Platforms are among the largest, most profitable, and defensible businesses. Amazon, Apple, Alphabet, Meta, Microsoft, and Nvidia all have platforms at their core (more on them later). And many other companies have built dominant competitive positions anchored by platforms within their respective markets: Airbnb, Alibaba, Booking, ByteDance, Coupang, DoorDash, eBay, Mastercard, Meituan, MercadoLibre, Nasdaq, Naver, PayPal, Pinduoduo, Roblox, Sea Limited, Tencent, Uber, Visa, etc.

Yet despite their importance, the term “platform” is used so loosely that it has lost much of its meaning. Commentators, executives, investors, and entrepreneurs routinely apply it without a clear definition. And nearly every other software company now claims the label, often to make its prospects appear more promising.

In this post, we offer a clear and precise definition of platform businesses to cut through the confusion and focus attention on their key features.

Definition and examples

Our definition is simple (see our 2015 academic paper for more details):

Platform businesses enable direct interactions between customers who have chosen to join the platform.

Despite its simplicity, however, the definition carries some important nuances, as we will see.

To unpack it, let’s start with a few obvious examples. eBay is a platform because it enables buyers and sellers who have deliberately joined it, to find each other and trade directly, on terms they largely control, with limited involvement by eBay.

The same is true of thousands of other online marketplaces: Airbnb, Alibaba, Craigslist, MercadoLibre, OpenSea, Pinduoduo, Swimply, Upwork, etc.

Google Search enables search users and advertisers to find each other and transact through clicks and conversions. Similarly, Google’s AdSense/Display Network connects publishers and advertisers: publishers control ad space, advertisers set bids and targeting, and Google makes the match and delivers the ads.

Payment platforms (e.g., Alipay, American Express, PayPal, Venmo, Visa) enable transactions between consumers and merchants who have signed up. They focus on the payment step, rather than discovery. By contrast, food delivery platforms like Deliveroo, DoorDash and Zomato combine discovery, payment, and delivery for consumers, restaurants, and drivers.

Platforms are not limited to facilitating commerce. Dating apps (Bumble, Hinge, Tinder), social networks (Instagram, Reddit, TikTok, X), and messaging apps (Discord, Telegram, WhatsApp) enable communications between users.

Videogame consoles like PlayStation and Xbox are platforms that enable users to play games produced by third-party developers for those consoles. This was true even before the consoles had their own game marketplaces (PlayStation Store and Xbox Marketplace), which are platform businesses in their own right.

Most platforms today are digital, but they need not be. Shopping malls (the Dubai Mall, Roppongi Hills) and trade fairs (the Consumer Electronics Show) also qualify: they connect consumers with various retailers or suppliers without producing or owning the relevant products.

Platforms can involve one type of customer (e.g., Telegram, WhatsApp, Zoom) or multiple types. eBay connects buyers and sellers; Android connects users, device makers and app developers; TikTok connects creators, viewers, advertisers, and merchants via TikTok Shops.

What makes platforms special

Platforms have two important properties which follow directly from our definition. First, they exhibit network effects: as more users join, the platform becomes more valuable to each user because there are more users to interact with. A social network is increasingly attractive to every user as more users join. A marketplace’s appeal to each buyer increases with the number of sellers it hosts—and vice versa.

Second, because value is created through direct interactions among users, platform businesses tend to have significantly lower cost structures than their non-platform counterparts. Unlike traditional media companies (Disney, Paramount) who spend large amounts of money on content creation, social media platforms (TikTok, YouTube) produce no content—their users do.

Together, these features create enormous leverage, driving the scale, profitability, and defensibility of successful platform businesses. While we have already enumerated many examples that illustrate this point, Craigslist arguably comes closest to the Platonic ideal: in 2019, it generated over $1 billion in revenue with about 50 employees, margins near 90%, and a website stuck in 1995.

What platforms are NOT

Just as important as what our definition includes is what it excludes.

First, platforms must enable direct interactions. This excludes retailers, which buy products and resell them in their own name.

“Direct interactions” means customers—not the intermediary—set key transaction terms and are responsible for carrying them out. When the intermediary controls pricing, merchandising, and inventory (as retailers do), it is not a platform. This is why platforms require far less capital and scale more easily than retailers, be they online or offline. Businesses like 7-Eleven, Carrefour, Chewy, Costco, Trader Joe’s and Zappos are not platforms: they take inventory risk and control pricing, display, and returns. By contrast, eBay and flea markets do not take inventory and allow sellers to set terms.

Similarly, Netflix is not a platform because it acquires or produces content, and controls consumer access through subscriptions. YouTube is a platform because it enables third-parties to create and manage their own content channels for viewers.

The same distinction applies to services. Upwork and Fiverr enable direct interactions between freelancers and clients. By contrast, outsourcing firms like Infosys employ workers, set prices, and control delivery, making them service providers, not platforms.

Second, platforms require all relevant customer groups to join. For this reason, B2B vendors that help businesses interact with their customers are not platforms unless the businesses’ customers also join the vendor.

Joining a platform requires an explicit decision: paying for access, buying a device, or incurring meaningful effort (e.g., creating a profile or driving to a mall).

For example, Adyen sells systems that allow its business customers (e.g., McDonald’s, L’Oréal, The North Face) to accept payments from their customers, but the latter do not connect with Adyen directly. McDonald’s customers do not join Adyen in any meaningful way. Similarly, Kea provides restaurants with an AI-powered communication system that allows them to handle phone orders. But the restaurants’ customers need not even know Kea exists. In both examples and many others, the end user is not a participant, so these are not platforms in our sense of the term.

Shades of gray

Above we have seen examples of pure-play platform businesses (Airbnb, Alibaba, Craigslist, Discord, eBay) and of businesses that are clearly not platforms (7-Eleven, Adyen, Chewy, Infosys, Kea, Zappos). But reality is oftentimes more complex: many businesses are neither 100% platforms, nor 0%, but instead fall somewhere in-between.

First, the distinction between platforms and retailers is a continuum, not a brightline. Amazon.com sits somewhere in the middle: roughly 60% of sales come from third-party sellers, while the rest come from its own retail operations. Spotify is also somewhere on the spectrum: it acts like Netflix (reseller) for licensed music, but more like YouTube (platform) for podcasts. Furthermore, it allows some direct interactions between users via playlist sharing. Ride-hailing services (Bolt, Grab, Uber) fall in-between as well: drivers can use their own cars, choose work schedules, and work for rival firms, but platforms set prices.

Second, some businesses evolve from products into platforms. Amazon Web Services (AWS) began as a B2B vendor of cloud services, but added a marketplace in 2012, connecting AWS customers with third-party developers—essentially, an AWS app store.

Many other software-as-a-service providers (Intuit, Salesforce, Twilio) followed similar paths, becoming partial platforms by creating marketplaces around their flagship products. The extent of the transformation depends on how much incremental value these marketplaces add beyond the core product.

Shopify went even further. It began as a B2B vendor of tools for building and running online stores, later it added an app store for complementary third-party apps, and in 2020 launched Shop.app, which creates a direct connection with its customers’ customers. The app enables consumers to track orders, discover and purchase products from Shopify merchants, and check out seamlessly across Shopify-powered stores. As a result, Shopify evolved from a one-sided product business into a three-sided platform connecting merchants, third-party developers, and end-users.

What about franchising businesses like Hilton or McDonald’s? The classic franchising arrangement—in which the franchisor provides brand licensing and operational systems to franchisees—is not a platform, but rather a B2B vendor. End consumers transact with franchisees, not directly with the franchisor, even though they recognize the franchisor’s brand.

However, when the franchisor layers on a system-wide loyalty program (as both Hilton and McDonald’s do) or a centralized booking or ordering app, it introduces a platform element to the business that can generate some network effects.

From platform businesses to platform companies

In general, we prefer talking about “platform businesses” rather than labelling whole companies as platforms. A single company can operate multiple platform businesses, as well as non-platform business lines. That said, it is still meaningful to describe a company as platform centric when its platform businesses account for a substantial share of its value. The difficulty is that this contribution is not always easy to measure.

Some cases are straightforward. The aptly named Meta Platforms is almost entirely built around platforms (Facebook, Instagram, WhatsApp). The same is true for Alphabet: most revenues and profits come from its advertising platforms—Google Search, Google AdSense/Google Display Network, and YouTube—and the Google Play marketplace linked to Android.

Microsoft and Amazon are also clearly platform-centric, though less purely so. Microsoft’s core platforms—Windows, Xbox, LinkedIn and GitHub—remain essential, and even its main products—Azure and Microsoft 365—have platform features (collaboration, APIs for third-party developers). Similarly, both Amazon.com and AWS are platform businesses to a significant extent, as explained above.

But what about companies like Apple and NVIDIA?

On the surface, they both appear primarily as product companies. Apple’s revenues are dominated by devices (iPhone, Mac, iPad, Watch), and NVIDIA’s revenues come entirely from selling high-performance graphics chips (GPUs). Yet both are best understood as platform centric. Apple’s iPhone would be far closer to a commodity smartphone had it not turned iOS into a platform for millions of third-party applications and opened the App Store marketplace. Likewise, NVIDIA’s dominance in GPUs today does not rely solely on hardware performance. Just as important are the network effects and switching costs created by CUDA, the software platform that lets developers build applications for NVIDIA chips—even in the absence of a centralized app store or direct monetization of these third-party applications. Without their respective platform components, it is difficult to see how Apple and NVIDIA would be two of the top ten most valuable companies in the world.

The key point is that platforms need not dominate revenue to be central to a company’s success. A firm may appear product-centric in its revenue mix yet rely on a platform layer as the foundation of its competitive advantage.

Concluding thoughts

Defining platforms clearly is not a matter of academic pedantry, but an essential step in better analyzing business strategies and future prospects. The goal is not to classify businesses and companies in neat buckets labeled “platforms” and “non-platforms”. That exercise is both uninteresting and futile, as we have seen that real-world nuances make it hard to make such clean distinctions.

Rather, the objective is to identify the fundamental drivers of platform businesses and use them to articulate strategic decisions. Should a firm move toward a platform model or remain a reseller? What are the trade-offs? How can a particular product be turned into a platform and what is the upside of doing so? How much can a specific company’s platform component contribute to its competitive edge?

We explore these questions and many others in our various Substack posts.

***

Let’s start with two obvious examples: TaskRabbit vs. Airbnb.

TaskRabbit’s network effects are local. Meaning that they only exist within limited geographical areas. A tasker in Boston only cares about TaskRabbit clients in the Boston area. And vice versa, only Boston area taskers are relevant to a Boston-based client. Contrast this with Airbnb. Its network effects are clearly global. A host in Paris, France may receive guests from any country around the world, so the value of Airbnb to such a host is clearly higher the more travelers Airbnb has in all countries, not just in France. Conversely, a traveler from Brazil derives higher value from Airbnb when the latter has more hosts in all the countries the traveler might want to visit.

Note that the notion of local vs. global network effects is not about whether the company is local or global. TaskRabbit is present in several countries in North America and Europe, yet its network effects are local. This might seem like an obvious point to make, but we have seen startups with local network effects that are eager to claim a national or international presence, implicitly (and incorrectly) suggesting their network effects are stronger because they are active in many cities.

What about ride-hailing apps like Uber? At first glance, their network effects appear mostly local: Uber drivers in San Francisco only care about Uber riders in San Francisco and vice versa. However, to the extent that Uber riders travel and use Uber in various cities, that creates some amount of country-level network effects. San Francisco’s Uber drivers may also care about Uber’s Boston riders if many of them go to San Francisco. Thus, Uber’s network effects extend across cities which receive significant numbers of visitors from each other. That being said, we still expect the majority of Uber riders in a given city to be locals. And Uber’s global network effects are likely considerably weaker because a much smaller fraction of Uber riders in a given country are visitors from a different country.

