7 marketplace challenges AI can help us solve

Online marketplaces are a US$3.3 trillion industry and an attractive space for businesses to operate in. Marketplaces are category builders, have incredible impact on seller and buyer communities, and even shape consumer behaviours and economic trends. However, there are also unique challenges in building or scaling a marketplace business. Often you have a chicken and egg problem, you face constant trade-offs on geography, category or customer segment level. If you’re a horizontal you start losing to verticals; you open the supply too much and your quality goes down or findability for buyers becomes a problem. If you optimise too much for seller success, your relevance and buyer trust can deteriorate.?

I’ve been working in product and building e-commerce marketplaces for the last 15 years, and I’ve often been asked about AI and what impact it will have on the marketplace industry. I’m extremely bullish on the possibilities AI can unlock not only to sustainably grow a marketplace business but to introduce a step change in terms of user experience. Using AI is nothing new for marketplaces - whether for discovering an item to buy, listing a product to sell, pricing, or fraud detection, AI is behind the scenes enhancing the product and user experience. However, in the last few years we’ve reached a point where many underlying AI technologies have become more effective and accessible. In addition, talent has become more widely available and many startups are in a better position in terms of tech-stack and data foundations to actually take advantage of AI applications.?

So what are some of the major marketplace challenges AI can help us solve?

#1 Network Effects?

It all starts with network effects. Understanding how your network effects work and how you can intentionally trigger them can create defensibility and accelerate growth. The lifeblood of network effects is marketplace liquidity - the efficiency with which a marketplace matches buyers and sellers. Network effects are not uniform. Depending on the particular type of network effect there are different factors that drive marketplace liquidity. The most common network effects for e-commerce marketplaces? are: hyperlocal, cross-border, category, and cross-category.?

Hyperlocal network effects are present in ride-hailing or food delivery marketplaces (like Uber or DoorDash) where most transactions occur in hyperlocal clusters and consequently there is a strong correlation between density and liquidity. The higher the number of participants within a relevant locale, the higher the liquidity of the marketplace. With cross-border network effects in global marketplaces (such as Airbnb or Catawiki), every addition of supply makes the platform more valuable globally, and the higher the global supply and less friction during the cross-border transaction process the higher the liquidity becomes. Category network effects occur when users benefit from depth of supply within the specific category - most often in vertical e-commerce or real estate portals (such as Rightmove or CarGurus) where for buyers it is very important to know they have access to all available market inventory. For category network effects, the bigger the distance with any competitor in terms of inventory size, the higher the liquidity and the stronger the network effects. In cross-category network effects there is an increase of value when participants make transactions in more than one category - the higher the cross-category or cross-segment (buyer to seller) behaviour the stronger the correlation with liquidity. This is most common in C2C platforms. Finally, it’s important to remember that a business can have several network effects at play, so they’re not mutually exclusive.?

So how can AI help in managing liquidity and network effects??

• The most traditional use of AI for liquidity management is powering the buyer discovery journey - search & relevance, ranking & best match, recommendations and personalization - all areas where AI systems are fairly mature.?

• AI can be an effective way to intentionally distribute the visibility of the marketplace supply - deciding which sellers and objects to prioritise or take visibility from or even not allow on the platform in the first place. In this context I hate the phrase ‘marketplace equilibrium’ because it’s not an equilibrium - if you’re balancing too much the needs of everyone you’re doing a lousy job in growing the segments and liquidity pockets you should care about most. Also important to remember is that there are never only winners - if you give value to a particular segment you’re inevitably taking value from others so try to be conscious about those consequences. To do this effectively with AI and create repeatable network effects it is important to first translate the underlying insights of the network effect dynamics into optimization goals for the systems that distribute and expose content and make decisions on ranking - recommendation engines and personalization features, search, CRM etc. AI can be extremely effective to narrow in on a specific set of customers and liquidity pockets and optimise visibility for priority segments - so the sharper the segments and goals, the more effective AI can be in achieving them.?

