Messy, noisy, and colorful, with an endless array of products touted by an army of vendors, online marketplaces are not unlike real-life bazaars. Shoppers need to be on their toes to navigate through the information overload, but in return, they can get everything they want in one place. And while IRL they must fend for themselves, online, they can rely on recommendation systems – that is, personalization for marketplaces.
In a digital context, all that color and noise translates into, you guessed it, data. Tons of data, rich and unstructured, creating an environment that’s equal parts challenging and rewarding for recommendation systems. What are these hurdles and how can they be cleared by personalization providers to bring about a win-win-win for marketplace platform, vendors, and consumers alike? Let’s see.
Why Marketplaces Exist in the First Place
In order to understand the importance of personalization for marketplaces, it helps to get the basics straight. An online marketplace facilitates the sale of products in a business-to-consumer (B2C) setup by aggregating the offerings of third-party vendors on one platform. (Bundled together with this definition, we’ll also refer to auction sites like eBay and classified sites like Craigslist, although these operate, at least in part, on a consumer-to-consumer (C2C) basis.)
As opposed to a manufacturer selling its products online, or to an e-retailer running a web store, a marketplace is much more of a tech entity whose real asset is not a product or a service but the platform itself. It is characterized by
- Multiple vendors, overlapping offers. On a marketplace, most products are being sold by more than one vendor. So even a very precise search, whittled down to a single item, can yield several different results.
- A generalist approach. With the exception of vertical classified sites that zone in on one specific category like used cars, online marketplaces sell practically everything under the sun.
- An immense range of available products. Hundreds or thousands of vendors are present on an average marketplace, each with a separate web store’s worth of products. No surprise, then, that the overall product catalog constitutes a mind-boggling figure. For example, N11, Turkey’s leading marketplace site and a Yusp client, carries more than 20 million different products.
Provided that everything is running smoothly, an online marketplace makes a lot of sense for all parties involved. The marketplace operator can skim off some of the vendors’ profit in exchange for letting them use its platform. It saves vendors the trouble of maintaining their own web store and they, in turn, take care of the logistics of retail on an overall staggering scale.
By aggregating the offers of multiple vendors, the marketplace also saves a lot of legwork for shoppers who can compare and pick the best deals in one spot. To some extent, it vouches for the trustworthiness of its sellers, and makes efforts to standardize their product-related content for easier browsing. The marketplace operator also takes control of many other aspects of the customer experience, like providing a seamless search function, or ensuring that shipping costs don’t remain hidden till the very end of the checkout process.
The Pivotal Role of Personalization for Marketplaces
The main reason why personalization plays a crucial role in the life of marketplaces is simple: size matters.
As mentioned earlier, the product catalog of online marketplaces is larger by orders of magnitude than that of a standard e-retailer. And the bigger the product catalog, the more sense it makes to use personalization. Lots of products and lots of users generate lots of data, and a data-rich environment creates ideal conditions for running a recommendation system.
For classified service providers, personalization is extremely useful in maximizing the value of ad revenues by boosting page views, click-through rate (CTR) and time spent on the site. By helping to connect the right buyers with the sellers, it increases the satisfaction of these latter, encouraging them to advertise more.
In case of other types of marketplaces, personalization earns its keep by increasing the number of sales, and generating higher CTR and visit frequency on the site, thus helping to maximize the value of commission fees collected from vendors. By serving relevant suggestions and additional cross- or upsell recommendations, a personalization engine can be a major force in converting more visitors into customers, or one-time buyers into serial shoppers.
Although some marketplaces also treat their vendors to tailored services, by default, personalization is aimed at end users, and managed centrally. That is, the deployment and maintenance of a recommendation system is handled by the marketplace operator. Together with the personalization provider, they decide on business rules, on recommendation placements and logics. While vendors have minimal control over this (in some cases, they may be able to opt in or out of certain customer experience features, including personalization), they too benefit from a soundly functioning recommender system.
The Challenges and Opportunities of Marketplace Personalization
In many aspects, the personalization of marketplaces is very similar to that of e-retail. For instance, KPIs in the two fields overlap almost entirely. The efficiency of a marketplace personalization engine is usually assessed in terms of gross merchandising volume (GMV) – this is the equivalent of revenue generated by recommendations, a metric typically used in the context of e-retail personalization.
However, the challenges facing personalization providers for online marketplaces are quite unique. The complexities and gigantic proportions of marketplaces present extraordinary difficulties for recommender systems, but also opportunities for growth that are not to be found elsewhere.
Multiple Vendors, Similar Products
The fact that many similar or identical products are offered by various sellers poses a double challenge for marketplace personalization.
First of all, it’s difficult to establish whether near-identical products are just similar, or exactly the same. And if they are the same, whether one item’s metadata (its description, interaction history, and ratings) can be applied to its clones. This identification process is especially important in tackling the cold start problem.
Best practices vary case by case. For N11, the Turkish marketplace, Yusp developed a three-tier product identification system: it assigns a GroupID to batches of almost identical products; a GTIN (Global Trade Item Number) to all completely identical products; and a unique ProductID to each available item. For example, all iPhone X 256 GB models for sale on the marketplace get the same GroupID; within this group, all phones in rose gold get the same GTIN; and each vendor’s iPhone X 256 GB in rose gold is tagged with a different ProductID.