Why does the distinction between local vs. global network effects matter? Because all other things equal, global network effects are broader and last longer as the platform expands, and ultimately provide far more defensibility for platforms than local network effects. Think of it this way. Every time TaskRabbit launches in a new city (let alone a new country), its network effects essentially start from zero. A tasker in Austin, Texas couldn’t care less about TaskRabbit’s installed base of clients in other cities. Nor does a client in Vancouver care about the many taskers that TaskRabbit has in US cities. Contrast this with Airbnb. Every time Airbnb launches in a new city or country, its entire installed base of hosts and travelers from all existing countries and cities is relevant and gives it a massive advantage over any local competitor. Once Airbnb’s network effects are working globally, any new entrant needs to enter and build liquidity globally to compete. This is a large part of the reason Airbnb has few (if any) meaningful international competitors. Meanwhile, TaskRabbit has many competitors, both in the US (Thumbtack, Handy) and internationally (Airtasker in Australia, Bark in the UK, now also in the US). And Uber ended up losing many of its international competitive battles: against Didi in China, against Grab in South-East Asia, etc. Uber’s large installed base of riders and drivers in the US gave it no real network effects advantage in those regions, and ultimately the regional competitors won out because they had the advantage of better local knowledge.

As a quick application of these distinctions, consider Boatsetter, a marketplace that allows users to rent private boats (with or without a boat captain) from boat owners. Does it enjoy local or global network effects? We don’t have any particular insight into the behavior of Boatsetter renters, but a key question we would ask in order to evaluate the strength and defensibility of Boatsetter’s network effects is: how many Boatsetter renters are local vs. visitors? The higher the percentage of Boatsetter renters that are out-of-town visitors, the more global the network effects are, and so the stronger and more defensible they are.

To further refine our understanding of local vs. global network effects, consider Shappi, a platform which allows buyers in South America (Ecuador only for now, presumably expanding to other countries soon) to purchase products only available in the US and have them delivered by verified travelers. In short, a buyer in Ecuador might buy an iPhone online and have it delivered to Shappi’s US warehouse. Shappi then assigns a verified traveler that plans to travel to Ecuador soon (travelers share their planned trips with Shappi), ships the iPhone to the traveler, then the traveler takes the iPhone to Ecuador when traveling there and delivers it to a Shappi representative, who then delivers it to the buyer. Typically, a traveler would deliver multiple products (for multiple buyers) to Ecuador and Shappi pays them based on the number and weight of the products they deliver.

Are Shappi’s network effects local or global? In some sense, they appear global because they cross international borders: Shappi enables interactions between buyers in Ecuador and travelers from the US to Ecuador. A buyer from Ecuador values Shappi more when Shappi has a larger network of travelers from the US to Ecuador. And vice versa, a traveler from the US to Ecuador values signing up with Shappi more when it has more Ecuador-based buyers, because that means opportunities for higher payout.

However, it is quite apparent that the network effects are mainly tied to specific “routes”, such as US-Ecuador. Assuming Shappi expands to other South American countries, a buyer in Columbia would care about travelers from the US to Columbia, who are most likely distinct from the travelers from the US to Ecuador. Conversely, travelers will most likely not change their travel plans based on Shappi delivery opportunities. A traveler who plans to go from the US to Columbia only cares about the number of Columbia-based Shappi buyers and not about the Ecuador-based Shappi buyers.

This means Shappi is not nearly as defensible as a platform with truly global network effects like Airbnb. One could start a competitor to Shappi for the US-Peru route and be evenly matched with Shappi. Meaning that Shappi’s existing network of buyers and travelers for the US-Ecuador route gives it no advantage on the US-Peru route. In this sense, Shappi looks more like TaskRabbit: its network effects are largely “local”, where here “local” means specific to a route.

One could argue that some travelers may go to different destinations, so a Shappi traveler that went to Ecuador this month might go to Columbia next month. First, we doubt there are many such travelers that use Shappi. If there are, Shappi would be well advised to focus on expanding to routes that share travelers in common. Second, even if there were many such travelers, Shappi buyers still only care about US travelers coming to their specific country, so the network effects remain far less global than Airbnb’s.

Key take-aways

There are many shades of gray between local and global network effects. In general, whether a platform’s network effects are closer to local or global depends on several factors, including the nature of the products or services being traded, transportation costs, language barriers, the share of users that use the platform in more than one location, etc.

Thus, marketplaces for digital products have inherently global network effects – see for example Open Sea, Uniswap. So do marketplaces for digital services or services that can be performed remotely – see Upwork, Fiverr, Magic.

Meanwhile, when it comes to marketplaces for physical products, the key question to ask is: how much demand for trade is there between different locations? An important factor that determines the answer is the magnitude of transportation costs relative to the value of the products being traded. Etsy has global network effects because it facilitates the trade of relatively small items (e.g. jewelry) which can be shipped internationally at reasonably low rates. B2B wholesale marketplaces like Alibaba and Faire.com also have global network effects because they facilitate the trade of bulk orders. By contrast, the network effects of car marketplaces for individual consumers (e.g. CarGurus, eBay Motors) are largely local.

Finally, for in-person services, the questions to ask in order to determine whether a platform’s network effects are closer to local or global are:

• Do participants on one side interact with participants on the other side from many different locations or just one? (see the discussion of Shappi above)

• For every given location (e.g. city), what is the share of participants on each side that are local vs. visitors from out-of-town (or out-of-country)?

More fundamentally, in our view all these cases boil down to figuring out whether the platform in question accumulates a network effect advantage as it expands to new locations. To what extent does the network from one location provide an advantage when “opening up” a new location? The answer is “a lot” for platforms with global network effects like Airbnb, Alibaba, Etsy and Upwork, vs. “not much if at all” for platforms like TaskRabbit and Shappi. For ride-hailing platforms like Uber or food-delivery platforms like DoorDash, the answer is “somewhat” when it comes to opening up new cities in the US, and “not much” when it comes to opening up new countries.

This discussion also suggests that companies can be strategic in how they expand, in order to maximize the instances of global network effects as much as possible. For example, Gable connects companies that wish to rent collaborative workspaces for their remote employees with hosts who own suitable spaces. On the surface, Gable’s network effects appear to be mostly local, but it can choose to pursue companies that are present in multiple cities and expand to those cities where its existing company-clients have workers, thus deliberately taking advantage of inter-city network effects. And in some cases companies can take action to at least partially expand the reach of their network effects from local to global. For example, eBay Motors offers remote inspection and shipping services via partnerships, which can facilitate more cross-country transactions.

***

The combination of user data and AI often creates data feedback loops. This means that as a firm gathers more customer data, it can feed that data into machine learning algorithms to improve its product or service, thereby attracting more customers, generating even more customer data. Think, for example, of search engines: the more people search on Google and click on the links provided, the more data Google gathers, which allows its algorithms to provide more accurate and relevant search results, attracting even more users and searches, and so on.

We have written here and here about the various conditions under which such data feedback loops can lead to sustainable competitive advantage. But the strength of these data feedback loops can vary greatly, and companies can make deliberate choices in their products or services to strengthen them. How to do so is the focus of this piece.

Not All Data Feedback Loops Are Created Equal

Some products have naturally very strong data feedback loops. Think of smart thermostats, where every temperature adjustment by a user provides a valuable data signal that the device can use to achieve better personalization. Or Google Maps, where every user’s choice of route and the time taken to reach the destination help the algorithm improve its route recommendations and traffic predictions. Or Spotify, whose recommender system learns directly from users’ choices of which recommended songs to include in their playlists and how often they listen to those songs. The reason these feedback loops are so strong is that users reveal clear and unambiguous signals of their preferences in the natural process of consuming the product, which are leveraged to further improve the product or the service for those users.

At the other end of the spectrum are products with naturally weak data feedback loops. Their usage is hard to track or does not reveal useful information about user preferences, or it is difficult and slow to gather informative feedback from users.

This is obviously true of traditional “dumb” products like cars, furniture, and clothes: They are not digitally connected, so the only way to create any data feedback loops from customers is to manually collect feedback via focus groups and surveys, which can only help for future product releases. And it is also true of products where the feedback loops involve very long cycles of learning and improvement such as financial institutions’ credit scoring systems (feedback mainly comes from defaults, which take years to materialize), or venture investing (it takes years to figure out which startups succeed and which fail).

Less obvious is that even some digital products that collect lots of user data may have weak data feedback loops. This means most of their value comes from pre-programming and in-house data training and does not increase much, if at all, via learning from users. For instance, popular wearables such as Fitbit, Whoop, Nutrisense, and Oura have fairly weak data feedback loops, even though they provide lots of data insights.

Consider the latest Fitbit tracker, the Charge 5, which has an impressive list of features. The tracker measures a wide array of data from the user (heart rate, speed and distance of movement, sleep, skin temperature) and provides valuable insights such as heart rate variability, time spent in different heart rate zones, a cardio fitness score, a readiness score for working out, a quality of sleep score. However, these insights do not appear to get better with more usage or more users. They are straightforward summary statistics of what is being measured by the tracker, or the result of comparisons between user data measured by the Fitbit tracker and relevant reference points based on pre-existing research, which are pre-programmed in the Fitbit system.

For example, Fitbit explains that “the cardio fitness score is determined by your resting heart rate, age, sex, weight, and other personal information.” Similarly, the readiness score is based on the user’s recent sleep patterns as measured by the tracker — which are presumably compared to some stable reference points. In other words, these are estimates of the “true” state of a user, but they do not improve with more usage data, simply because Fitbit has no way of figuring out how close its estimates came to the “true” state and adjusting them accordingly.

Of course, there may be some limited ways in which wearables could improve the value provided to each user based on usage data from that user or other users. At the most basic level, any connected product can learn which user interface is more engaging (e.g., via A/B testing), which is technically a form of data feedback loop, albeit a limited one — table stakes for most products.

And there are correlations that wearables could analyze to recommend behavior. For instance, a correlation between the time users go to bed and the quality of their sleep, which can be used to recommend an optimal time to go to bed; or between the time users exercise and the quality of their sleep, which can be used to recommend an optimal time to exercise. In these cases, as users adjust their behavior based on the recommendations, the wearable gets some feedback on whether this is helping or not, and so can further improve by learning from users. Still, the wearable provider will struggle to make reliable recommendations without additional data about or control over the user’s relevant environment.

Finally, consider the data feedback loops of large language models (LLMs) such as Open AI’s Chat GPT or Google’s Bard. These models ingest enormous amounts of data from the web and use machine learning models to generate answers to user questions. In the early versions of these models, their capabilities are largely determined by pre-launch “in-house” training and testing — the quality of their answers improves only to a moderate extent with more users or more usage. Indeed, as of this writing, there are two main mechanisms that create data feedback loops around LLMs:

1. Users click on the thumbs-up or thumbs-down buttons at the end of the answers. (Most users probably ignore this most of the time.)

2. Users ask follow-up questions, which could signal whether an answer was satisfactory or not. (This inference is likely difficult in most cases.)

Note the difference with regular search (on Bing or Google), where the user’s choice of which displayed links to click on provides a much clearer signal of the relevance of the search results. Of course, things may change drastically as the LLMs are updated and start incorporating more reliable ways to generate data feedback loops — a topic we discuss in the next section.

How to Enhance Data Feedback Loops

The key question is then, for those products where the data-enabled learning feedback loops are not naturally strong, what can be done to enhance them?

(Re)design the product to create natural data feedback loops.

Ideally, one would want to (re)design the product or service in such a way that customers, in the natural course of using the product, are creating data that signals how useful/effective the product is for users. This data can then be used by the provider to improve the quality of the product or service.

For example, LLMs could add new features that allow users to save and organize the responses they found most helpful into folders of favorites (akin to bookmarks) and delete those they don’t want to keep. They could add a document-creation feature (akin to Microsoft Word) where they would copy and edit LLM responses. They could also create challenge games (AI vs. users) and leader boards where the AI and users seek to answer questions, and users vote on the answers. And so on. The idea is to create opportunities for users to provide reliable signals of the perceived quality of the LLMs’ answers, which can then be used to improve their algorithms. Over the past week, Google’s Bard took a step in this direction by introducing the option for users to modify each response by selecting whether they want it to be shorter, longer, simpler, more casual or more professional.