• Multi-objective recommendation (MOR) systems is another area of interest. In terms of liquidity management and network effects, MORs allow optimization for the interest of multiple sides of the marketplace - buyers, sellers and the platform itself. This is an extremely handy tactic as it allows us to go beyond the typical content relevance goal and can add weight on additional goals (even conflicting ones). Examples include: optimising buyer relevance, seller success and platform profitability at the same time; managing demand concentration or fragmentation;? or managing transactional monetization within platform advertising where often we see trade-offs between buyer experience and findability and partner/advertiser success (for marketplaces that offer ad monetization products).?Further research on MOR.

• Similarity models and substitute recommendations are important foundational models for managing excess or gaps in supply. Similarity models determine how close/similar objects are, and what attributes they have in common. Substitute recommendations provide alternatives to buyers when a product is out of stock or supply levels are low. Unlike simple relationships used in similarity models, substitute recommendations require embedding of contextual knowledge and modelling different properties of substitution, which make them more complex to do well. On the opposite side, if we have a lot of the same or similar supply, we need a way to avoid overwhelming buyers with endless versions of identical items. Here AI models can help identify representative selection (without showing everything we have), inject a level of diversity ensuring we show relevant yet distinct items, and rank and re-order objects in a way that very similar items are not displayed closely together. This is an important factor in scarcity management and overall experience quality perception.?

************

#2 Verticalization

Is a vertical or a horizontal marketplace better at satisfying user needs and which one has a deeper moat? Horizontal marketplaces usually operate within several categories - Amazon, Ebay, and Etsy to a certain degree are such examples. They can become household brands with immense traffic as they address many purchase occasions, often unlocking multiple network effects. Vertical marketplaces, on the other hand, focus on a particular category - for example, Reverb for musical instruments, StockX for sneakers and streetwear - and they become really good at solving category specific pain points and with that offer superior user experience.?

Horizontal marketplaces as category builders can expand a niche market thanks to access to mass audiences and habit-building impact over their users, but over time they become victims of their own success and start facing impossible trade-offs, leaving many categories vulnerable for vertical newcomers. The principal challenges for a horizontal platform in these circumstances are how to retain that specialist perception and category trust, and provide the best findability experience that can lead to superior liquidity to any vertical alternative. This is where AI can help give horizontals a better fighting chance (there are still challenges with brand perception and product portfolio trade-offs, but better user experience can be achieved).?

There are two areas, which have an outsized impact in terms of category trust and findability - personalization and product taxonomy. Let’s look at each one and see how AI can enable better user experience and verticalization.?

Personalization for Verticalization?

Personalization introduces the ability for any platform or marketplace to adapt to a better version of itself, fit for each individual user and purchase/use occasion. In the context of verticalization we need to think beyond traditional content personalization and think about personalization across three levels - algorithm, context and display. For a multi-category marketplace this could mean knowing which discovery angles are relevant in each category and displaying them only and where they are needed. For example, for art buyers and art purchase occasions - artists, artwork type, art movements and style, visual search, can be extremely important. Having such contextual knowledge shows buyers that the platform understands the domain and their needs and knows how to guide discoverability and decision-making process - this is key in building category relevance and trust. This part is usually where vertical marketplaces win over the horizontals, as horizontals need to optimise and spend resources on product investments, which make sense for most categories and end up offering generic navigation and little category depth. Other important parts of the category context include understanding the typical user journey and needs for the specific category or purchase occasion: decision attributes and their weights, user jobs and tasks at hand,? category and cross-category network effects, the importance of displaying the obvious content vs. adding serendipity etc. All those factors are important considerations which should shape category experience and how we assemble pages and sessions for users.?

In order to be successful at verticalizing a horizontal through personalization we need to think of the following three layers:?

• Algorithm level –> relevance and weight of category attributes, importance of diversity, popularity, freshness, substitute factors for our ML recommendation models or ranking capabilities.

• Context level –> category context awareness, balancing the needs of early exploration or fast findability depending on user state, importance of scarcity or abundance, ability to model user states across multiple sessions or purchase occasions.?

• Display level –> the ability to construct and assemble pages and navigation dynamically, content first adaptive UIs to give immersive category experience.?

