Relevance is crucial for e-commerce search and discovery. A successful e-commerce system should not only offer relevant results but also provide a pleasant and engaging shopping experience. Still, relevance remains the foundation, as poor relevance is invariably associated with a negative customer experience.

This section will provide a glimpse into the relevance issues prevalent in leading e-commerce platforms. I will explore the challenges they face in addressing these problems to deliver a more engaging shopping experience.

Three Queries

It’s not easy to assess the quality of e-commerce search, as it involves many factors. In the following, I will examine three illustrative queries, though these examples don’t fully represent the systems’ total capabilities.

First, let’s examine the query “red wine $30.” While “red wine” alone is a general, high-frequency head query, adding “$30” turns it into a specific, low-frequency tail query. I will analyze and compare the search results for this query on Amazon, Walmart, and Ebay, observing how they have changed over a period of more than five years.

As depicted in Fig. 1, the search results for “red wine $30” have shown some improvement over the past five years, but there’s still significant potential for enhancing their relevance further. This is true for most platforms, with the exception of Walmart. However, even Walmart’s results falter in relevance at positions 4, 5, and 6 (not displayed above), where unrelated products like cosmetics and sleepwear appear instead.

This query underscores multiple prevalent challenges in query understanding, which is a critical element of e-commerce search:

1. Identifying the Head Term: Recognizing the head term of the query, like “red wine” in this instance, is important. It allows the search engine to return items under the red wine category, including those like Cabernet Sauvignon, which don’t explicitly mention “red wine” in their names. Without this, search engines might mistakenly prioritize unrelated products, such as “red wine glasses,” simply because they contain “red wine” in their product names.
2. Recognizing Attributes: Search engines often face difficulty in correctly identifying product attributes within queries. In the above example, some search platforms incorrectly feature products with a red color, partly because some items in the catalog are labeled as being “wine red” in color.
3. Interpreting Matching Criteria: The query includes “$30” as a price point, generally understood to mean close to $30 rather than exactly $30. However, the accuracy of interpreting this price criterion varies across platforms, with only Walmart’s results suggesting an accurate interpretation.

Next, let us take a closer look at two broader queries: “mother’s day gift” and “insomnia.” In these cases, users are typically looking for recommendations and advice. For example, the query “insomnia” might imply a deeper need, something along the lines of, “I’m having difficulty sleeping. Can you suggest products that might help me?”

The search results on Amazon, Walmart, and eBay offer some relevance, but they don’t always fully address the users’ needs.

The results shown above are quite narrow. For example, Ebay lists items containing “insomnia” in their names. This is very limiting. In fact, all three e-commerce platforms have an extensive array of insomnia-related products in their catalogs, including sleep aids, natural supplements, aromatherapy products, ergonomic pillows, light therapy lamps, calming teas and beverages, books on cognitive behavioral therapy, etc. Users would benefit more from a holistic understanding of these options, rather than a mere selection of products with “insomnia” in their titles.

Furthermore, displaying products prominently without adequate context can undermine the trust users place in e-commerce platforms. As shown in Fig. 2, on Amazon and Walmart, sleep aid medications or supplements appear at the top of search results. This can typically be attributed to two main reasons. First, these products are ranked higher by algorithms that use some non-transparent criteria (like the total number of all user clicks, which may not be meaningful for a particular user), creating a sense of unpredictability or randomness for the user. Second, they could be featured products from brands that bid on the keyword “insomnia.” These featured products often have lower relevance. For instance, Walmart’s recommendation of Motrin may be relevant for pain-induced insomnia, but Motrin, primarily a pain reliever, is not a dedicated sleep aid.

Three Tasks

The primary and longstanding challenge for e-commerce in improving user experience can be succinctly captured in the phrase, “tail experience is the head experience.” This encapsulates the difficulty of maintaining relevance across a vast array of queries, especially those new and obscure.