Fitbit and other wearables could add standardized fitness tests or challenges (e.g., run one mile or do three sets of repetitions of 20 sit-ups) and measure users’ total time, heart rate, and other biometrics before and after. This would allow the wearables’ AI to more accurately predict users’ readiness for working out as they observe more data from more users. Which in turn should induce users to rely on the wearables more and more for deciding when to work out. Of course, this still relies on assumptions about unobserved user behavior, but the key point is that the devices need to perform some sort of quasi-experiments on users in order to obtain data that can help them learn and improve.

Integrate with other products to create data feedback loops.

In many cases, redesigning the actual product or service to engineer feedback loops may be challenging (as illustrated by the wearables example above). An alternative way of achieving the same goal is to integrate your offering with other existing products that customers already use or could use.

For example, Fitbit (or Whoop or Oura) could create an integration with smart thermostats to enable the wearable to automatically control the ambient temperature during a user’s sleep. This would allow the wearable’s AI to adjust the temperature and determine its effect on the quality of a user’s sleep as measured by the wearable. The more a customer uses the wearable (i.e., wears it to sleep), the closer the wearable can come to figuring out the ideal temperature pattern for any given user. This can be achieved by automatically experimenting with many different temperature patterns as the user sleeps.

Or the wearables could integrate with Peloton or Tonal, which provide standardized workouts, so that Fitbit’s AI could correlate a user’s biometrics with the type and intensity of workouts. The advantage relative to the option described above (where the wearables are simply asking users to perform specific workouts) is that here the actual workout behavior can be observed.

Similarly, LLMs could integrate with whatever software or tools their answers are being used in. For example, they could integrate with content creation/editing software (e.g., Google Docs, Substack, Salesforce), which would allow them to observe which parts of their answers end up being used in the content created by users and how users modify their answers, and use that data to improve their answers.

Ask users for feedback in an incentive-compatible way.

Still, for many products, finding such ways of making user feedback inherent to product usage — directly or via integration — may be hard or impossible. Short of that, the next-best option is to explicitly ask users for feedback. Most online products and services do this to a certain extent. As mentioned above, LLMs ask users for thumbs-up or down after every answer generated; Netflix asks users for a thumbs-up or down after each piece of content they watch in order to improve its recommendation system; Amazon.com asks buyers to rate the products they purchased; Airbnb asks both travelers and hosts to rate each other, and so on.

Of course, the difficulty with asking for explicit feedback is obtaining useful information while providing some benefit to users that makes up for the inconvenience cost of providing that information. Nobody likes being badgered with surveys that “only take three minutes of your time,” or seeing feedback polls (“how likely are you to recommend us to your friends?”) pop up every other screen – unless they derive a clear individual benefit from them.

In addition to making requests for user feedback as easy and painless as possible, it is also important to clearly communicate to users how their feedback will benefit them individually (e.g. “by providing ratings of the following specific features, Booking will be able to recommend you the most suitable hotels for your future stays”). Just saying “we’d love to hear your feedback” or “rate these features to improve our recommendations” isn’t enough. While some altruistic users may be motivated by knowing their feedback benefits the company or future users, many will not be.

The key is to emphasize the customized nature of the learnings derived from user feedback. And such customization should be meaningful. For example, each Netflix user should be able to ultimately see how their movie ratings influence the future movie recommendations provided by Netflix. One can even take this one step further by giving users some editorial control over how the AI systems process their preferences. For example, Netflix could allow users to edit their preferences on movies, which they may update over time, and Booking.com could allow users to edit what features they care most about when booking travel. All of this ensures that users have an incentive to provide more informative feedback.

Include humans in the loop.

An important way to manufacture a data feedback loop while minimizing the feedback burden on users is to include humans in the loop to complement (or even replace) user feedback. A good example of this is Alexi, an AI-powered legal research service offered to law firms. Customers can submit legal questions along with any relevant case facts, and Alexi sends back a legal research memo containing an answer, complete with summaries of the relevant case law and litigation. The memos are generated by Alexi’s AI but then reviewed and amended (when necessary) by Alexi’s in-house legal team — the humans in the loop. Thus, Alexi’s AI gets the benefit of learning from customer queries and the implicit feedback provided by the in-house team’s corrections, without burdening customers. This works well here because customers do not expect instant answers. (Alexi promises an answer within 24 hours, a very reasonable turnaround in this context.) Nor do they ask hundreds of questions a day like Chat GPT or Bard users might.

Another example of good use of humans in the loop is Grammarly, an AI-powered writing assistant that helps users improve grammar, spelling, punctuation, and style. Grammarly provides suggestions in real time, while users are writing in any application or site (e.g., word processing, email, social media, communication apps). Users have the option to accept or reject the suggested edits, which obviously helps improve the algorithm and creates a data feedback loop. However, Grammarly also uses human reviewers to check the suggestions made by the AI model and review ambiguous cases (a user rejects a suggested edit, but it is unclear why he did so), instead of asking users for feedback.

Other examples of AI-based products or services that use humans in the loop to complement or replace user feedback include content moderation on social media platforms (human moderators get involved in the most complex cases that the AI has trouble with), AI-powered radiology services such as Gleamer (human specialists complement AI diagnostics and would correct any issues in the diagnosis the system automatically writes up), and AI-powered security services such as Deep Sentinel (the AI detects threats and decides when to escalate to human guards, who can then decide whether the AI made the correct call).

Of course, the main weakness of the humans-in-the-loop approach is that it does not scale well to hundreds of thousands or millions of customers, or to services with very high frequency of usage and expectation of quick turnaround times. Not many companies can afford to hire armies of over 15,000 content moderators as Meta does for Facebook. This is why most companies employing humans in the loop try to minimize the time required from humans in the process. Nevertheless, the humans-in-the-loop approach can be very effective in the early stages of a product/service, when the learning curve is steepest. The hope is that the need to involve humans decreases rapidly over time as the AI system learns.

Concluding thoughts

The increasing availability of artificial intelligence, including machine learning algorithms, means that deliberately creating data feedback loops is now possible for most products and services. For some products, it is easy; for others, one needs to find more creative ways to engineer the data feedback loops — via integrations or minimally intrusive incentive-compatible requests for user feedback. When they are strong, these feedback loops can create a form of network effect (more users bring more data, which makes the product better, in turn attracting more users, and so on) and – provided other conditions are met – compounding competitive advantage.

In essence, API platforms provide a way for customers to more easily integrate with multiple service providers’ offerings. Put differently, they provide the “information pipes” for the needed connections. Thus, instead of a customer (say, a fintech company) having to build their own integrations with all the service providers (in this case, potentially thousands of financial institutions so that all the fintech company’s customers can transfer information from their accounts to the fintech company), they just have to integrate once with an API platform (Plaid), which takes care of all the “plumbing”.

On the surface, API platforms look like amazing network effect businesses. For example, Plaid claims it connects over 5,500 fintech applications (Acorns, American Express, Betterment, Coinbase, Venmo, etc.) to over 12,000 financial institutions in the U.S. alone (big banks like Chase, Wells Fargo, Bank of America, as well as smaller banks and credit unions). One quarter of all U.S. adults has already connected an app with Plaid, and that number grows daily.

But do these API businesses really have powerful and defensible network effects? Indeed, are they really platforms at all?  

One-sided outsourcing service vs. two-sided platform

To answer these questions, it’s useful to take a closer look at Plaid. One can think of Plaid as having gone through two stages.

In the first stage, represented in figure 1 below, Plaid was essentially a one-sided outsourcing service. Specifically, it took on the work of any financial app wanting to allow its customers to link the app to their bank accounts. This might be so the app could verify their account balance, or analyze their recent transactions, for example. We say “one-sided” because this was typically done without the banks actively participating. For example, Plaid would offer a virtual browser connected to the bank’s website, where the consumers of the fintech app (say, Robinhood) could log into their bank account for the purpose of sharing some of their bank account information with Robinhood. After the consumer logged in, Plaid could scrape the account details, bank balance or past transactions off the page, and send the information back to Robinhood on behalf of the consumer.

In this first stage, when the banks were completely passive, defining a network effect doesn’t even make sense: we can’t talk of more banks joining Plaid. Instead, Plaid’s main defensibility came from the fixed costs of building the virtual browser and the scraping technology that worked with each of the fintech customers’ banks, including updating it every time a bank changed its user interface. There was certainly no defensibility from network effects. The main task was to recruit fintech apps that wanted to outsource the work of linking to their customers’ bank accounts to Plaid. By not having to also get banks on board, Plaid solved the chicken-and-egg problem, and was able to focus on just attracting fintech apps initially. This also means that given our definition of platform businesses, Plaid was not a platform business at this stage.

In the second stage, represented in figure 2 below, Plaid got the banks to make deliberate efforts to provide improved and continuously updated APIs. This ensured the linkages between fintech apps and banks always worked, even when banks changed their user interface, and that customers’ data did not have to be saved by Plaid (which was better for security reasons). An example of a bank working closely with Plaid is U.S. Bancorp, and their agreement ensures that the bank’s permissions hub and Plaid’s portal will be constantly in sync. The impetus for U.S. Bancorp (and other banks) to do this is that they compete to offer their customers the best service: an increasingly important part of that is ensuring they seamlessly and securely work with the multitude of fintech applications their customers expect to be able to use with their bank account.

The result is that at stage two there is a cross-side network effect running between fintech apps and the banks. The more banks join Plaid (and keep their APIs working), the more valuable Plaid becomes for fintech apps since their customers can link their services to their bank accounts smoothly. The more fintech apps join Plaid, the more attractive it becomes for a bank to make an effort to integrate with Plaid, since the investment in keeping their APIs updated and working smoothly has a greater payoff for their customers.

Once banks start making relationship-specific investments in Plaid (like U.S. Bancorp), these cross-side network effects provide real defensibility. But the defensibility is somewhat limited in our view. A key driver of Plaid’s value is that many banks rely on arcane information technology systems, which require significant effort to connect or integrate with. As the banks’ systems are updated, they will likely be able to provide standardized APIs that will also work with Plaid competitors, and indeed may mean that eventually fintech apps can bypass API platforms altogether. Moreover, Plaid does not provide discoverability: fintech apps and banks do not use Plaid to discover each other, which further limits the scope for network effects.

It is worth contrasting Plaid with another API platform, Zapier, which has more compelling network effects in our view. Zapier helps customers (mainly business users) connect apps like Gmail, Slack, Dropbox, Discord, Twitter, Zoom, Airtable, Stripe and so on, via APIs, so that they can more easily move information between these apps. It supports over 3,000 such apps and over 1 million users. Users cover a wide range of firms (and their employees) that want to automate tasks that work across combinations of different apps (called zaps). An example of a zap would be monitoring whenever a Twitter profile tweets something that mentions the company’s name, and Zapier sends a notification to a Slack channel so that employees can follow up from the company’s twitter account if necessary. Cross-side network effects arise between users and apps.

In contrast with Plaid, the integrations that an app may need to work with other apps are not so easily standardized. This is because of the difficulty of different software combination working together seamlessly. With thousands of apps, there are millions of different possible integrations between any two. As a result, Zapier needs to rely on apps to self-serve to a certain extent, joining Zapier and constantly updating their APIs. As Zapier says on their website, “two new apps join our Platform every day, and your Zapier integration connects you to them.” Zapier also provides discoverability through their app directory, which means the network effects are not limited to integrations. Apps can also get certain benefits, including being promoted by Zapier, by earning a higher status in Zapier’s partner program. An app’s status level in the program is determined by how many people use its integration and the “health” of the integration.

In addition to a stronger and more defensible cross-side network effect, Zapier also enjoys a same-side network effect between the apps themselves, given they have to work together. An app is more likely to join Zapier, the more other (complementary) apps are on Zapier, because that means there are more apps to integrate with, which in turn increases the value of the app in the eyes of its users. Finally, unlike the banking APIs on Plaid, there is almost no limit to the number of ways different apps could work together, especially as new apps are added every day. However, similarly to Plaid being vulnerable to banks creating their own standardized APIs, Zapier is vulnerable to app providers (especially the large ones like Dropbox, Notion, Slack and Zoom) providing their own APIs to facilitate integrations with other complementary apps.

