Similar Products and Visual Search: How Image Search Is Changing Ecommerce

Traditional ecommerce search assumes that shoppers know what they are looking for and how to describe it.

If they need white sneakers, they type “white sneakers.” If they are looking for a cordless drill, they enter the name of the tool. If they want a particular smartphone, they search for the brand and model.

But real shopping behavior is often very different.

A customer may see a lamp they like in an interior photo, take a screenshot of a jacket on social media, photograph a chair in a café, or simply find an image of a product without knowing what it is called.

This is where a traditional search box starts to struggle.

How do you describe the exact shape of a handbag? What is the correct name for a particular shade of a dress? What should you type if you do not even know which product category the item belongs to?

One solution is visual search — searching for products using an image and finding visually similar products.

This is more than an unusual interface feature. When implemented properly, it changes the way customers interact with an ecommerce catalog.

What is visual product search?

Visual search allows shoppers to use an image instead of a text query.

A customer can upload:

• a product photo;
• a screenshot from social media;
• an image received in a messenger;
• a photo taken with a smartphone camera;
• an image from another website or source.

The system analyzes the image and searches for the most relevant products within the online store's catalog.

Imagine that a shopper sees a chair they like.

They do not know the manufacturer, collection, or even the correct name of the design style. Instead of trying to formulate a query such as “beige upholstered chair with wooden legs,” they simply upload the image.

The system compares it with the store's catalog and returns products that are visually closest to the original image.

For shoppers, this can be much more natural: sometimes showing what you want is easier than describing it in words.

How does a computer understand what is shown in an image?

A computer does not actually look at an image in the same way a person does.

Modern machine-learning models transform an image into a set of numerical features — essentially a mathematical representation of what is shown.

This representation may contain information about:

• shape;
• color;
• texture;
• composition;
• individual visual elements;
• overall style;
• semantic similarity to other objects.

The same process can be performed in advance for all product images in an ecommerce catalog.

When a shopper uploads a new image, the system creates its mathematical representation and searches for products whose representations are the closest to it.

This means that an uploaded photograph does not need to be identical to the product image in the catalog.

For example, a dress in the customer's photo may be worn by someone outdoors, while the store's product image may have been taken in a studio against a plain background. A good visual model can still recognize the visual and semantic similarity between them.

This is one of the major differences between modern visual search and older image-comparison systems that relied heavily on matching pixels or basic image properties.

Visual search and similar products are not the same thing

The technologies are closely related, but they solve different problems.

Visual search usually begins with an external image.

A customer brings an image to the store and essentially says:

“I want something like this.”

Similar products, on the other hand, usually start inside the product catalog.

A customer is already viewing a product and wants to see alternatives that are genuinely close to it.

For example:

I like this sofa, but it is too large.

These sneakers look good, but I want another option.

This bag is out of stock — show me something similar.

Instead of displaying a generic “You may also like” block, the system analyzes the product and searches for its closest alternatives.

So there are two different journeys:

image → product catalog

and

product → similar products in the catalog.

They may rely on similar technology behind the scenes, but they solve different problems for the customer.

Why traditional search is not always enough

Text search has an inherent limitation: shoppers have to use words that somehow match the way the store describes its products.

Suppose a product is listed as:

“Loft-style pendant light.”

A customer might search for:

• lamp above a dining table;
• black light fixture;
• metal lamp;
• hanging ceiling lamp;
• café-style light.

The customer and the catalog are talking about the same object, but they use different vocabulary.

A good ecommerce search engine can partly solve this problem using morphology, synonyms, typo correction and semantic search.

But some characteristics are simply difficult to express in words.

For example:

“I want roughly this shape, but a little lighter.”

For visual search, that can be a perfectly reasonable query.

Where visual search is particularly useful

The value of the technology depends heavily on the type of catalog.

Not every ecommerce business benefits from visual search to the same extent.

Fashion and footwear

Fashion is one of the most obvious use cases.

A shopper sees an outfit or a particular item they like and wants to find something similar.

However, they may not know whether the item should be described as a bomber jacket, overshirt, oversized blazer or something else entirely.

