Visual search

What it is

Visual search covers any query where the primary input is an image. It includes exact product matching, similar item recommendations, text extraction from photos, and identification of objects, places, plants and animals. Most implementations now allow text to be added to the image, which makes visual search a form of multimodal search.

Why it matters

It creates demand from moments when someone cannot name what they want, which is common in fashion, interiors, DIY and spare parts. For ecommerce, it can shorten the path from inspiration to purchase, and it rewards clean, distinctive product imagery rather than keyword density alone. It also affects attribution, since visual entry points are often poorly labelled in analytics and can be undercounted.

How it works

Models convert images into numerical representations and find near matches in an index, then rank them using commerce and relevance signals. Practitioners support this with consistent photography standards, plain backgrounds for at least one image per product, multiple angles, correct image sitemaps, descriptive alt text and file names, Product structured data, and complete feeds including identifiers. On site visual search, where shoppers upload a photo to find products in your own catalogue, is a separate build using a vision search provider or in house model.

When it applies

It applies wherever appearance drives choice, or where the item is hard to describe in words. It is most useful for large catalogues with visually varied products and for businesses whose customers browse on mobile.

Examples

• A shopper uploads a photo of a patterned cushion to a retailer's site and sees matching items in stock.
• A cyclist photographs a worn brake pad to find the correct replacement part.
• A homeowner photographs a tile in a showroom and finds the same range online with prices.

How it is measured

• Image pack and image tab impressions and clicks in Search Console
• Usage and conversion rate of any on site camera or upload search feature
• Proportion of catalogue with at least one clean, plain background image at required resolution
• Match accuracy in sampled tests, measured as correct product returned in the top results