Algolia acquires »

Read More

Best Practices to Improve Marketplace Product Search & Discovery

Best Practices to Improve Marketplace Product Search & Discovery

Forty three percent of visitors to e-commerce sites begin their journey at the search bar, and when it comes to search, few online businesses have a greater challenge than marketplaces. The scale of marketplace operations is staggering: millions of products divided into hundreds of different categories; terabytes of data related to product descriptions and images; hundreds, if not thousands, of separate vendors over which the marketplace has limited control.  

How, given all this complexity, can a marketplace manage to get customers to the products they want to buy? And what if those customers only have a vague idea of what they’re looking for and need to browse? 

Just performing basic functions – getting customers to a specific product like a shower curtain rod or a product category like men’s trousers – is a task that can defeat many search engines on a regular basis. The idea of optimizing the results for every search doesn’t even come under serious consideration.

Because of the scale of their operations, most marketplaces let themselves be guided by the Pareto principle – 20% of your catalog accounts for 80% of your profits – and optimize only for that top tier of items: the best sellers, those with the highest margins or perhaps lowest return rates. 

In this article, we’ll talk about how modern, search-as-a-service solutions allow marketplaces to set the bar much higher. Specifically, the focus will be on how quality data, along with AI-based technology for improved relevance and smarter ranking, can improve the search experience for customers, boost conversions across 100% of the marketplace catalog, and level the playing field vs. giants like Amazon. 

Data quality

Data quality is one of the first things we discuss with new clients. The effectiveness of the AI (machine learning) algorithms that optimize search results – and conversion rates – is entirely dependent on good, clean data. As one HBR article put it, “Poor data quality is enemy number one to the widespread, profitable use of machine learning.” 

The difference between high quality search results and substandard results often comes down to the data. The better the quality of the data, the better the outcome, both for the customer and the bottom line. 

Marketplace product schema

The foundation for maximally useful data is your search index. Whether you create it using a web crawler or an API, it can and should be enriched in order to optimize search results.

We have identified twelve best practices for enriching your search index, and they make a significant difference. In addition to standard best practices, marketplaces have special challenges with data they must cope with. Although they may rigorously enforce a data structure (schema) on sellers — title, product descriptions, images, SKU, tags, etc. – we often find that customers have data quality issues — poor product titles, descriptions, or product categorization. 

Seller-generated content, which is very popular in marketplaces, presents its own challenges: spelling, slang, image quality and more. 

AI-Powered site search 

Once the product data is in good shape the real challenge begins: connecting queries with that data for maximum conversions. This task has become increasingly complicated. 

There was a time when consumers typically wrote one-word queries, and search engines were built word by word to deliver matching results. That’s changing with the advent of AI-powered search. Google and others are conditioning people to write longer, more elaborate queries. Keyword-based search engines of the past still work for exact match searches, e.g. a brand name search like “Nike running shoe”, but struggle with longer or more nebulous queries or questions like, “something to protect my feet while running.” 

Here’s an example of what can go wrong, even with a simple three-word query. Not long ago, I searched for a “shower curtain rod” on a popular home furnishing site, and the results I got were for shower curtains. Multiply this kind of failed search relevance by the literally millions of SKUs that marketplaces carry and you have a very costly problem.

A few months later, I went back to the site and found that the shower curtain rod problem had been fixed. What happened? If I were to guess, someone reviewed their search analytics, saw the mistake, and wrote a rule for their site search engine to recognize “shower curtain rod” as a distinct query from “shower curtain.”

Unfortunately, this approach doesn’t scale. Marketplaces are constantly changing with new products, titles, descriptions, and metadata. There simply aren’t enough hours in the day to write rules for every query that yields irrelevant results. Without a better way to solve the problem, marketplaces lose one of their great advantages — their ability to offer very large numbers of popular and hard-to-find longtail products. 

concept search
An example of concept search — customers can get great results even without using keywords.

