1990s–2000s
Keyword and metadata search
The earliest image search engines couldn't analyze pixels at all — they indexed images the same way they indexed web pages, by matching text: filenames, surrounding page content, alt attributes, and manually applied tags. This is why an undescribed or poorly named image was effectively invisible to search, a limitation that keyword-based image search still has today whenever no one has bothered to caption the image.
Late 2000s
Content-based image retrieval arrives
Around the time TinEye launched in 2008 as one of the first commercial reverse image search engines, content-based image retrieval — analyzing an image's actual visual content rather than surrounding text — moved from academic research into production search. This is also when perceptual hashing techniques matured, enabling engines to catch resized or recompressed copies of an image, not just byte-identical files.
2011–2017
Reverse search goes mainstream
Google Image Search added a 'search by image' option in 2011. Combined with the growth of TinEye's index and the launch of Yandex's and Bing's own visual search tools, reverse image search shifted from a niche technique to something ordinary users reached for — to find a larger version of a photo, identify a product, or check if an image was being used without credit.
2017–present
Deep learning and embedding-based visual search
The rise of convolutional neural networks and, later, transformer-based vision models changed what 'similar' could mean. Instead of matching low-level visual features or near-duplicate pixels, modern engines convert images into embeddings — numeric representations that capture semantic content — so a search can return conceptually related images, not just copies of the same file. Google Lens, launched in 2017, is the clearest mainstream example: it can identify an object, plant, or landmark it has never seen an identical photo of before, something keyword or pure content-based retrieval could never do.