Concluding thoughts

In evaluating new startups that claim to be API platforms, it is important to analyze each in detail to understand the power and defensibility of their network effects, if any. For network effects to exist and provide defensibility, the service providers involved (e.g. banks in the case of Plaid, apps in the case of Zapier) must make relationship-specific investments (e.g. customizing and updating APIs for that particular platform), which they wouldn’t want to duplicate for a rival platform with significantly fewer customers. Otherwise, if the service providers don’t have to take any active (and costly) steps to participate on a given API platform, then the latter looks like a glorified outsourcing service – it might benefit from important scale economies, but certainly not from network effects.

Moreover, like for any business, it is useful to evaluate the scope for extending the network effects on these API platforms, in order to enhance growth and defensibility. One way in which API platforms like Plaid and Zapier can do so is by offering a service to their customers’ customers, i.e. to end-users. For instance, Plaid has recently begun offering U.S. consumers the ability to monitor and control the information that is being shared by the banks and fintech apps they use. By forming a direct connection to end-users, Plaid can add a third side to its platform, and build a stronger moat around its business. For instance, it could go further and enable consumers a way to discover fintech apps that work with their bank accounts, which would create an additional network effect between consumers and fintech apps.

In today’s post we further illustrate these tradeoffs with another decentralized platform, Audius. It raises a number of interesting issues that we didn’t discuss in the previous post.

What is Audius?

At a high level it is a decentralized music platform developing a community-owned and artist-controlled music sharing protocol. One can think of it as a blockchain-based SoundCloud alternative, providing a way for artists to distribute their music directly, and to cut out intermediaries (such as music labels) that take the vast majority of the revenue streams created by artists. Founded in 2018, today it has more than six million unique users a month and over 100,000 artists.

The key defining characteristic of Audius from our perspective, compared to, say, the earlier version of SoundCloud, which also tried to cut out intermediaries in the music industry, is that it is a decentralized platform. As explained in its whitepaper, its main stakeholders are:

• Artists (musicians): Using Audius, artists can generate timestamped records for their creative works and register them to the Audius network. Because the network is decentralized and secured by blockchain technology, the idea is that the metadata attribution remains safe, verifiable and immutable.

• Fans (users listening to music): For a user that isn’t into crypto, the music player works like any other music player. You can create an account, follow your favorite artists, and listen to their tracks.

• Node operators: these are professional operators of two types. The first type of node operators host content and make it accessible on behalf of artists, whereas the second type index the content and provide an easily searchable interface for retrieving metadata.

• Apps: Audius is by design committed to open and free APIs, which third-parties can use to build apps on top of the network, allowing users to interact with the content in all kinds of ways on all kinds of devices (e.g. a music racing app). Audius (the company) has its own music streaming app, but the majority of listening on Audius is now via third-party apps, including third-party music players.

• The Open Audio Foundation: a foundation formed to steward the long-term growth of Audius. It is endowed with a fixed supply of tokens for incentives and growth programs.

The Audius token ($AUDIO) serves to align the incentives of these stakeholders by motivating them to take actions that increase the value of the entire platform. For instance, artists can earn tokens by uploading tracks that cross a certain level of engagement, and users can earn tokens by curating playlists that become very popular or by building API integrations. And, as with other decentralized platforms, the token is also used for governance. For a governance proposal to pass, at least 5% of all staked $AUDIO must vote on the proposal and more than 50% of the votes must be in favor of the proposal.

Tradeoffs

A fundamental benefit of decentralization is the commitment power to not exploit platform participants – or more generally, to not change the rules on them without giving them a say. This means decentralized platforms are most likely to thrive in industries where participants have been burned by centralized platforms (through increased fees, changes in algorithms, being de-platformed etc.).

Audius is a good illustration, as it was designed to stand in sharp contrast to Apple Music, Spotify and SoundCloud. The difference with Apple Music and Spotify is quite obvious since those services do not allow any meaningful direct interaction between artists and their audience (e.g. Spotify only allows artists to send one email to their audience a year, in a standardized format). The comparison with SoundCloud is more interesting. As we noted earlier, SoundCloud also initially set out to provide artists a way to directly connect with users. The difference is SoundCloud was a centralized platform, that could change its design and governance rules at any time – and it did. Some of the changes SoundCloud made over the years that upset its users and artists include:

• It revamped its website in 2013 to emphasize usage and the interests of mainstream artists over the interests of its original (indie) artists who had helped grow the platform in the first place.

• In order to appease music labels so it could offer a paid subscription service that could rival Spotify, SoundCloud went to great lengths to reach agreements with the major music labels. It allowed certain copyright holders to unilaterally remove paid subscribers' content without recourse. For instance, Universal Music Group had the right to take down any file or any account (premium or free) on SoundCloud, regardless of whether it violated copyright laws.

• SoundCloud phased out groups in 2016, upsetting many indie artists, for whom groups were the most effective way to share music on SoundCloud.

The primary driver of these changes was SoundCloud’s decision to try to compete with Spotify and Apple Music in the mainstream music market, given the much higher revenues and valuations these services were attracting. To do so, it catered to the interests of labels and investors rather than those of its original artists and users.

Audius’ decentralized structure is designed to avoid precisely this type of situation. After the launch of the Audius main service in October 2020, Audius the company has no longer had any ability to make unilateral changes to the service. Any change must be voted on by the network according to the governance system briefly described above. In this sense, Audius is clearly committed not to behave opportunistically.

On the other hand, decentralization also carries some downsides. One is that decentralized platforms cannot be very nimble in making design improvements due to the need to get stakeholders to vote on each decision, including some relatively minor ones. For example, in the case of Audius, the user experience leaves a lot to be desired: something as simple as resetting a password is quite confusing and not secure, and Audius’ own music player is rather limited. As mentioned above, Audius gets around this limitation by enabling other developers to produce their own music players on top of the Audius database, some of which are much better than Audius’ own player.

The other downside we discussed in our previous post is the lack of legal certainty. Could Audius be sued if there is a copyright dispute over the illegal distribution of music on one or more of its distribution nodes, or would the operators of the nodes be sued? Similarly, who bears responsibility for commercial deals between Audius and other companies? For instance, Audius signed a deal with Tik Tok in 2021 to allow artists to easily upload tracks from Audius to their Tik Tok accounts. This seems to have been negotiated by Audius the company without requiring stakeholders to vote. What happens then if TikTok and Audius get into a complicated commercial dispute? Having a centralized authority to do the negotiation and to enforce the terms of a commercial deal might make more sense.

Implications of the openness of Audius’ model

As noted above, in addition to being decentralized, Audius is also an open platform in that it allows third-party apps to freely build on top of its network. Its long-term focus seems to be on building the underlying content ledger and governance infrastructure, and allow user-facing apps to be built on top.

It is worth emphasizing that it is not openness per-se that distinguishes Audius from centralized platforms. Clearly, centralized platforms could choose to be open to a certain extent, like OpenSea and Android have done. Audius is different in that it is credibly committed to remaining open --- indeed, changing the openness of its APIs would require the approval of its stakeholders (which is unlikely to the extent they benefit from openness). By contrast, there are many examples of centralized platforms that started with open APIs to encourage the building of their networks, only to close them off later (most prominently, Facebook and Twitter).

Of course, a platform may want to close APIs if the apps building on top of it become too powerful and it fears losing critical control points. This can be an issue for Audius’ open design at some point in the future. Imagine a for-profit music player built on Audius that becomes very popular and dominates usage by artists and fans. Such an app could (at least in principle) start convincing artists to host their music exclusively on its service and not on Audius, thus eventually diverting both users and artists away from Audius’ network. We have discussed a similar risk in the context of turning products into platforms by opening the door to third-parties. If this were to happen, the Audius founders and executives would be unable to unilaterally close off access to such an app. However, if they could garner the support of a voting majority, they could still do so. Thus, the possibility of dealing with such threats is not ruled out by decentralization. Rather, under decentralization, it would not be up to the whim of the CEO or a board of directors to decide the fate of third-party apps.

Concluding thoughts

Despite the apparent benefits of a decentralized platform for music distribution over existing centralized platforms, we don’t think Audius will disrupt the major music labels or services like Spotify and Apple Music anytime soon. Since Audius leaves it to artists to build and manage their own audiences, it does not work well for those artists that want a lot of support and upfront capital to try to become the next big star. Similarly, the next big manufactured band trying to follow in the footsteps of AKB48, Backstreet Boys, BTS, and the Spice Girls is unlikely to be born on Audius. And to the extent there is popular demand for such music, there will always be a whole industry built around music labels and streaming services to support them. Rather, Audius will likely co-exist with such services, by addressing a different market segment – artists that cater to more niche audiences and who previously found it difficult to make a living from their music.

Finally, while Audius’ white paper paints an appealing vision of how artists’ and users’ interests can be protected via decentralization, many details remain to be put into practice and stress tested. Currently, on-chain voting seems to be restricted to node operators and the Open Audio Foundation, although anyone holding $AUDIO tokens can delegate them to a node operator for the purpose of voting. This means that for now individual users and artists can only affect Audius governance decisions indirectly via off-chain discussions in Discord and by delegating their tokens to particular node operators (a form of representative democracy). It remains to be seen how robust the Audius network will be in cases where different stakeholders have conflicting interests. Looking through the record of 48 on-chain proposals to date, there has yet to be a single proposal where there were votes both for and against. Democracy works just fine when everyone shares the same view.

How can Corning’s Gorilla Glass be transformed into a platform by reaching out to its customers’ customers?

This week we circle back to one of our favorite topics, product-to-platform. In particular, we discuss how Corning’s Gorilla Glass can be transformed into a platform using the method of reaching out to customers’ customers. Recall this means creating some sort of marketplace between its customers and its customers’ customers (e.g. Open Table, Shopify).

Why Corning’s Gorilla Glass? Because it is an industrial, non-digital and B2B product, which a priori would seem quite difficult to transform into a platform. That makes the exercise all the more interesting. And it is meant to illustrate the point that even non-digital and boring products can benefit from (partially) moving in the direction of platform businesses. The goal is not to become an Alibaba or an Amazon or a Facebook, but to increase defensibility and accelerate growth by incorporating some platform features.

Gorilla Glass?

Gorilla Glass is the best-known screen glass for laptops, tablets, smartphones, wearables, and other devices. In particular, its customers include Apple, Samsung, Acer, Dell, Lenovo, etc. These customers incorporate the Gorilla Glass in their devices, which they sell to end-users.

Interestingly enough, Corning has been advertising Gorilla Glass to end-users for years, even though it does not sell anything to them. Here are some examples:

In fact, one of us initially became aware of the Gorilla Glass after seeing its full-page ads in The Economist magazine. This is of course an example of ingredient branding, reminiscent of Intel’s classic “Intel Inside” marketing campaign.

Turning Gorilla Glass into a platform

So how can Gorilla Glass be turned into a (partial) platform?

The first method for doing so, opening the door to third-parties, seems quite limited in this case. There are no obvious complementary products or services to the Gorilla Glass that could be best supplied by 3rd-parties to Corning’s customers alongside the glass.

The second product-to-platform method, connecting customers, is arguably the least relevant here. It’s hard to think of any interactions/transactions between Corning’s customers (large OEMs) that would benefit from being mediated by the Gorilla Glass.

It’s the third method for turning products into platforms, reaching out to customers’ customers, that is most interesting here – and the primary reason we chose this example.

Note in particular that reaching out to customers’ customers is significantly more difficult for Corning than for other B2B products we have discussed (e.g. Shopify, Indigo Agriculture, Open Table, Teachable). This is for two main reasons:

• As good as the Gorilla Glass is, it is not the most important component of the final product (smartphones, laptops). Nor is it a digital component that is a natural candidate to become a platform, like, say, an operating system or a microprocessor or a software tool that mediates the customers interactions with their customers (e.g. OpenTable, Shopify, Teachable).

