The failure keyword search cannot fix
A customer holds a photograph of a ring, a bracket, or a jacket. They do not know the collection name, the SKU, or the trade term for the setting. They type an approximation and get nothing, then leave. Every zero-result search in a large catalogue is a customer who was ready to buy.
Visual search addresses this directly. The customer supplies the image, the system converts it to a vector representation, and the catalogue is searched by visual similarity rather than by matching words.
The catalogue is the project, not the model
The embedding model is largely a solved, commodity component. The work is on your side of the fence: every product needs consistent, adequately lit imagery from comparable angles, and the index needs rebuilding as the catalogue changes. Inconsistent product photography is the single most common reason these projects underdeliver.
It follows that the honest first question is about your image library, not about AI. If half the catalogue has one poorly lit photograph on a busy background, fixing that is the higher-return project, and we would tell you so at the audit rather than after the build.
Measure zero-result rate, and always filter by stock
Similarity accuracy is hard to evaluate and easy to game. Zero-result rate is not. It is directly measurable, it maps to lost revenue, and it is the number that should justify the spend. Track it before and after.
One non-negotiable: results must be filtered against live stock. A visual search that surfaces beautiful matches you cannot ship converts a search failure into a fulfilment complaint. The index reads the catalogue; the results read availability.
Use this before you buy or build.
- Product imagery is consistent in lighting, background, and angle across the catalogue.
- The index has a defined rebuild trigger as products are added or changed.
- Zero-result rate is baselined before the build so the gain is provable.
- Results are filtered against live stock, never surfaced from the index alone.
Visual search is mostly a data-quality project wearing an AI badge. If your product photography is inconsistent, fix that first and the search will follow.
