The impact of LLMs on marketplaces

Published on
September 26, 2023
Bart Dessaint
No items found.
The impact of LLMs on marketplaces

Historically, enduring marketplaces have developed in parallel with new underlying infrastructure. As highlighted in our TechCrunch piece five years ago, the rise of Craigslist was made possible by the concurrent popularisation of email – and likewise eBay by the graphical browser; Uber by location services on smartphones; and Airbnb by the reputation systems of social networks.

LLMs and generative AI stand to be the next platform shift, enabling us to both interpret data and generate new content with unprecedented ease. We’ve recently shared our investing framework for generative AI applications, and now turn our attention to the next decade’s marketplaces. LLMs will enhance current offerings, allow for completely new exchange, and alter the economics of this business model.

Improving the Status Quo

The basic purpose of a traditional marketplace is to match supply and demand, and facilitate exchange. It accomplishes this by sourcing and onboarding suppliers, enabling product discovery, facilitating the transaction itself, and assisting with post-purchase support.

Supply Side

Supply sourcing: overcoming the cold start problem

Marketplaces add value by aggregating a fragmented supply base, and founders often tell us that sourcing supply can become an extremely time-consuming task, especially in the early days. This is the much-discussed “cold start problem”: it's hard to provide value to the demand side until you have supply, and vice versa. There are two distinct challenges:

  1. Sourcing: finding the largest and most relevant sources of supply. In the early stages, this is often done manually, which is time-consuming and expensive. At scale, this is a marketing exercise where the goal is to attract suppliers at the lowest possible cost.
  2. Onboarding: raising awareness, presenting the value proposition, and convincing suppliers to join the platform. In some verticals, this involves numerous touchpoints via phone calls and emails, leading to high supply-side CAC and suboptimal top-of-funnel conversion rate.

How might LLMs eliminate this friction?

  1. At the sourcing stage: Marketing use cases for LLMs are well-documented, making it easier to write personalised copy and content. If suppliers need to be discovered proactively, LLMs can be used to make web scraping and crawling simpler.
  2. At the onboarding stage: LLMs could expedite the process of onboarding a supplier, by offering a lifelike chat interface that assists sellers, gathers information, and provide clarifications on how to get started. To facilitate cross-border trade, they can offer on-demand translation services to assist international sellers – with LLMs now outperforming specialised translation tools. To ensure continuous improvement, they can also analyze data related to onboarding and identify bottlenecks.

These new LLM-powered tools could reduce CAC during sourcing, and increase conversion during onboarding. The risk is that this may also lead to less sticky supply, as it might become easier to participate in multiple platforms over time.

Supply creation: augmenting creators, or replacing them?

Many marketplace suppliers are limited only by their own creativity, especially on maker- and creator-focused platforms like Etsy. LLMs might help these suppliers ideate and generate new product concepts, or aggregate and reimagine existing content.

LLMs are creating a new tension for certain marketplaces – like those for freelance work or digital goods (such as artwork and books). As these marketplaces have amassed extensive and detailed data on the goods they have sold over the years, LLMs might be used to directly generate some of these digital goods, personalising based on customers' purchase history and browsing behaviour.

This will raise some questions of IP and copyright, and will have knock-on effects on the creator economy: how does the role of the human supplier change?

Etsy's success was partly driven by a simple insight, that consumers love purchasing items crafted by 'real individuals' rather than faceless businesses. In this regard, we expect that LLMs will assist creators more than replace them outright. On the other hand, some services marketplaces might not even need a human supplier at all – indeed, the threat of generative AI was reflected in Fiverr’s declining stock price.

Marketing / listing: seamless content generation

Listing is an obvious application for LLMs. The user’s goal is to list their products in the most engaging possible way, and generative AI products can assist with this:

  1. Text - LLM-powered tools can be used to generate rich listings, with minimal input needed from the user. Several e-commerce platforms have already embraced this: Shopify’s Magic feature helps to generate engaging product descriptions for brands. Over time, we think this will be table stakes to reduce the friction of listing products.
  2. Image - Similarly, with recent advancements in computer vision combined with generative models, users can generate images of their goods with any background they desire, and adjust style and lighting to their preferences. This can be done at large enough scale to tailor images to individual customers’ preferences. On the other hand, marketplaces will need the guardrails to combat entirely AI-generated imagery, as it’s easier than ever to offer deceptive representations of the goods sold.
  3. New formats? Traditionally, marketplace listings take the form of text and imagery. With generative AI, the ability to generate new formats, such as audio, video, and 3D, opens up exciting opportunities for suppliers and marketplaces alike. This technology allows them to showcase their goods in a much more interactive and engaging way.