• Corning’s customers are very large companies (much larger than Corning), who would likely thwart any blatant attempt to take control of their customers via a broad-scope platform. Again, this is very different from Shopify, Indigo Agriculture, Open Table and Teachable, where customers were very small businesses or individuals.

One might ask whether the “Gorilla Glass” marketing campaign doesn’t by itself create a platform by reaching out to customers’ customers. The answer is not really. Ingredient branding can be very powerful in creating brand awareness in the eyes of end-customers, which then puts some pressure on Corning’s customers to include Gorilla Glass in their devices and possibly to pay a premium for it relative to alternatives. This undeniably creates some defensibility, but does not generate network effects.

Either way, we believe an “ingredient” supplier like Corning can go further in terms of reaching out to its customers’ customers. Specifically, Corning could prompt end-users to register their device featuring Gorilla Glass on its website or app, where they would 1) receive some credit or discount for buying other devices with Gorilla Glass, and 2) be able to browse and discover (and possibly purchase) such devices.

Two key points:

• Only offering a monetary benefit (e.g. insurance or warranty for the glass) that covers the current product purchased by the end-user would not be enough. That would add value to the end-user and possibly allow Corning to extract some additional money from them, but would not create any network effects. It is indeed important that the monetary benefit comes from buying multiple products that feature Gorilla Glass, so that in a sense, the end-user adopts Gorilla Glass as they would adopt a platform.

• Only having a website or app that lists devices featuring Gorilla Glass would not be enough either. This would simply be an extension of the “Gorilla Glass inside” marketing campaign – great for raising awareness and possibly getting customers to pay a premium for Corning’s products, but not really moving in the direction of a platform with network effects.

So does the proposal above create network effects? Yes: any device maker that chooses to rely on Gorilla Glass benefits when more end-users accept the Gorilla Glass offer, as they are more likely to find out about its products and might be more inclined to buy them given the financial reward from buying another device with Gorilla Glass. And conversely, end-users obviously are more likely to take the offer when there are more brands and devices that use Gorilla Glass.

As we’ve discussed in our post on reaching out to customers’ customers, the biggest challenge with this method for transforming products into platforms is overcoming potential resistance by the company’s customers. Would Apple, Samsung, Lenovo and other Corning customers resist the approach described above? We think it’s unlikely. As explained earlier, Corning offering end-users some credit or discount for buying other devices with Gorilla Glass can be argued to increase the sales of Corning’s customers. Of course, in theory, an individual customer like Apple could worry that Corning’s offer might make some Apple users more likely to consider competitors’ devices. However, we don’t think this would or should be a realistic source of concern: Corning’s customers are large companies with well-established brands, for whom the risk of commoditization by a possible Corning platform is minimal (by contrast with, for example, Shopify’s position vis-à-vis its small business customers).

Concluding thoughts

While our product-to-platform proposal is unlikely to turn Gorilla Glass into the screen tail that wags the smartphone dog, it would help enhance its defensibility and possibly accelerate sales growth. Furthermore, our proposal for Corning can be applied (with suitable adaptations) to any other component supplier or manufacturer, whether or not they have resorted to ingredient branding (e.g. Gore-Tex clothing technologies, NutraSweet’s sweeteners, Microban’s anti-microbial technologies, etc.).

In this post, together with economist and blockchain expert Hanna Halaburda, we evaluate the advantages and drawbacks of decentralized platforms, relative to traditional, centralized platforms. 

Examples

It is useful to start with three examples of decentralized platforms: a marketplace for products (Rarible), a marketplace for services (Braintrust), and a social network (Minds).

Rarible

As we’ve discussed in a previous post, Rarible is an NFT marketplace that has opted for decentralization to compete with OpenSea’s centralized approach. Since its founding in early 2020, Rarible has steadily moved towards becoming a full-fledged DAO. It incentivizes people to build on top of their open protocol, rewarding the early builders with their token RARI which was released in July 2020. 

The primary function of the RARI token is to give holders the power to submit and vote on proposals that impact the marketplace (the token determines voting power, but need not be spent in order to actually vote), to moderate creators, and to curate featured artwork. Currently, every week, 41,250 of RARI tokens are distributed among applications building on Rarible Protocol (including Rarible.com), proportional to sales volumes in a program called Marketplace Liquidity Mining (MLM). Some of these are distributed as rewards to buyers and sellers for NFT trading, with the exact formula outlined here. Buyers and sellers get an equal split of the tokens; verified users get 15% more tokens;  for trades of NFTs created outside of Rarible, buyers and sellers get 50% less. Other tokens have been reserved for airdrops and for investors and Rarible team members. 

An interesting challenge noted in a current proposal being voted on is that most community members convert their tokens into Fiat currency (as they say, “people need to eat!”). As a result, these types of participants end up with little or no voting power. The proposal is to grant a group of five community members 20,000 tokens and to grant the two co-founders of Rarible another 20,000 tokens for the purpose of voting on governance matters. This is viewed as an interim measure until key contributors have built up sizable holdings of tokens themselves. And this illustrates a broad issue with decentralized platforms: how can they sustain incentives for participants to hold tokens for voting purposes if they don’t expect the tokens to appreciate in value, i.e. if there is no speculative reason to hold tokens? 

Braintrust 

Founded in 2018, Braintrust is a two-sided marketplace matching top technology talent (developers mainly) with clients (enterprises) who want to hire them. Its blockchain-based, decentralized approach stands in direct contrast with its centralized counterparts such as Fiverr and Upwork. Braintrust users can earn its BTRST token by building features of the platform, inviting and vetting talent, and referring clients. After discussing ideas in the Braintrust Discord group, proposals for governance changes can be voted on-chain or off-chain. 

On-chain voting can only concern the modification of a limited number of parameters (this is restricted by the BTRST smart contract). Right now, these are: a) admin permissions, b) the marketplace fee charged to employers/clients (there is no fee charged to talent), and c) the categories featured on the platform (currently, they are engineer, designer, product and other). Proposals that pass an on-chain vote are then executed automatically on the smart contract.

Proposals regarding any other governance aspects (not covered by on-chain voting) are handled via off-chain voting in a dedicated discussion tool. Users need to submit proposals, allow several days for discussion, distill into specific recommendations, then vote. Outcomes are decided using rules detailed here. 

As of right now, there is no clear commitment regarding the implementation of proposals that pass an off-chain vote, but Braintrust plans to add more and more decisions to on-chain voting.

Minds

Launched in 2015, Minds’ stated goal is to provide a decentralized alternative to traditional centralized social networks like Facebook. In a 2018 Wired article, Minds was actually described as the “Anti-Facebook”.

Like other decentralized platforms, Minds has its own utility token MINDS (issued on the Ethereum blockchain). Similarly to Braintrust, users can earn the token by posting content (more popular content earns more), being active contributors, finding and fixing software bugs, referring new users, etc. In turn, users can spend tokens to boost viewership of their content on Minds.com beyond their current connections/subscribers (akin to buying Facebook ads to have one’s content reach a wider audience), or to pay other users for access to their content (in the form of tips or subscriptions). 

The governance of Minds is an interesting blend of centralized and decentralized. The terms of service that define what content is allowed and what content is objectionable are entirely defined by Minds.com. It is worth noting that Minds keeps its content guidelines quite short and aims to be committed to free speech. In particular, it does not try to define hate speech or mis-information, as traditional social media networks do, ending up with long and unwieldy content rules. 

The enforcement of content moderation on Minds is mostly decentralized and relies on a jury system. Specifically, when a user reports a piece of content or a channel for inappropriate content, the following sequence is put in motion:

• Minds staff can accept or reject the report.

• If the report is accepted, the user under investigation is notified and given a chance to appeal.

• If appealed, a jury of Minds users votes on appeal - the decision of the jury is final.

For each report, the jury consists of 12 randomly selected Minds active users who are not subscribers to the channel of the user under review. If 75% or more of the jury members vote to accept the appeal, the administrative action is overturned. Otherwise, the administrative action is upheld.

Trade-offs

With these examples in mind, let’s start with the advantages of decentralized platforms.

Credible commitment to providing value in the future 

Perhaps the most obvious and important benefit of decentralization (especially when implemented via smart contracts) is that it can act as a credible commitment not to hold up participants in the way some centralized platforms do. The most important forms of hold-up by traditional centralized platforms are increasing fees over time and/or introducing new types of fees (e.g. DoorDash with restaurants, PropertyGuru with listing agents, and Amazon with sellers as detailed in our earlier post), and unilaterally changing the governance rules and their enforcement in unpredictable ways (e.g. Apple with rules for third-party apps in its App Store, Facebook/Twitter with hate speech and misinformation). 

By contrast, well-designed decentralized platforms can commit (e.g. via smart contracts) to clear and transparent design and governance rules, which can only be changed in restricted ways via majority voting. It is therefore unsurprising that the transaction fee charged to clients is one of the first three parameters subject to on-chain voting on Braintrust as discussed above - indeed, transaction fees are presumably a major concern freelancers have with centralized platforms. Similarly, a key way in which Minds differentiates itself from traditional social networks is through its decentralized content moderation: what users hate the most about Facebook and Twitter is the inconsistent and ad hoc enforcement of their content moderation rules. 

Aligning incentives

Equity is often issued in centralized platforms to align incentives, for example between founders and senior employees. Issuing tokens can be seen as playing a similar role. However, they may align incentives better than equity. This is because tokens can be issued to a much wider group of stakeholders, including users and developers (not just employees). The ability to hold tokens and participate in governance voting can enable the creation of vibrant communities, encouraging stakeholders to invest time and resources to improve and promote the long-term interests of the platform. For example, this can be one way to address the usual chicken-and-egg problem of starting a marketplace, as we discussed here.

There are at least a couple of reasons why tokens can be issued to a wider group of stakeholders: 

• The primary reason is regulatory arbitrage. There are numerous securities laws that would make it prohibitive to issue small amounts of equity for users and developers. Because tokens are often framed as utilities and not securities, they circumvent these existing securities laws (at least for now). 

• A secondary reason is that blockchains and smart contracts may make raising funds via tokens more efficient. This is especially true when compared to private companies raising equity. For example, the process of closing a private equity round, where all investors settle at roughly the same time contingent on others also settling, may be more efficiently handled by a smart contract on the blockchain. And by having everything transparently on the blockchain, it will be more obvious when a new set of investors tries to introduce opportunistic terms that align with the founders to exploit existing outside investors, thereby reducing the likelihood of such behavior. In some sense, issuing tokens is like going public, and the blockchain has made that option possible even for early-stage companies without revenue. 

We now turn to the drawbacks of decentralized platforms relative to centralized platforms:

Ruling by committee

The most obvious flaw of decentralized platforms is that they rely on decision-making by (very large) committees. This creates several problems. 

First, participants may have diverging or heterogeneous objectives that are hard to reconcile in a unified strategy. This can lead to disputes that pits the different interest groups against each other, possibly resulting in lobbying and politicking, or worse, the forming of competing factions that do not work in the common interest. 

Second, participants may engage in opportunistic behavior. For instance, speculators could seek to gain control over governance and vote for actions that drive up the token price in the short term but are not in the long-term interests of stakeholders. This is much harder to do with traditional private companies that issue equity, since this would typically require obtaining control of the board and replacing the existing CEO, something that is hard to implement when the founders decide which investors can participate and whether they get board seats.  

Third, decision-making by committee is inherently slow. By contrast, a centralized platform operator making decisions unilaterally is much more likely to come up with a clear objective and execute on it faster, something that can be crucial in fast-changing competitive environments with uncertainty. 

Fourth, voting on decisions suffers the usual free-riding problem of voting. If it is costly to evaluate proposals and to vote, and no single vote is likely to determine the outcome, participants may just rely on others to do so, free riding on their efforts. This can lead to low participation in the voting process.