Addressing this challenge involves three key tasks: first, accurately understanding any user intent; second, acquiring a broad range of rich content that meets all possible user intents; and third, customizing the presentation of this content in ways that are uniquely engaging for each individual user.

Intent

It’s well-known that e-commerce search engines struggle with natural language queries, but why do they falter even with seemingly simple queries like “red wine $30”?

The root of the problem is their dependence on a multitude of machine learning models, each tailored for specific aspects of search. For example, just for query understanding, we might need the following ML models:

• Language Detection
• Speller
• Stemming & Lemmatization
• Query Segmentation
• Query Classification
• Entity Linking
• Query Tagging and Annotation
• Query Relaxation
• Query Rewriting

Developing these models and ensuring that they work seamlessly together is a challenge. More often than not, they fall short in capturing the full spectrum of user intents and the complex nature of products. For example, interpreting a query like “2018 red wine $30 California” requires understanding properties of wine such as type, vintage, location, and price. With each product category having its own unique attributes, scaling these machine learning models for all product types and their nuances becomes a significant challenge.

The emergence of neural information retrieval and, more recently, LLMs, could be a major breakthrough. They make it possible to use one universal model to understand all forms of user intent, whether conveyed in natural language or as a collection of keywords. Despite potential new challenges (e.g., latency), this approach offers a solution to the scalability problem in creating machine learning solutions for e-commerce search and discovery.

Content

To serve various user intent, it is critical that e-commerce systems collect and manage relevant product content effectively. This involves maintaining an extensive product catalog and taxonomy, discerning which products meet particular user intents, and offering well-informed, authoritative explanations of product relevancy.

In many e-commerce systems, less than 10% of products ever surface in search results. A major cause is the lack of detailed product information. A non-alcoholic beer, for instance, may not show up in a search for “alcohol-free drink” if it’s not explicitly described as such in its name or description.

Furthermore, matching broad or vague queries like “insomnia” to suitable products such as lavender oil or epsom salt requires specific and extensive knowledge covering a wide range of products and intents. It’s also essential to have knowledge that can explain why specific products meet particular needs (For example, epsom salt is considered beneficial for insomnia because it contains magnesium sulfate, and magnesium is known to play a role in promoting sleep. Epsom salt baths can provide a soothing experience that may help reduce stress and promote a feeling of relaxation, which can be conducive to better sleep.) Explainability is a key factor in building trust with customers and ensuring that recommendations don’t seem random or unhelpful.

Developing detailed and intricate content is a daunting task, as it usually demands specialized expertise, significant resources, and a substantial amount of time. The emergence of LLMs could be transformative in this context, as they offer a wealth of knowledge that can be harnessed to enhance e-commerce experience.

Presentation

Current e-commerce platforms typically display products in one or more lists or carousels, which often fail to account for the context and nuances of varying user intents. This uniform approach is not only unengaging but also less effective in addressing customer needs.

Different context calls for different forms of presentation. For specific queries like, “iPhone 13 Pro Max” or “Lucerne Milk, Reduced Fat”, an ideal presentation would feature the requested product alongside related complementary items. For broader searches such as “mother’s day gift”, providing a comprehensive overview of related categories or obtaining more detailed user preferences is essential. In the realm of fashion, more sophisticated interfaces may be necessary to allow users to specify their desires more precisely, or to offer virtual try-on experiences.

But when we consider the broader picture and compare e-commerce with traditional in-store shopping, a significant gap becomes evident in how products are presented to the customer, leading to stark differences in customer experience.

E-commerce offers convenience but misses the immersive sensory experience of in-store shopping. Customers in physical stores relish the excitement of trying on new outfits, the thrill of discovering innovative gadgets, and the delight of serendipitous finds. Replicating these tactile and inspiring aspects in e-commerce is a challenge. While advancements in technologies like generative AI, multimodal models, and AR/VR are making strides towards mimicking the real-world shopping experience, there is still a considerable journey ahead in fully capturing the dynamic and sensory-rich environment of traditional shopping.

‍