Fortunately, with the arrival of sophisticated AI, there is a better way. Neuralsearch® provides improved relevance for longtail catalogs and semantic-based search out-of-the-box. It actually combines keyword search with concept search to deliver superior performance compared with companies that stack vector search technology on top of keyword search. 

AI-powered search nearly eliminates the need to write relevance rules. It understands that a query for a “shower curtain rod” is different from “shower curtains” without training. Moreover, it handles common synonyms automatically. For marketplaces, it’s a huge savings in time and effort. For customers, it’s a better experience. A customer can type “shower curtain rod,” “shower curtain bar,” or “the thing that holds up a shower curtain,” and expect to get great results. 

Ranking for conversion

Relevance isn’t the only issue marketplaces need to address in search results. Ranking is also very important. Better search relevance translates into higher conversion rates. Smarter ranking takes optimization to the next level. There are three important strategies that work both alone and in combination: personalization, dynamic boosting, and A/B testing.

Relevance and ranking adjustment
Relevance and ranking adjustment to improve results for visitors.


With, personalization can be integrated into the search or browsing process itself. Personalization gives you even more control over which search results shoppers see and in what order. What’s more, customers increasingly expect it. According to one study, 71 percent of consumers expect companies to deliver personalized interactions.

Automated rules-based ranking instructs the search engine to rank specific items higher over a given time period based on factors such as age, gender, location, purchase history and whatever other data may be available. Generally speaking, the more data you have about an individual, the better you can personalize the results of that person’s search. Even if you have no data on a particular visitor, you can still personalize results based on geo-location, click stream, device type, time of day, season, and even the weather. 

Dynamic boosting

Search results can automatically be improved even more over time with signal boosting. implements this feature using the branch of AI known as reinforcement learning, which uses feedback (clicks, purchases, signups, etc.) to “teach” the search engine which results are most likely to lead to a conversion, and then push them to the top. 

Instead of making large changes infrequently, this approach – what we call dynamic boosting – makes frequent incremental changes. Dynamic boosting has many upsides. It improves results continuously and surfaces other potential results faster. Poorly performing results also tend to fall away quickly.

dynamic boosting
Dynamic boosting to automatically push important results to the top.

With dynamic boosting, you can not only define which events matter, but also control and adjust the weight of the boost for each type of event.

You can boost individual products or categories based on attributes like brand, material, features, etc. For example, if everyone that searches for “shower curtain rod” wants one specific item, you can ensure that item is pushed to the top of every search. Or, maybe you want to promote certain kinds of results within the product category, e.g. using margin or materials. 

A/B testing

Pushing the item most likely to yield a conversion to the top of the results list should be an important goal for every marketplace professional. A recent study of over 80 million Google searches indicates that over 25% of visitors click on the first organic response. The best way to determine what that should be is A/B testing

Most people are familiar with the idea of split testing for buttons, headlines, or images, but A/B testing applies to search results as well. And it works. No less of an expert on the subject than Jeff Bezos has stated, “Our success at Amazon is a function of how many experiments we do per year, per month, per week, per day….”

For marketplaces, split testing search settings can massively improve customer experience and revenue. It’s largely how Catch, an Australian marketplace, boosted search revenues more than 40%. 

The bottom line

To begin where we started, 43 percent of visitors to an e-commerce site begin their journey at the search bar. Improving search results has a direct effect on the bottom line. Amazon has figured out a formula that’s catapulted them to success. They also have more than 1500 people working on their custom Lucene-powered search and C++ search libraries. Smaller marketplaces lack such resources, but they don’t need them. Search-as-a-service is all it takes.

Don't just take my word for it! Try free for 14-days or schedule a demo with our team to learn how some of the world's largest marketplaces are boosting search revenue with AI-powered search.

AI ecommerce conversion ebook

Similar articles

Best Practice

8 Actionable Tips for Improving Your Search UX and Usability

Best Practice

Why Good Site Search Is Important for Government Websites

Best Practice

How to Fix the Out of Stock Product Problem Using Onsite Search