An image eliminates much of this problem.

Visual search becomes even more useful when combined with traditional filters:

find a similar dress → size M → under €100 → black.

Instead of browsing through thousands of products, the customer quickly narrows the catalog down to a relevant selection.

Furniture and home interiors

When shopping for furniture, appearance can be more important than the exact model name.

A customer may want:

• a similar sofa;
• a chair with the same general design;
• a light fixture in a similar style;
• a table with a comparable shape;
• furniture that matches an existing interior.

Trying to describe all of these visual characteristics through a text search can quickly become cumbersome.

An image communicates many of them immediately.

Jewelry and accessories

Shape and small visual details are especially important for jewelry and accessories.

A shopper may be looking for earrings similar to the ones they saw in a photo.

They may have no idea what the particular clasp, cut, shape or decorative element is called.

That can be difficult for traditional search but perfectly suitable for visual search.

Flowers and gifts

Flower shops provide another interesting use case.

A customer could upload a picture of a bouquet and ask the store to find something similar in terms of colors, shape or composition.

An exact match may not even be necessary.

Sometimes the customer is looking for the same visual mood rather than the exact same bouquet.

Auto parts and tools

This category is more complicated.

Relying solely on an image can be risky because two parts that look almost identical may have completely different specifications or compatibility requirements.

However, visual search can still be useful as the first step in narrowing down the options.

The system could identify the approximate type of part and then ask the shopper to specify the vehicle, model, year or technical characteristics.

In this scenario, visual search does not replace a technical catalog. It helps the customer reach the right section of that catalog more quickly.

That distinction is important.

Why an online store needs a similar-products feature

Consider a common ecommerce scenario.

A shopper lands directly on a product page from a search engine, advertisement or recommendation.

The product is almost right, but:

• the required size is unavailable;
• the preferred color is missing;
• the price is too high;
• the product is out of stock;
• one specification does not fit;
• the shopper simply wants to compare alternatives.

At this point, the store risks losing the visitor to the Back button.

Out-of-stock products are a particularly good example.

The shopper has already demonstrated interest, yet the store effectively tells them:

“This product is unavailable. Please start searching again yourself.”

A genuinely useful similar-products block changes that experience:

“This product is unavailable, but here are the closest alternatives.”

The shopper does not have to return to the beginning of the catalog and repeat the discovery process.

Why a traditional “Similar Products” block may not be enough

The phrase “Similar Products” has existed in ecommerce for many years.

But very different systems can sit behind that heading.

In the simplest implementation, the store displays several other products from the same category.

Open a red evening dress and the website displays five random dresses.

Technically, they are similar.

From the customer's perspective, they may not be similar at all.

Another approach is manual linking. A store manager specifies alternatives for every product.

This can work well, but it quickly becomes impractical for a catalog containing thousands or tens of thousands of items.

Rule-based systems provide another option:

same category + same brand + price within ±20%.

This is more useful, but products that satisfy these rules may still look completely different.

AI-based visual analysis adds another dimension: actual visual similarity.

Similar products are not necessarily competitors

Some online retailers hesitate to show too many alternatives on a product page.

The reasoning is understandable:

“Why distract the customer if they have already opened the product?”

And poorly configured recommendations can indeed interfere with conversion.

But a good recommendation system should not simply display more products.

It should assist the shopper at the moment when alternatives become useful.

For example:

Required size unavailable → show similar models.

Product out of stock → show the closest alternatives.

Too expensive → show visually similar products at a lower price.

Wrong color → show similar items in other colors.

In this context, similar products are not advertising. They are part of the shopping journey.

Visual search becomes particularly valuable in large catalogs

In a store with 50 products, a customer can realistically browse a significant part of the range.

In a catalog containing 20,000 products, that becomes impossible.

This creates an interesting paradox:

the larger the catalog, the greater the chance that the right product is already there — and the harder it can become for the customer to find it.

This is why ecommerce is gradually shifting from simply “displaying a catalog” toward understanding customer intent and reducing the number of options they need to process.