Naive voters

Another problem with decentralized platforms is that it requires participants to have a certain level of sophistication and knowledge in order to productively participate in governance decisions. That works well for, say, Braintrust, given their workers are tech developers who know software and are likely to have a reasonable understanding of how DAOs work. It would likely work a lot less well with, say, a decentralized version of Uber, where the average driver may not be in a good position to make good and informed decisions on governance issues. 

Legal uncertainty

The flip side of the regulatory arbitrage that can be exploited by decentralized platforms is the regulatory and legal uncertainty they face. Founders and investors may not know if they are acting within the law and whether they face potential liabilities for their actions. They also face the risk of regulatory changes that could undermine their whole structure. Many questions remain unanswered. When do tokens become securities? What taxes are payable? Can DAO’s own property or enter into standard contracts? Can they be sued?

Concluding thoughts

Based on the tradeoffs above, it is apparent that the decentralized model may be more suitable for certain types of businesses and centralized for others. At a high level, there are two main conditions that create a particularly favorable environment for decentralized platforms to succeed:

• A market where network effects push towards winner-takes-most competition among platforms and there is a severe risk that users will be exploited by the winning platform (through higher fees and/or adverse changes in governance rules). This makes decentralized platforms particularly appealing to potential users, especially if they are also good at addressing the inherent cold-start problem. 

• Platform participants have fairly homogeneous preferences and interests. This is probably true for Braintrust provided it remains mainly focused on top-tier software developers on the supply side of its marketplace. It is likely less true for Rarible given the huge variety of NFTs and diversity of buyers, sellers and developers building on it. 

Ultimately, we think the decentralized platforms we’ve discussed above (Rarible, Braintrust, Minds) are unlikely to cause their centralized counterparts to go out of business anytime soon. Yet, they can probably co-exist and attract a certain set of participants, particularly those that feel exploited by the traditional centralized platforms, or that have an intrinsic preference for decentralization. Indeed, there are many reasons why people may seek decentralization: strong aversion to any form of censorship, strong preference for transparency and transaction security. Blockchains and smart contracts enable such decentralization to an unprecedented extent, while also allowing for monetization, which was a challenge for historical decentralized platforms. 

Finally, it is worth pointing out that the centralized vs decentralized choice is not a 0-1 decision. Blockchains and smart contracts have opened up a whole continuum of degrees of decentralization that platforms can consider. Unless one is trying to attract decentralization maximalists, the optimal choice will usually be to decentralize only the design and governance decisions for which transparency and commitment matter a great deal (e.g. the level of fees charged to participants, the rules for de-platforming participants). Other decisions (e.g. competitive strategy, hiring and firing, resource allocation, etc.) are best kept centralized.

Traditional platforms like eBay are centralized because the corporation that owns the platform alone determines and enforces design and governance rules for that platform. In contrast, the notion of decentralization associated with blockchain-based platforms means that some significant design and governance rules are fixed via smart contracts that the corporation has no control over once they are put in motion, or via the voting of stakeholders (e.g. the buyers and the sellers in the case of a product marketplace).

Important design and governance decisions for eBay include:

1. which buyers and sellers are allowed to operate on the platform

2. what and how sellers are allowed to sell

3. how the sellers and their listings appear on the platform

4. how buyers and sellers are allowed to communicate

5. how transactions are conducted (e.g. determining whether a transaction is completed or not, and so when and if the buyer’s payment is released to the seller)

6. dispute resolution

7. the ratings & reviews system for sellers

8. the fees charged to buyers and sellers

Of course, there are many important decisions regarding transactions that are controlled by sellers rather than by eBay, which is why it is a platform and not a reseller of the sellers’ items. eBay lets sellers determine what inventory to put on or take off at any time, what pictures to display for their items, what prices to set, how to do the delivery, and how much customer service the seller has to offer, including whether to issue refunds for returned items.

Let’s now compare this to the situation on OpenSea.

OpenSea can control (1) and (2) to the same extent that eBay does, but up to now it has chosen a much more hands-off policy, possibly due to the early stage of its marketplace and resource constraints. For (1) and (2), OpenSea retains the right to remove sellers or listings that it finds problematic. There can be subtle copyright issues of sellers who create or list collections that appear similar to existing prominent collections, and OpenSea has had to remove some such sellers or collections. Sometimes scammers create and sell “copies” of popular NFTs by changing something small, like one character in the name, which could confuse novice buyers. OpenSea does have a blue verification checkmark, which appears next to accounts or collections that OpenSea has verified and deems safe. However, probably due to the surge in NFT projects and the lack of human resources necessary to do this at scale, right now blue checkmarks are mostly given to celebrity accounts and collections that have achieved more than 100 ETH in sales. The vast majority of OpenSea accounts remain unverified.

OpenSea has been criticized for its very loose governance policy, including not removing offensive listings and not doing enough to filter out scams. It is conceivable that as it matures, it will tighten up its policies and start screening sellers and collections more tightly, which would bring it closer to how eBay currently operates.

For (3), OpenSea does highlight a few projects more prominently on its front page, but otherwise lets users search and order results according to 10 different criteria, with the default being most recently listed. By comparison, the default ranking on eBay is “best match” and uses an algorithm that takes into account seller ratings and other factors. eBay does allow its users to also search and order results according to five other criteria. While OpenSea doesn’t collect ratings and reviews, and use that to rank projects or sellers, it is certainly something it could consider adding in the future.

On (4), OpenSea does not enable communications between buyers and sellers, but if it did, there is nothing stopping it adding rules of conduct similar to eBay’s.

On (5), (6) and (7), because transactions are executed via smart contracts on the blockchain, OpenSea does not need to approve transactions, hold the buyer’s money in escrow, offer dispute resolution services, or create a ratings and reviews mechanism to help buyers ascertain the reliability of sellers (all of which eBay does). Indeed, the whole point of blockchain-enabled contracts is that the certificates of ownership of digital assets are automatically authenticated, and transactions are automatically filled. Of course, there is still the possibility of scams discussed under (1) and (2) above, but such scams are only possible pre-transaction and rely on confusing the buyer about what it is they are purchasing. This is something OpenSea clearly warns buyers of in its terms of service.

Finally, on (8), OpenSea has full control over the fees charged to buyers and sellers, just like eBay. The main difference is that OpenSea’s fees are significantly lower: 2.5% transaction fees charged to sellers vs. eBay’s complicated array of seller fees, with the most common rate being 12.55% plus 30 cents on each sale.

To sum up, the key differences are as follows:

• eBay has more control over how transactions are being conducted and disputes resolved, simply because OpenSea can rely on the underlying blockchain infrastructure to take care of completing transactions.

• OpenSea does less in terms of verifying accounts and listings, but this is a choice. Nothing (other than lack of resources) prevents it from doing the same amount of verification of who/what is allowed to sell as eBay does. And OpenSea most likely will do more on this front in the future given pressure from users.

Thus, there is no meaningful way in which OpenSea is more decentralized than eBay. In particular, OpenSea gives none of its governance decisions to platform participants, so fundamentally it’s no different from a classic centralized platform like eBay.

This begs the question: why do some people still refer to OpenSea as a decentralized marketplace? What they may have in mind is that OpenSea allows creators who mint their NFTs on OpenSea to move them anywhere they want, i.e. they are not tied to the OpenSea platform. This is in contrast to some other NFT marketplaces like Nifty Gateway, which promote curated and exclusive NFTs that after being minted can only be traded on Nifty Gateway. This walled-garden approach is the familiar business model of many online games like Fortnite, where items (e.g. skins) cannot be traded outside of the game. But allowing trading of NFTs outside of the marketplace where they originated is a property of openness and not decentralization. OpenSea, as the name suggests, has chosen an open model.

For OpenSea to become truly decentralized, it would need to give up some control over some of the governance rules discussed above, for example by issuing tokens and allowing token holders to vote on those rules. Competing NFT marketplace Rarible has done exactly this by issuing its $RARI token and setting up the Rarible protocol and governance rules. The tradeoffs involved in going from centralized to decentralized are really interesting in our view, and will be the subject of future posts.

Recall that a product or service exhibits network effects when the value to a user increases in the number of other users that buy the same product or use the same service. How does this definition apply to blockchains?

To answer this, it is useful to first note that there are three main types of constituents for layer-1 blockchains like Ethereum: validators (miners for proof of work or stakers for proof of stake), developers and end-users. In principle, there can be same-side network effects among the members of each of these three groups, as well as cross-side network effects between any pair of groups – so six total possible sources of network effects. We consider each in turn.

Ethereum’s network effects

a) Same-side network effect between users

End-users “adopt” a blockchain by opening up a wallet denominated in that blockchain’s token (cryptocurrency). So for end-users, adopting a blockchain is synonymous to adopting that blockchain’s token. And the more people have wallets corresponding to a specific blockchain’s token, the more attractive that token becomes as a medium of exchange (e.g. buying goods/services with that token from sellers who accept it). This is of course the classic network effect enjoyed by any currency.

However, once the number of users and transactions on a blockchain becomes large enough, the same-side network effect between users can turn negative because of congestion. This happens sooner on proof-of-work blockchains like Bitcoin and Ethereum. The more users conduct transactions on a proof-of-work blockchain, the slower the entire network, because it takes longer on average for transactions to get validated.

Another issue with the same-side network effect among users is that cryptocurrency wallet providers (e.g. Coinbase) are making it increasingly easier to hold and transact many different cryptocurrencies, i.e. to “multihome” across many blockchains.

b) Same-side network effect between developers

The more developers build projects/applications on top of any given blockchain (e.g. dapps on Ethereum), the more tools, software, modules and expertise become available for building for that blockchain (e.g. on GitHub/Ethereum), and so the easier it is to develop new applications for it, which attracts more developers to do so. This is very similar to the network effects benefitting programming languages like C++ or Python.

c) Same-side network effect between validators

Validators compete to validate transactions, so there is a straightforward negative effect between them that takes the form of business competition. Everything else equal, each validator would like to face less such competition.

d) Cross-side network effect between developers and users

Higher user demand for a blockchain makes developers more willing to invest in developing apps because it signals to them a larger pool of potential users for their apps. And a wider selection of apps attracts more people to adopt that blockchain.

However, it is worth pointing out that this effect is significantly weaker than the network effect between users and developers on a centralized platform like Apple’s iOS. This is because, unlike the case of iOS with its centralized App Store, developers do not get discovered by users by virtue of adopting a popular blockchain like Ethereum.

Furthermore, just like in (a), once the number of developers on a blockchain becomes large enough and induces a sufficiently high amount of usage, the network effect turns negative due to congestion. At scale, congestion can be significant and overwhelm the positive network effect: for much of 2021 to date, as Ethereum experienced much wider adoption and usage, congestion issues seemed to dominate any positive benefits coming from this greater usage. This has driven some developers (and users) to consider alternative blockchains built on proof-of-stake (e.g. Solana). Indeed, the main advantage that blockchains based on proof-of-stake consensus have over the ones based on proof-of-work is speed (significantly less congestion), albeit arguably with some sacrifices in terms of decentralization.

And there is one more problem with the users-developers cross-side network effect for blockchains. To the extent that users tend to follow blockchain-based apps wherever they are, they will be willing to adopt the blockchain adopted by the developer. Yes, that may require users to add a new token to their wallets, but that is a small friction in most cases, especially with wallet providers making it increasingly easier to hold multiple tokens. This means there is no real coordination problem of switching blockchains if developers find an alternative blockchain is better.

e) Cross-side network effect between users and validators

All other things equal, adding more validators to any given layer-1 blockchain (miners for proof of work or stakers for proof of stake) makes it safer for users to hold its token and engage in transactions using it. The increase in safety comes from the fact that it is more difficult for any individual validator to manipulate transactions when there are more other validators. Conversely, the more users hold tokens and engage in transactions, the more validators will be attracted since there are more revenue opportunities from validating transactions.