Next up, the personalization engine has to select which of the competing, similar products to feature in its recommendations to a given user. Understandably, all vendors want a place in the limelight with their products, but if there are more available alternatives for a given item than a shopper can reasonably process, then the recommendation system has to compile some kind of shortlist.
It will highlight vendors that users have favored in the past, by comparing their products based on buyer events and choosing the most sought-after ones. That is, unless the marketplace imposes business rules to promote certain vendors or items. It’s also worth noting that recommendations aren’t entirely popularity-based: the shortlist of items selected for a certain user may also be influenced by their browsing history or known preferences.
With N11, Yusp has created a function enabling users to rate their shopping experience with a given seller, from choosing a product all the way to delivery. This explicit user feedback is then taken into account when ranking alternatives for a product recommendation. The better overall user experience a vendor provides, the higher they’ll be placed on the list.
Unstructured or Semi-Structured Data
Diverse product descriptions and inconsistent categorization complicate things further for the personalization of marketplaces.
Because the details of the items on sale are provided by individual vendors, or, on classified sites, by actual individuals, these descriptions vary greatly in terms of standards and granularity. All this data is not stacked and squared away, which would make it easier for a recommender system to sift through.
Assigning the items to product categories is also left to the discretion of the sellers, and here again, interpretations can diverge. What’s a jogging stroller, for instance? Is it baby gear? Is it a vehicle? Or sports equipment? Each vendor makes an educated guess or tosses a coin and shelves items accordingly. Since their efforts are not synced, the resulting arrangement may look tidy – everything is categorized somewhere – but in fact, it’s a labyrinth the personalization engine has to navigate through.
To be able to serve relevant recommendations under these challenging circumstances, personalization providers sometimes resort to micro-clustering. This type of algorithm collates the metadata (descriptions, ratings, user actions) of near-identical products into a virtual item for easier analysis and selection.
Different User Intents and Purposes
All kinds of shoppers flock to online marketplaces. Some are regulars; others make the odd purchase once in a while. Some shop for a single household; others buy in bulk for their business. Some people top up their pantry supplies with a steady stream of FMCGs, while others visit marketplaces to procure one big-ticket item. Their priorities, their budgets, their decision-making criteria are all different. Personalization engines need to pick up the pattern and adjust their recommendation strategy, or else they’ll fail spectacularly (cue the Amazon toilet seat story that made waves a few years ago).
How can we avoid the blunder of recommending an item the user has no intention of buying again? Instead, how can we serve recommendations that will keep them coming back? User profiling is the answer.
For a giant eBay affiliate marketplace in Turkey, Yusp identified several main user groups based on add to cart events. “Small” users with single, one-off purchases belonged to one group; “light” users with small, infrequent purchases and “midsize” users were assigned to distinct groups; lastly, there were “large” customers with sporadic, bulk orders.
Having observed the buying behavior patterns in each group, we tailored our personalization approach accordingly, determining whether or not to recommend similar items, or products from the same category as the one recently purchased. This way, we could make sure that the average shopper who bought a television would not be bombarded with more TV offers; but that the gadget dealer who regularly buys iPhones to resell in his store would keep seeing iPhone recommendations.
Of course, there is more to marketplaces than iPhones and smart TVs. Of the millions of products available, the overwhelming majority are long-tail items – not in high demand, not bought in large quantities, but, thanks to their sheer aggregated volume, still responsible for a large chunk of the marketplace’s revenue. On classified sites, even popular items are technically long-tail, as only one or a few of them are offered by any given seller.
In case of long-tail products, the role of marketplace personalization is even more crucial in helping to connect supply with demand. Relevant recommendations are much more likely to attract buyers interested in a niche product than casual browsing or even search would.
In addition, because there are so many product categories on a marketplace, there are more opportunities for upsell and cross-sell recommendations. And the heavy traffic of such sites guarantees that even a tiny uptick in conversions brought about by these recommendations translates into significant profit.
Personalization for Marketplaces Means High Involvement
For an online marketplace, personalization is a core business function. Real-time, relevant recommendations are an indispensable ingredient of its success.
As a consequence, marketplace operators are deeply invested in personalization. They are demanding clients with high standards and ever-changing requirements. They are always on the lookout for new recommendation features, often tinkering with their own platform, which in turn necessitates further adjustments by their personalization provider.
Multiple stakeholders along the marketplace’s supply chain – vendors and category managers, often representing clashing interests – render business rules extremely complex. With a product catalog that’s constantly in flux, seamless and instant synchronizing is a must. Integrating the vast quantities of data a marketplace churns out with the personalization system and ongoing maintenance requires close, daily collaboration between client and provider. Frequent A/B testing – of new features against existing ones, or of in-house solutions versus the personalization engine’s functions – is a key to progress. Analytics-as-a-service plays an increasingly important role: the personalization provider helps its client interpret performance stats as actionable results.
Here at Gravity R&D, we have considerable experience in personalization for marketplaces. Yusp, our recommendation engine, is used by online marketplaces and classified sites around the world. Read more about what we do, or get in touch with us – we’ll be happy to hear from you.