Filters do this through product attributes.

Search does it through words.

Recommendation systems do it through behavior and context.

Visual search does it through images.

These tools do not necessarily replace one another. The most useful experience emerges when they work together.

Can visual search replace categories and filters?

No.

There is currently little reason to abandon traditional catalog navigation entirely.

If a customer knows they need:

205/55 R16 tires,

uploading a photograph would make little sense.

If they are looking for:

a 55-inch OLED television under €1,500,

filters are likely to be much more efficient.

But when the request is:

“I like this type of lamp — find something similar,”

an image may be far more natural than text.

The future of ecommerce search is therefore unlikely to depend on a single search method.

Search is becoming increasingly multimodal.

A shopper can potentially combine:

• text;
• images;
• product attributes;
• filters;
• browsing history;
• personal preferences.

For example:

upload a photo of a chair → find similar products → show only green options → under €500 → currently in stock.

This is where the technology starts to provide practical value rather than simply looking impressive in a product demo.

Catalog quality still matters

AI cannot automatically fix a poorly maintained ecommerce catalog.

If half of the products have no images, photographs are low quality, categories are inconsistent, attributes are missing and duplicate products are everywhere, the algorithms have less reliable information to work with.

Introducing intelligent product discovery therefore also makes catalog data quality more important.

A modern search or recommendation system may use:

• product images;
• titles;
• descriptions;
• categories;
• attributes;
• brands;
• prices;
• availability;
• behavioral signals.

The better the underlying catalog, the more useful intelligent search can become.

The biggest mistake is adding AI simply because it is AI

Almost any ecommerce feature can now be marketed with the word “AI.”

But customers do not care which model is running on the server.

They care about something much simpler:

Did they find the right product faster?

If visual search requires five steps, takes too long and produces irrelevant results, browsing the catalog may still be easier.

If a similar-products block displays random items, it may be better not to have one at all.

These technologies should therefore be evaluated using ordinary ecommerce metrics:

• Do shoppers actually use the feature?
• Do they open products from the results?
• Do they find relevant items?
• Do they continue browsing?
• Do they add recommended products to the cart?
• Does the feature contribute to completed orders?

AI is useful when it reduces friction between customer intent and a purchase.

What changes inside the online store?

For many years, an ecommerce website was essentially an electronic catalog.

Home page → category → subcategory → filters → product page.

That structure is not disappearing.

But another model is emerging alongside it:

the shopper describes or shows the desired result, while the store determines how to reach the most relevant products.

That is a significant change.

The catalog stops being just a hierarchy of categories.

It gradually becomes a network of products and relationships that the system can understand:

• this product looks similar to another one;
• these products interest similar customers;
• this product can replace an unavailable item;
• this product matches the uploaded image;
• this product corresponds to the shopper's current interests.

Those relationships can make even a very large catalog considerably easier to explore.

How we are developing this in Ewonta

We are also working on this area within the Ewonta ecommerce ecosystem.

Upcoming tools are being developed to provide online stores with visual product search, visually similar product discovery and personalized recommendations.

The goal is not simply to add another button with an “AI” label.

The practical objective is to shorten the distance between:

“I want something roughly like this”

and the right product page.

The ecommerce catalog is indexed, product images and data are analyzed, and the resulting relationships can then be used for visual search and recommendations.

These capabilities are planned for use both in online stores and mobile commerce applications.

You can learn more about the AI tools being developed for Ewonta here:

AI for ecommerce — Ewonta:
https://ewonta.shop/en/ai

Visual search is unlikely to make the traditional search box disappear.

Similar products will not eliminate categories.

AI recommendations will not make filters obsolete.

But all of these technologies address one of the biggest problems of a large ecommerce catalog:

the right product may already be available, yet the shopper cannot find it.

Traditional ecommerce has spent years teaching customers how to navigate catalogs: select a category, configure filters and formulate the right search query.

New product-discovery technologies reverse that relationship.

Instead, the online store gradually learns to understand what the customer is actually trying to find.

And that may become one of the most important changes in ecommerce product discovery over the coming years.