This network effect is important in the early stages of a new blockchain, but from the perspective of users, it arguably plateaus fairly quickly, once a critical number of validators and/or corresponding hashing power has been achieved.

f) Cross-side network effect between developers and validators

Similarly to e) above, adding more validators also makes any given layer-1 blockchain safer for developers to build applications on top of it. And the more applications are developed for a blockchain, the more attractive it becomes to validators.

Also, similarly to e), this network effect probably runs out of steam quite quickly for developers, and so does not provide much long-term defensibility.

Price effects are not network effects

One possible confusion around the network effects of a blockchain is the fact that users, validators and developers holding a particular blockchain’s token stand to benefit financially when more other participants come onto the blockchain. This pushes up the token’s price given supply is limited. But an increase in the price of the token doesn’t necessarily mean the blockchain generates more utility for participants, which is what is relevant for it to have a network effect.

Undoubtedly, the price of a given blockchain’s tokens serves important roles, including acting as a signal of the blockchain’s popularity. However, creating a network effect is not one of them. Saying a blockchain has a positive network effect because its token price goes up as that blockchain attracts more users would be like saying coffee has a network effect because its price goes up as more people want to drink coffee. That is just a normal demand-supply effect, and not a network effect. This is why any claim of network effects should focus on utility, taking prices as given. Even if the reason for the price increase stems from speculative beliefs rather than an increase in real demand, this does not create a network effect. Otherwise, we would have to say that fine art, stocks and tulips all enjoy network effects, which is obviously silly. Higher expected prices increase the value of holding an asset, not the value of consuming it. The former does not offer defensibility for the underlying product, which is the whole point of network effects.

Concluding thoughts

Understanding the various network effects surrounding blockchains is not straightforward, and it is easy to conflate them with demand effects and the role of self-fulfilling beliefs in asset pricing. This is why we thought it useful to disentangle and discuss them in this post.

So how defensible are the network effects described above? Put differently, how confident can we be that Ethereum is protected by a first-mover advantage?

To us, out of the six network effects discussed above, the strongest and most defensible for Ethereum right now is the same-side network effect among developers. There are more developers working on Ethereum projects than on any other blockchain. To the extent there are more new Ethereum projects than new projects on any other blockchain, its advantage will continue to grow in this respect.

Overall, therefore, we think Ethereum has quite strong defensibility, at least for the next few years, primarily driven by the network effects arising from its developer ecosystem. The congestion issue that we discussed earlier can be managed to a certain extent by making use of popular layer-2 solutions like Polygon, which make Ethereum more scalable. And it seems we are not too far away now from the long-anticipated launch of Ethereum 2.0 which will involve a switch to proof-of-stake, thereby addressing the congestion concerns head on.

That being said, new layer-1 blockchains like Solana rely on more popular programming languages, which could partially nullify the developer expertise network effect advantage enjoyed by Ethereum. Developers are also increasingly multihoming. More fundamentally, one can imagine a future in which it becomes seamless to make transactions across different competing blockchains via tools that ensure interoperability. This would make network effects work at the level of the entire blockchain economy rather than at the level of any individual blockchain, meaning no particular blockchain would enjoy defensibility, Ethereum included.

The chicken-and-egg problem is often the toughest hurdle for founders of marketplaces and other network effects businesses to overcome. Indeed, most marketplaces never get past this stage. But the problem is not insurmountable: all successful marketplaces (indeed, network effects businesses more generally) have found ways to solve it, each using their own specific set of hacks to fit their circumstances (e.g. Airbnb, Alibaba, eBay, Facebook, PayPal, etc.).

By now, a number of investors and academics have written about a wide range of tactics that founders can consider using to solve the chicken-and-egg problem. Rather than simply repeat them here, in this post we group them in four broad categories, so that one can more easily grasp the fundamental principles underlying the various strategies. This may also be helpful in uncovering new variations of these broad strategies as applied to different circumstances.

1. Single-side mode

The idea here is to attract an initial base of users on one side, by making them willing to join even if they expect few (or even no) users on the other side. In some cases, it may be enough to make it extremely easy for one of the two sides to sign up (e.g. creating a profile for them) combined with only charging fees when they make successful transactions, which means there is no downside to signing up. But in many other cases, this is not enough, either because signing up requires more of an investment, or because signing up is insufficient to create any value for the other side (users need to be active, not just have accounts). In such cases, users may need to be incentivized to actively participate. This requires offering something that has sufficient standalone value for them to do so; i.e. that does not depend on there being any users on the other side.

There are many ways to do so.

• Pay users to join: PayPal initially paid users $20 to open an account and $20 to refer a friend (later dropped to $10, then $5). This also gave them some initial funds to spend in PayPal.

• Employ/hire the supply side: Uber started off by hiring drivers by the hour to drive in specific areas of San Francisco at specific times in case anyone called for a ride. After generating enough demand, it was able to convince drivers to move to the current revenue-sharing model, without any fixed guaranteed payment.

• Offer free services: Zillow’s free Zestimate allows homeowners to get an estimate of their house’s value – the estimates are more accurate if users update information about it. Similarly, Credit Karma offers users the ability to check their credit score for free. And Hipcamp started off by offering information about public campsite availability to attract the demand side, before opening up for private hosts to list their properties.

• Offer one side a compelling standalone product: OpenTable started off by selling restaurants an electronic reservation system to manage bookings (regardless of where they came from). After signing up many restaurants, it was able to launch the OpenTable website to consumers, who could use it to book tables at all restaurants that had installed OpenTable’s reservation systems. Similarly, Amazon started by selling books (and later other products) as a retailer to attract consumers, before opening up to third-party sellers.

This last strategy, of offering a standalone product, is often referred to as single-player mode. The terminology comes from video games. By making a game attractive for users to play on their own, the game developer can attract a large enough installed base, at which point it can turn on multiplayer mode, enabling player interactions.

It is worth noting that selling a product to one side first is not always part of a deliberate strategy to seed one side of a marketplace. For instance, it is not clear whether Jeff Bezos had the two-sided marketplace vision in mind when he started selling books online. And certainly Steve Jobs had no plans for an App Store when Apple launched the iPhone with only its own apps. In such cases, firms later realize the opportunity to turn their popular products into platforms, something we’ve written about at length starting with this post.

2. BYO demand (or supply)

The bring-your-own-demand approach relies on incentivizing suppliers (or creators) to bring their own existing customers (or fans) to the marketplace. The appeal for suppliers is having a better platform for interacting with their existing customers (e.g. managing schedules, payments, records). Examples of bring-your-own-demand include Clubhouse, Patreon, and Playbook.

This strategy is similar to the single-side mode approach described above. The difference is that here suppliers invite (or onboard) their existing customers onto the marketplace. In other words, supply brings its own demand, rather than the marketplace going after demand by itself (e.g. OpenTable going after consumers after having attracted restaurants with its initial reservation management system).

Of course, a very similar logic applies to bring-your-own-supply, i.e. incentivizing buyers to bring their own existing suppliers to the marketplace. Examples of this approach include Marketboomer and Fairmarkit. More generally, it can be applied by any sourcing/procurement marketplace.

The bring-your-own demand/supply strategy does however have an important limitation. If customers/suppliers mainly stick with the suppliers/buyers that onboarded them (i.e. discovery of new suppliers/buyers is limited), the marketplace is not very defensible. Just like the marketplace convinced suppliers/buyers to onboard their customers/suppliers, a copycat with a slightly better software platform can convince suppliers/buyers to switch and take their customers/suppliers with them. We have discussed this issue in this post. Therefore, the key to building a defensible moat is to quickly move to enable each supplier’s/buyer’s customers/suppliers to discover other suppliers/buyers, and also attract new customers/suppliers.

3. Fake one side

Whenever we see an early-stage marketplace that hides its supply side, we immediately assume it’s because it lacks a critical mass of suppliers. Such marketplaces might ask buyers to submit RFPs, which they promise to send to their “large base of suppliers”. In practice this means, after receiving a request, the marketplace tries to manually find suppliers who can fulfill it. Ideally, these suppliers are eager for extra jobs, and are onboarded onto the marketplace in the process. Sometimes however, the marketplace founders might need to act as suppliers themselves at the very beginning to get things going: the founder of Cambly reveals here that the initial English tutors on Cambly were himself and his co-founder.

Another version of this approach is to post suppliers’ offerings without their approval, provided the marketplace is confident it can manually fulfill orders from them if required. DoorDash famously did this in its early days around the Stanford campus, with the founders posting menus of nearby restaurants, taking the orders, picking up the food and delivering it to customers themselves. Zappos used a similar strategy for online shoe orders.

It is important to note that faking one side does not necessarily involve outright lying or mis-representing availability. For instance, in Taobao’s early days, Alibaba employees were listing and buying listed items from each other to create a sense of activity. Founder Jack Ma even listed his watch. Airbnb enabled its hosts to automatically post their listings on Craigslist, to boost their demand, thereby piggybacking on the liquidity of Craigslist. More generally, a marketplace can just guarantee to fulfill demand, and leave the logistics of how they actually do so unclear at first.

However, some marketplaces have engaged in more suspect practices. Our own experience using OfferUp in its early days was that most of the so-called “buyers” seemed to be bots sending requests for more information (and then going silent), presumably to create the impression of lots of buyer activity. Apparently, we were not alone in drawing that conclusion. According to James Currier of NFX,  “Reddit used fake persona accounts to plant interesting questions. And PayPal built a bot that posed as a human, purchased things on eBay, and insisted on paying with PayPal.”

4. Change expectations

The most fundamental reason why the three broad strategies described above work is that they break the spell of “unfavorable expectations” that plague early-stage marketplaces: buyers don’t join because they don’t expect sellers to join, and vice versa. The three strategies above are ways to make it worthwhile for one side to join regardless of its expectations for the other side: because they get some standalone value, or because sellers can bring their own buyers, or because buyers are “fooled” into thinking there are some sellers present.

Rather than making users willing to join regardless of expectations, another approach is to try to change users’ expectations in the first place, from unfavorable to favorable. If users have favorable expectations and there is some advantage for users to come on board early (e.g. gaining a head start in building up their reputation on the marketplace), they will have an incentive to join even if others haven’t.

One way to do so is attract a “whale” on one side, i.e. a large or particularly sought-after supplier (or superstar). Once the whale comes on board, everyone expects the marketplace is going to work, and so users are willing to make the necessary effort to participate. This is inspired by the standard shopping mall tactic of first signing up one or two key anchor stores, which ensures there will be enough interest from shoppers in the future, which in turn attracts other retailers, who otherwise might have been hesitant to commit to the remaining slots. This can be a particularly valuable tactic for B2B marketplaces: getting commitments from a handful of major buyers or sellers may be enough to get everyone else to come on board.

A modern-day, blockchain-era version of this strategy is to offer early users/adopters cryptocurrency tokens together with a compelling white paper. The tokens may be worth little initially, but if there is enough belief in the project, people will participate on the expectation their tokens will be worth a lot more in the future, when many users participate. These beliefs can be self-fulfilling if more demand for tokens drives up the token’s price, and early users recruit others to the blockchain to further increase demand for tokens. Note the token approach has a built-in incentive for users to join earlier rather than later – more upside from token value appreciation and possibly also more generous token distribution by the platform.

Other strategies, hacks and considerations

One way to simplify (though not enough to fully solve) a two-sided chicken-and-egg problem is to reduce it to a one-sided chicken-and-egg problem, by focusing on users that can act as both buyers and sellers. Examples of platforms that have used this approach in their early days include Airbnb, Craigslist, eBay, Etsy and Taobao.

There are many other hacks that are insufficient on their own but can be used to complement the strategies described above, although these apply to almost any startup trying to get off the ground. Examples include: creating the illusion of exclusivity and the fear of missing out by initially making membership invite-only (e.g. Clubhouse, Gmail, Facebook), making use of free press (e.g. Swimply, the Airbnb for swimming pools, S’more, the love-is-blind of dating apps), holding events to attract and on-board potential users (e.g. PartySlate did this with its own launch party), referral programs (e.g. Airbnb, PayPal), etc.

