What visual search is
A shopper sees something — in the street, on social media, at a friend’s home — and has no idea what it is called. A text query such as “beige belted coat” returns hundreds of loose matches, while a photo gives an exact reference point. Visual search takes an image instead of words and returns catalogue products that look like it.
How it works
- Detection. The model finds the product in the picture and crops it out — the armchair, say, rather than the whole room.
- Embedding. The image is converted into a vector — an embedding that describes shape, colour and texture.
- Nearest-neighbour search. The system looks for catalogue images with the closest vectors, just as semantic search does with the meaning of text.
- Filtering and ranking. Out-of-stock items are removed, and results are refined by category, price and size.
Multimodal models add one more step: the shopper can combine the photo with text — “the same, but in blue”.
Where it already works
| Service | What it does |
|---|---|
| Google Lens | Nearly 20 billion visual searches a month (Google, October 2024); for products it shows reviews, prices across retailers and where to buy |
| Yandex Smart Camera | In the Yandex app, the “Products” mode recognises an object and shows where to buy it |
| Ozon | Photo search in the app and on the website; clothing and footwear are searched this way most often |
| Lamoda, Wildberries, Yandex Market, Megamarket | Photo search: Lamoda since 2017, Wildberries since 2019 (updated in September 2025), Yandex Market and Megamarket since 2023 |
Yandex is the dominant search engine in Russia and several CIS markets; Ozon, Wildberries, Yandex Market and Megamarket are Russian marketplaces, and Lamoda is a Russian fashion e-commerce platform.
Where visual search pays off
- Clothing, footwear, accessories — easier to show than to describe.
- Furniture and home decor — style, shape, upholstery colour.
- DIY and spare parts — identifying a part without knowing its name or part number.
- Electronics and groceries — less benefit, because people choose by model and specifications.
Visual search works well alongside similar items and complete the look: a product found by photo becomes the starting point for recommendations.
What to check in the catalogue before launch
- Photos: the whole product, a neutral background, no text or logos on top, several angles.
- Resolution: use marketplace requirements as a benchmark — on Ozon that means at least 1000×1000 pixels.
- Attributes: colour, material, category — to filter out items that look similar but do not fit.
- Availability: results full of out-of-stock items are worse than an honest “nothing found”.
- Metrics: share of search sessions started from a photo, results CTR and conversion to purchase compared with text search.