In conclusion, moving beyond algorithm or ranking level personalization can help us build truly immersive category experiences but will require a broader set of capabilities and need to incorporate overall user experience and platform interactions.?

Product Taxonomy?

The second capability, which has an outsized impact for building category trust and superior findability is the product taxonomy within the specific category context. Product taxonomy is the logical and structural organisation of products and consists of product categories and attributes, and how they relate to each other. The purpose of such taxonomy is to enable better discovery and category navigation and filtering, and feed structural data to various ML models. Without a solid taxonomy a marketplace cannot have effective findability, robust similarity models, recommendations or pricing capabilities. Poor taxonomy can also impact category trust as it won’t convey the right context to buyers.?

Beyond taxonomy data, tone of voice and product and attribute naming can be extremely important in building empathy with the specific buyer or seller segment. Are you talking to experts and trying to convey specialist knowledge or do you want to speak to everyone with accessible language? These are the types of considerations for the taxonomy components that gets exposed to end-users.

AI has two roles when it comes to product taxonomy:

• AI as a consumer of the product taxonomy. All models need good structural data. They can only be as good as their underlying data and taxonomy. Let’s say you’re building a price prediction model for watches, you need to know which product attributes influence price variation and how different product versions vary. Without clean and well-structured data about watch brands, models, material, movement types, case size you can’t achieve high model accuracy and have useful predictions. In this case you need to think about designing your product taxonomy that can be effective for algorithms to use.?
• AI in service of the product taxonomy. There are also ways AI can enhance how taxonomy data is being collected and how it is being used. Creating relevant tags and product data points that allow us to automatically classify and categorise our products is a good task for ML models. Autofilling seller listings from text or image input; building and maintaining search vocabularies with LLMs, generating discovery angles, informing related and substitute product versions, managing SKUs at scale - all types of tasks AI can make marketplaces more effective.?

In conclusion, AI can make a step change in the ability to verticalize a horizontal platform. However, deep contextual knowledge of the category dynamics, personalization of the overall experience and immersive UIs, and high quality product taxonomy are the prerequisites.?

************

#3 Product recommendations for long term buyer satisfaction?

Catawiki is a marketplace for special objects and we cater to collectors as well as a growing number of passionate enthusiasts - people with a passion or spontaneous interest for anything art, interior, luxury or collectables. As our inventory grows we see an increasing need to improve our personalization and product recommendations. So often we ask ourselves what makes a good product recommendation, especially in the space of wants and not needs??

We see that two factors are important in product recommendations - relevance and diversity/novelty. Most people have multiple passions, they can entertain several at a time or switch from one obsession to another, so balancing the familiar with the unexpected is very important. Next to that, people can be in different states - one where they want to go into a rabbit hole and stay there for a while or be ready for a new adventure.? Having the contextual knowledge and the ability to determine the state someone is in is very important yet quite tricky, especially when you have both collectors and casual shoppers.?

So, can we build pathways that can not only satisfy a one off desire but also take people on a journey of indulging a lifetime of passions. And how to build discovery journeys that are optimised for long term satisfaction and fulfilment? This can be difficult for traditional recommendation systems, which are often optimised on short term rewards - if you click on a few Rolex watches you’ll inevitably see a lot of Rolex watches in your product recommendations - this approach might be effective for a while, but not for long. So we need to figure out how much of these highly relevant yet obvious results to show you, how much serendipity to throw in the mix, and what level of serendipity - other luxury watches, or classic cars and even Bordeaux Grand Cru wines. Shall we keep it just to products you’ve somehow shown interest in previously or bring totally new worlds to you. And most importantly how to manage explore-exploit trade-off, exploitation being selecting products based on the current knowledge and shown signals and exploration involving trying out new things. To find answers with traditional ML techniques and a/b tests can be very slow and inefficient as combinations are endless. This is where new techniques such as Reinforcement Learning (RL) can be an effective tool. RL can simulate many scenarios and combinations of sequential recommendations. It adapts to a changing environment, continuously learns from user interactions across multiple sessions and discovers the path that leads to the highest cumulative reward - not just one click on a Rolex watch but long term buyer satisfaction and retention. The most exciting work in this space has been done by Tony Jebara, first at Netflix and more recently at Spotify and I can’t wait to see more RL applications within the ecommerce and marketplace space.?

