Visual search

What visual search is

Visual search lets customers search with a photo instead of words. The image can be their own photo, a screenshot or a picture from a social network. The shop has to translate that image into similar products without the customer being able to name them. That is exactly where the value sits: visual search solves cases in which the customer does not know the term for what they are looking for.

How it works technically

The image is converted into a vector, a sequence of numbers describing its visual content. Because similar images produce similar vectors, similarity becomes computable. The search then runs not over words but over proximity in vector space.

For that to work, every product needs such a vector. Product images are therefore pre-processed and the vectors stored in a vector database. New products and replaced images must pass through that step again, otherwise they cannot be found through visual search. Notably, fine differences are hard: a photo taken in everyday conditions reliably matches on shape and colour, but material and workmanship cannot be determined from it.

Where it helps in an assortment

• Spare parts and accessories: The classic case. The customer holds the part and searches for exactly that one.
• Reordering: An item is used up or lost and the customer no longer has the label.
• Large catalogues: With tens of thousands of items, word search struggles because too many similar names exist.
• Highly visual goods: Furniture, decoration and fashion with clear traits, where shape and colour drive the choice.

What it forces on data quality

Visual search imposes a requirement that has nothing to do with search as such: product images must be consistent. The same item photographed once cut out on white and once in a scene produces different vectors and therefore two apparently different products. Anyone introducing visual search cannot avoid an image standard.

The second effect concerns legacy material. Older images with watermarks, frames or mixed backgrounds measurably hurt match quality. The usual answer is to enable visual search for part of the assortment only and keep that part current, instead of including everything and losing quality.

The benefit stays bounded and concentrates on the cases above. For standard goods with unambiguous names, such as a particular cable type or a well known branded item, visual search adds little, because the item is already found by entering its number or name. The effort does not pay off there.