These changes are particularly promising for consumer verticals that have a large volume of SKUs and frequent stock renewal – where the friction of listing is highest.

Demand Side

Search and recommendation: the end of scroll fatigue?

Navigating a traditional marketplace often involves tedious scrolling through many product pages. LLMs can augment recommendation engines, and enable tailored product discovery with a human, conversational touch.

Traditionally, boutique, luxury retailers sell in a much more personalised way than high-throughput online marketplaces. The former rely on clienteling — a personalised approach to cultivating long-term customer relationships — to boost conversion. This typically requires a large basket size to justify the associated human effort, so it has been limited to premium retailers.

LLMs can change this. The expensive and hard-to-find "agent-vendors" that marketplaces seldom or never utilised can now be readily available on their platforms. This transformation holds particular promise for high-volume verticals, like apparel, CPG, and even dating, where there is enough data to observe user preferences accurately, and the quality of product recommendations has an immediate impact on conversion.

Customer support: the rise of the virtual assistant

In the past, successful marketplaces have refrained from offering extensive customer support services to protect their unit economics, especially in high volume/low commission models. LLMs can allow for better customer service at lower cost – voice and virtual assistants, can now provide quick and personalised real-time assistance for returns, disputes, and warranties. Marketplaces now have the opportunity to transform the post-purchase customer experience, boost NPS and increase retention.

Autonomous pricing

When it comes to pricing, AI will have the greatest impact on complex purchases with a wide range of parameters and risks. Consider the purchase of a car: anyone who has ever bought or considered buying a vehicle online is aware of the complexity involved. The price one is willing to pay depends on multiple factors: both item-specific (e.g., colour, year, condition, options, mileage) and individual-specific (e.g., use, budget, financing, urgency of need).

Over time, one could imagine that buyers may be able to specify their preferences in natural language with an agent that infers the parameters and their weights. This bot would then run the negotiation with the supply side (or their own bots, which would rely on their own parameters such as available supply, minimum margin, and time-to-end-of-season) and bid on their behalf.

This could lead to more fine-grained dynamic pricing or staged service delivery in high-value transactions, optimising for the preferences of both sides of the marketplace.

New Marketplace Models and Tools

Now, let’s delve into the types of platforms that were previously hard to build or operate. We split these into two categories: new verticals of goods and services, and new types of interaction.

New verticals

LLMs can help us more efficiently aggregate supply, enhance listings and facilitate complex negotiations. This prompts us to ask: until now, which sectors of the economy have been either too expensive to transition online, hard to market, or complex to price?

Given low historic penetration rates, we focus on B2B segments, particularly those that involve high basket value and heterogeneous SKUs. For example, we can envision marketplaces for complex manufactured goods or multi-step human services, where traditionally they are bought offline. On the B2C side, we can project potential opportunities that share similar characteristics – such as more efficient travel services, or rare / unique items that have been difficult to price.

New interactions

Eventually, LLMs could serve as a new “front end to the internet” – a new way to discover anything one might need (as Google has been for the last two decades). This leads us to contemplate whether a comparable transformation is on the horizon for online marketplaces, and whether a new abstraction layer will emerge – a "meta marketplace". This could look like a personalised agent which understands the user intimately, searches all corners of the internet to enable seamless purchases across many destinations.

If this layer emerges, the role of marketplaces will change substantially. Will they adopt defensive strategies to avoid involvement with these new aggregators, aiming to sustain a more direct rapport with their customers?

Economic Considerations

We’re not yet sure what the long-term impact of LLMs will be on the business model of marketplaces. Our hypotheses are:

  • Larger aggregate GMV: if LLMs reduce the friction of listing and discovering products, they are likely to drive up the total volume exchanged.
  • Downward pressure on commission in the short term, owing to an easier onboarding and listing process, which could result in a lower barrier to entry, and lower switching costs on the supply side.
  • Ability to offer premium services on the demand side, offering access to advanced search features and premium pre-purchase customer support.
  • Ability to offer premium services on the supply side, with an opportunity for marketplaces to start charging suppliers a SaaS fee to access LLM-powered advanced capabilities, such as listing and pricing.


As investors, we’re excited by three categories of startups:

  1. “AI-native marketplaces” that serve existing verticals, and unlock new efficiency using LLMs ;
  2. Entirely new platforms made possible by LLMs – either new interactions or verticals;
  3. Tools and enablers for the above

Whether you are a marketplace founder leveraging LLMs – or have an altogether contrarian perspective – we would love to hear from you!