************

#4 Curation?

The single capability which I believe will truly reimagine e-commerce is curation. We live in an era of excess and curated marketplaces help buyers filter out the noise and find the best choice for them. Instead of offering the biggest inventory like amazon or ebay, marketplaces like Catawiki (special objects), Goat (sneakers) or The Fascination (trendiest brands) are selective by design. They focus on thoughtful selection and presentation of intentionally sourced products and sellers. Let’s look into how Curation adds value to the marketplace model and where AI can supercharge those capabilities:?

• Curation as a trust builder. Many curated marketplaces create greatest value in its verified sourcing and high-quality product selection. The curation process involves strategic decision-making on a type of content that aligns with the target audience and brand identity. Once there are clearly defined standards and criteria for selecting products AI tools help analyse large inventory and detect products that meet quality guidelines, making the process of sourcing more effective.?Authenticity is another valuable asset especially for rare, higher value items. Traditionally this has been resource heavy as it relies on human verification and physically reviewing an object. Innovative AI solutions are emerging, which aim to automate part of the process and improve scalability. AI-driven authentication systems have the ability to analyse the fine-details of a luxury watch or a designer bag and compare with a database of authentic pieces. These advanced systems assess different aspects of the watch or a bag, including its design, materials, labels, to ascertain its authenticity.?Computer vision, similarity and patterns detection and the use of microscopic photography are a powerful combination for such use-cases. Having proprietary models that continuously learn and adapt to new fraud patterns, and building large databases of authentic and fraudulent products is essential for a successful AI-driven authentication.?
• Curation as a discovery angle. Curation serves as a discovery tool, moving beyond conventional inventory categorization. Using curated collections introduces new discovery angles, appealing to varied mental models.?Let’s say a sneaker marketplace could think of complementary ways to organise its content: by use - running, workwear; by features - breathable, water friendly; or by themes (or hyper-reference to a culture) - minimalistic monochromatic, hispanic heritage, iconic kicks, japanese addictive labels - the possibilities are endless and way more enticing. To do that a marketplace needs deep category knowledge and authentic presence in certain subcultures - that’s why human experts, curators or editors are often at the heart of curation. AI complements human curation by making the selection & compilation process more effective. The approach can be algorithm first where algorithms create content pools on objective factors and human curation act as the ‘filter’ or human first where humans envision a need, create content hypothesis, compile content pools and algorithms augment with similar content; see an example from Spotify and their algotorial approach. Algorithms can also be used for trend awareness to continuously refresh and adapt the curated selection, ensuring they remain relevant and enticing.?

************

#5 Understanding purchase intent?

I would like to highlight one specific part of the findability experience where recent advancement of AI and LLMs are making significant progress and have made easier to build great experiences for buyers - query understanding and mapping purchase intent to the available supply.?

LLMs excel in natural language processing enabling buyers to use conversational queries and see precise search results that not only match keyword input but also respond to purchase intent. For example, a buyer can search for ‘housewarming gift’ and based on popularity trends a model can detect the most common products people buy for this purpose and recommend products across a variety of categories and within the typical budget for this purchase occasion. Another big unlock here is the ability to continue the conversation and advance further the discovery process - adding or changing the search criteria - ‘for small apartment’ or ‘for a culinary lover’.?

Query reformulation (QR) is another popular technique in product search, where AI can do most of the work and offer superior buyer experience. Reformulations aim to maximise the product coverage while maintaining relevance. Simple applications include synonyms & misspelling handling but other use cases can solve for context and diversity maximising not only the buyer side intent but supply visibility.?Amazon has been doing interesting work in the QR space.

From a user experience point of view I’m extremely excited about the possibilities here as for decades we’ve been talking to machines with a machine language and now we have a chance to build truly intuitive and human centred shopping experiences.?

