counter, via Wikipedia
Bing	"Tens of billions"	2018	Microsoft research paper
Google Images	"More than 1.1 billion"	Fiscal 2004	Google Form 10-K
Clearview AI (face search)	"Over twenty billion"	Oct 2022	Company figure, recorded by CNIL

TinEye is the only major engine with a public counter. A September 2026 review reported more than 85 billion, a secondary figure to check against the live counter. Google's figure comes from its first annual report, filed when engines competed publicly on index counts; it stopped publishing a homepage count in 2005. Even tens of billions of images is a sample, not a copy, of the visual web. An image an engine never crawled cannot be found there, which is why one picture returns different matches on TinEye, Bing and Google.

Face search: the same mechanism, applied to people

Face search engines turn reverse image search into facial recognition: they compute a biometric template — a numerical description of one face — and return other images containing that face.

Clearview AI, based in New York, sells face search to law enforcement and other clients; it is not a public engine. France's data-protection authority, the CNIL, described its pipeline in 2022: scrape publicly accessible images from social networks, websites, blogs and videos; compute for each face "a unique mathematical hash specific to the face"; store these in a searchable database; and return matching photos with their source URLs. The mechanism is the one described earlier; the object matched is what changes, and why European law treats it as biometric processing.

PimEyes is a consumer face search service sold by paid subscription. It launched in 2017 as a Polish start-up and has since passed to offshore owners, now named as EMEARobotics of Dubai. In October 2023 it added age detection intended to block searches using images of children.

Mainstream engines have kept apart from this category: TinEye performs no facial recognition at all. How Google Lens handles faces is a matter of Google policy not verified for this page, so no claim is made about it.

The regulatory record against face search

Four European regulators have fined Clearview AI, each finding that scraping faces into a biometric database breached the GDPR (the EU General Data Protection Regulation).

Fines imposed on Clearview AI
Regulator	Date	Fine	Orders and status
Garante (Italy)	10 Feb 2022	€20 million	Scraping ban; deletion
ICO (UK)	18 May 2022	£7,552,800 (stated as €9 million)	Set aside on jurisdiction 2023; jurisdiction restored on appeal 2025; outcome unconfirmed
CNIL (France)	17 Oct 2022	€20 million	Delete French data; €100,000 per day of delay after two months
Autoriteit Persoonsgegevens (Netherlands)	16 May 2024, announced 3 Sep 2024	€30.5 million	Four orders with penalties

The Italian order itemizes its total by GDPR article and quotes Clearview's claim of more than 10 billion facial images; France recorded "over twenty billion." Both counts are the company's own.

The UK case shows how contested jurisdiction is. The First-tier Tribunal set the ICO's penalty aside on 17 October 2023, holding that processing for foreign law-enforcement clients fell outside the law. The Upper Tribunal allowed the ICO's appeal in [2025] UKUT 319 (AAC), holding that "behavioural monitoring" "does not require an element of active 'watchfulness'," and remitted the case. That restores the ICO's jurisdiction but not, by itself, the penalty; the outcome was unconfirmed as of October 2026.

The four decisions total roughly €80 million. A Greek fine is widely reported but was not verified here. Clearview has contested European jurisdiction throughout, and whether it has paid any fine is not documented in these sources.

PimEyes has faced complaints, not penalties: in the UK, Germany and Illinois since 2022. A Hamburg complaint from July 2020 ended, after more than five years, in an information letter, the authority noting that the company appears to be based in Dubai and does not respond; the privacy group noyb sued the authority on 30 April 2026. No final fine against PimEyes had been found as of October 2026.

Limits and common misconceptions

"It finds every copy." It finds copies in one engine's index, and every engine crawls a different slice of the web.
"The oldest result is the original." TinEye's "oldest" sort shows the earliest copy TinEye crawled, not when the image was made.
"Similar means the same." Embedding-based engines return look-alikes. Visual similarity is not evidence of a shared source.
"Face search is just image search." The mechanism is shared; the legal treatment is not, because regulators classify face templates as biometric data.

Whether a service looks for copies or for meaning, and what it has chosen not to index, predicts its results better than any ranking.

Frequently asked questions

What is a reverse image search engine?

It is a search engine you query with an image instead of keywords, returning copies of the image, similar images, or pages about what it shows. The technique is called content-based image retrieval. TinEye is built only to find copies; Google Lens and Bing Visual Search are built to recognize the objects, text and places inside a picture.

How does reverse image search work?

The engine converts each image into a compact numerical description and compares descriptions rather than pixels. Older systems use perceptual hashes or local features such as SIFT keypoints; modern ones use neural-network embeddings searched by approximate nearest-neighbor lookup. Microsoft's 2018 paper describes Bing narrowing tens of billions of indexed images to a ranked shortlist in under 200 milliseconds.

Which reverse image search engine has the biggest index?

Nobody can say, because most engines do not publish one. TinEye is the only major engine with a public counter: more than 77.6 billion images in September 2025, per Wikipedia. Microsoft described Bing's image index as "tens of billions" in 2018. Google, Yandex and Baidu publish no current figure, and coverage matters as much as size.

What is the difference between TinEye and Google Lens?

TinEye finds copies of the same image, including resized or edited versions, and shows where each appears; it does not recognize objects or faces. Google Lens, launched in October 2017 and Google Images' default visual search since 2022, recognizes what is in a picture and blends that with web results. One answers "where else is this image?", the other "what is this?"

Why did regulators fine Clearview AI?

Clearview scraped publicly available photos, computed a mathematical template for every face and stored them in a searchable database without a lawful basis under the GDPR. Italy fined it €20 million in February 2022, France €20 million in October 2022 and the Netherlands €30.5 million in 2024. The UK's £7,552,800 penalty remains in litigation, and Clearview contests European jurisdiction.

How reverse image search engines work