************

#6 Driving seller success

Seller success is another obvious opportunity space where AI? (generative AI in particular) and LLM developments can make a difference: I will highlight three areas where AI enables marketplace sellers to be even more efficient and successful - demand insights, content creation, and buyer interactions.?

• Demand Insights. AI can analyse vast amounts of data to identify demand trends, allowing sellers to make informed decisions on type of inventory, popular products and marketplace liquidity needs. ML algorithms can forecast inventory needs based on historical sales, seasonality, and market trends, helping sellers to optimise stock levels and reduce excess inventory. Additionally, AI can provide insights into customer behaviour preferences, advise on competitive pricing, inform highest margin products and ROI on marketing investment, and enable sellers to adapt strategies and position effectively on the marketplace.?

• Content Creation. AI can enhance product images and videos, through image recognition allowing for automatic tagging for better discoverability, background removal or lighting enhancement and even deeper editing - showing product in context or use. Natural language processing can be utilised to generate compelling product titles and descriptions by analysing successful past listings and adapting for the highest conversion probability. AI-tools can suggest improvements, such as keywords for better search or language enhancement to increase buyer engagement. Effective pricing is another friction for sellers, which typically takes a lot of time and effort to do well. Based on comparable inventory and historical transactions, AI can suggest effective product prices, discount strategies, and positively impact sell rates for sellers.?

• Buyer Interactions. AI can assist marketplace sellers in buyer interactions by generating answers for typical buyer questions - answer buyer queries, provide product or installation information, guide customers though purchase process - uplifting user experience for buyers and saving sellers time. Additionally, sentiment analysis tools can monitor customer feedback and reviews and detect meaningful product improvement patterns or insights.?

************

#7 Localization?

In terms of technology or user experience, executing a localization strategy is not particularly difficult but is often an overlooked and underinvested area and many marketplaces end up with poor experience, especially for non-core markets. Why I get excited about the impact of AI in localization is the opportunity to make localization accessible and easier for many companies to execute well and scale globally. When you localise a marketplace usually you need to acquire local supply, create local marketing assets, translate platform and content in the specific language, build local trust, expand local payment and fulfilment capabilities etc.?

The emergence of LLM and generative AI is a game changer for the ability of a marketplace to localise effectively and adapt the experience for a new geographical audience. From basics such as machine translation of content, navigation or customer support scripts, to optimising search vocabularies and product taxonomies across GEOs and introducing abilities to generate and adapt content and marketing assets to different cultures and buyer preferences - all these are areas which LLMs and generative AI can add value.?

Estimating dynamically order arrival times is another area algorithms can do very well and with that enhance buyer trust. This is particularly important for global marketplaces where you have supply from different GEOs and very different delivery times - so managing buyer expectations and providing transparency on arrival times helps.?

************

In conclusion, the challenges faced by online marketplaces are diverse and complex, requiring innovative solutions and AI emerges as a powerful tool to address these challenges. As technology continues to advance, integrating AI solutions becomes not only beneficial but imperative for those seeking to satisfy user needs and stay relevant in the ever-evolving landscape of online marketplaces.?

There are many exciting applications of AI for a marketplace business: triggering network effects, shaping inspirational & fulfilling buyer journeys, building category relevance, making sellers more effective and enhancing cross-border shopping.?

Finally, here are a few thoughts on what is important for successfully deploying AI capabilities within an organisation:?

• Clear strategic intent and deep and widely available business and user knowledge -? the single most important foundation to build on. AI can only be as good as the insights and goals that shape it.?
• Not just data, taxonomy is the new gold for a marketplace (product taxonomy, ML models taxonomy, customer support and user feedback taxonomy, user segments taxonomy etc).?
• Build vs. Buy vs. Benefit - we don’t have only build or buy options, one can also benefit passively as AI tools and capabilities are being adopted independently by sellers or marketplace partners - educating them can accelerate such adoption without paying the price of building or implementing in-platform.?
• Finally, AI doesn’t come at no cost - both for the business and for our planet - so being thoughtful if and where to use, as well as understanding the actual costs and effects on the wider ecosystem are important and require leaders and founders to think beyond immediate economic benefits.?