For a marketplace, we’ve talked about how to attract one side without the other. But how to decide which side to focus on first? The usual advice is, all else equal, to focus on building the supply side first. In most cases, suppliers are fewer in number than buyers, and they are more willing to try things out in the hope of obtaining new customers. Having a sufficient variety of supply also tends to be more important in making a marketplace viable. But there are exceptions to focusing on supply first, and sometimes it may be easier to offer something of standalone value to buyers first (e.g. Amazon). Also, buyers might not care about the number of suppliers as much as vice-versa, so that attracting buyers first would then make more sense (e.g. the Credit Karma example mentioned above). Of course, having said this, in practice startups need to use somewhat of a “zig-zag” strategy: sign up enough users on one side to start offering some value to the other side, then attract some users on the other side, which then makes it easier to get some more users on the first side, and so on.  

Finally, another general piece of advice is to focus on a narrow segment/vertical first, and then expand out. eBay got a lot of its early traction with Beanie Babies; today, it has around 1.6 billion live listings at any given time, covering almost everything imaginable. Rover started with dog sitting in Seattle, and later expanded to pet sitting and dog walking, both across North America and parts of Europe. The idea is to get “liquidity” within a narrow vertical by getting a set of high-demand users on both sides before expanding out from there. Narrowing the segment also includes narrowing the relevant time period (so as to get more liquidity), a strategy stock markets and auction houses have long used. There are many benefits of narrowing down the initial target segment: (i) figuring out whether the unit-economics work before spending significant resources to expand to other locations/use-cases; (ii) creating a reliable playbook that can then be rolled out to other locations/use-cases; (iii) leveraging the sellers and/or buyers obtained in the first location/use-case to attract users in/for other locations/use-cases (when possible).

Concluding remarks

We have discussed four high-level strategies for solving the chicken-and-egg problem, each of which covers a wide variety of possible tactics and can be supplemented by many other hacks and shortcuts. It is important to realize that there are no wave-your-magic-wand solutions here: overcoming the chicken-and-egg problem for network effects businesses usually involves lots of hustle, out-of-the-box thinking and trial-and-error.

The benefits of classifying network effects

Before doing so, however, it is useful to start by reflecting on why classifying network effects might be useful, and briefly commenting on the NFX list.

We see two main benefits from doing the classification exercise. First, it helps reveal that network effects are more ubiquitous than one may realize. Some of the network effects listed by NFX (e.g. bandwagon, expertise, and tribal network effects) will not be obvious to most people. And by being aware of these, founders will see more opportunities to leverage network effects to build defensibility into their products. Second, network effects can work in different ways across the various categories, and understanding this can help uncover important differences in terms of the strength and defensibility afforded by the various types of network effects.

With this in mind, and in order to set the stage for network effect #16, here we briefly review and comment on the last three types of network effects identified by NFX, two of which were added after the initial list with 13 types was published:

• Bandwagon network effects (nfx #13): people are willing to pay more for some products that others buy because they want to be part of a new trend, or because they fear being left out. Tesla certainly benefits from this effect, as did Apple’s iPhone in its earlier days, as illustrated in this well-known satire and this one. Compared to most network effects, this one can be quite fickle though: what is cool today may no longer be cool tomorrow, and when too many people join, the coolness factor evaporates or can even work against the initial product – early-adopters may want to leave and chase the next novelty. Put differently, bandwagon network effects are unlikely to work at scale. Oftentimes, they look more similar to virality: creating a “trendiness” buzz around a product that makes everyone talk about it may help generate awareness, but does not constitute a sustainable source of defensibility. Especially when the features that made the initial product cool are copied by competitors with strong enough brands of their own (e.g. Blackberry).

• Expertise network effects (nfx #14): people get more value from a product that is more widely used, because this attracts more expertise to develop around the product, which benefits users who want to tap into that expertise. An example is Salesforce’s CRM software, where part of the value to a firm adopting it is knowing that its current and future employees are more likely to know how to use it, and that there will be many experts that can help (either via online communities or via direct hires). This is why once a product becomes an industry standard, it becomes difficult to unseat. As NFX points out, in the face of an incumbent with expertise network effects, competitors will need to try to minimize switching costs, so making their product as familiar as possible (e.g. Google Sheets’ approach to take on Excel), whereas an incumbent can focus on adding more features and making their product even more powerful.

• Tribal network effects (nfx #15): belonging to a tribe (e.g. the Harvard alumni club, the YC alumni network, or a BMW user group) is more valuable when there are more people in it. This is of course a type of social network, except the network effects between members of the same tribe tend to be stronger than with regular social networks. Indeed, members of the same alumni club are much more likely to do favors for each other than say random Facebook or Twitter users. While it is not easy for a firm to develop tribal instincts among its customers, some brands do seem to have produced a cult-like following (e.g. Comic-Con and Harley Davidson). An interesting implication of tribal network effects is that the existence of a competitor (a rival tribe) can actually strengthen them.

NFX’s list of 15 different types of network effects is not supposed to be definitive and there is plenty of overlap between many of the different types. In that spirit, we now propose network effect #16.

Internal network effects

One typically thinks of network effects as operating between clearly distinct entities that make independent adoption decisions – different consumers, different app developers, different firms (as either buyers or sellers on B2B marketplaces like Alibaba or Faire or AWS marketplace).

However, there are instances in which the network effect is between different employees within the same firm: each employee can individually decide whether or not to use a product or service, but as more employees use it, it becomes more valuable for everyone within the enterprise. This is what we call an “internal network effect”. And internal network effects have become increasingly important as a source of defensibility for software providers in the last decade or so due to the adoption of the bottom-up go-to-market approach to selling into enterprises, in which software is targeted at individual employees rather than at the company HQ. This approach may take longer (though selling to enterprises can also have very long sales cycles), but when it succeeds (i.e. when a large share of employees adopt a product), it makes it harder for an enterprise to switch to a competing product.

Slack, in its original form, is a classic example of internal network effects: the more people within an enterprise used Slack, the more valuable it became for everyone to join. Over time, people used Slack channels to communicate in all kinds of communities, and so what was originally an internal network effect grew to become external.

Many business productivity software products enjoy internal network effects, since they are designed for employees to use collectively, even when they are not focused on communications, like Slack. Examples include Asana (allows teams to manage projects and tasks), Monday.com (allows teams to plan, organize and track work), and many others. In these examples, the network effects are mostly internal; the only way they can include an external component is when employees change firms and carry their software preferences with them (we discuss this below).

Internal network effects are generally less strong and less defensible when compared to regular (external) network effects for several reasons:

• They are usually limited to each enterprise customer, so their scale is relatively small, especially if the target market is small and medium-sized enterprises.

• As in the case of regular network effects that are local (e.g. Uber, DoorDash), one has to start from zero and repeat the same playbook for each new “locality”; here each locality corresponds to each new enterprise customer, and requires “seeding” them and growing internal adoption each time.

• It is easier for users within the same organization to coordinate and switch to a rival product if it is superior, relative to the case when these individuals are not connected in any way. Indeed, a company’s HQ can mandate such a change. This means products built on internal network effects are less defensible than those built on regular (external) network effects, all other things equal.

Having said this, some enterprises have many thousands of employees, and when they have all learnt to use a particular software product, the switching costs of getting them all to adopt a new software can be high enough to create a powerful lock-in effect. While the company HQ could in principle still dictate matters, in practice the increased autonomy of individual employees in their software choices means this possibility is a lot less likely today. Winner-take-all at the level of the entire market is still less likely than with internal network effects, but winner-take-all is more likely within any given enterprise.

Moreover, as employees shift across employers, the advantage of a common software that is used by many large enterprises (in terms of reduced learning costs) comes into play. Each new enterprise may be more likely to adopt the software used by other enterprises because it becomes easier to attract new employees (an expertise network effect). This is particularly important when the software becomes a key part of the employees’ tasks in the enterprise and involves a significant learning curve (e.g. the choice of a programming language or technical software). And, as happened with Slack, once a product becomes industry standard, it may be possible to extend its use cases beyond each individual enterprise boundaries.

Concluding thoughts

In some sense, there is almost no limit to the number of ways one could characterize network effects. The real question is whether the characterizations lead to useful insights about the strength and defensibility of the relevant network effect, and more generally about firm strategy. Above we have argued that our proposal for nfx #16 meets that standard. What do you think? What would be your choice for nfx #16 or beyond?

If you enjoyed reading this, sign up now so you will receive our posts directly in your inbox as they are published.

Subscribe now

In the meantime, tell your friends!

The course

Put simply, if you enjoy reading our Substack articles, you will love our course.

Participants in the course will learn how to rigorously evaluate the prospects and defensibility of businesses with network effects. To achieve this, we bring to bear over two decades of academic research, as well as practical experience from having consulted with a dozen large corporations and angel invested in over 200 startups. The course will cover the perspective of investors considering businesses with network effects (e.g. marketplaces or products featuring data network effects), of founders building such businesses, and of executives working at established companies and considering product-to-platform transformations.

We will go in much greater depth and breadth than our articles. Participants in the course will have the opportunity to interact with a live cohort of investors, startup founders and senior executives, as we apply our frameworks and concepts to a wide range of examples, including their own businesses or investment targets. Guest speakers will include a venture capitalist who is focused on investing in network effects businesses, a founder who is building such a business, and an executive who has implemented a successful product-to-platform transformation.

All of this will take place in 5 sessions (October 11, 13, 18, 20 and 25), each running for two hours, 4-6pm PST. We have chosen times that should work well for both North America and Asia-Pacific, reflecting our own time zones (Andrei is based in Boston, Julian in Singapore). If you subscribe to our Substack newsletter, you can use the promo-code BINFXPC to obtain a 20% discount. You can see the full details of the course and register here.

The new platforms for executive education

If you click on the link above to our course’s webpage, you will notice that it is hosted by the startup ScholarSite. So why have we decided to run it via their platform? To understand, it is useful to take a step back and explain how courses like ours have been run until today.

Traditionally, such courses are taught as “executive education” via universities. The universities take control over pricing, marketing, and other logistics, and in return they keep the lion’s share of revenues and typically pay the professors who provide the content and teach the course a modest fixed amount for their time. A key selling point for such university-run executive education courses is that participants obtain a certificate with the university’s brand on it. Some participants find it valuable to feature these certificates on their LinkedIn profiles, or even claim an affiliation with the relevant university by virtue of having done the course.

Over the past decade, a few standalone platforms have emerged that help universities run their executive education courses online. Well-known examples include Emeritus and GreatLearning, as well as the executive program arms of Coursera and EdX. These platforms specialize in re-designing and marketing the courses for an online audience, typically sharing revenues with the universities.

We actually first tried to offer our course via our respective universities and one of these established platforms. In one case we were told we would need to record scripted videos, which in our view guarantees a mediocre learning experience. Furthermore, we also found out that doing such a course through our universities and one of these platforms would involve an administrative labyrinth that would take at least a year to traverse and would eventually leave us with a minuscule share of revenues.

These tight controls and rigid procedures are in contrast to how things work for the course we will offer via ScholarSite. The company’s mission is to empower academics to build their own independent brands. It disrupts the traditional executive education model by giving academics control over what, when and how to teach, how much to charge, as well as how to market themselves (although ScholarSite does help with marketing), and which participants to accept in the course. In other words, ScholarSite uses a true platform business model, in which most of the control is given to the content providers. Meanwhile, the company focuses on ensuring the quality of the scholars allowed to offer courses through its platform. And instead of paying a relatively modest hourly amount, ScholarSite lets instructors keep the bulk of revenues, just like Airbnb, Patreon, Uber, Upwork, etc.

A similar model has been adopted by Maven, another startup that enables creators to launch their own live cohort classes. Currently, there are two main differences between ScholarSite and Maven. First, ScholarSite is focused on leading academics that can bring their research to a non-academic audience, while Maven targets a wider range of content creators, many of whom already have large followings. And second, ScholarSite creates websites for the courses offered by their instructors and helps with marketing, while Maven is closer to the Substack model, which puts most of the responsibility for attracting an audience on the creators.

Over time, the second distinction might start to fade away. As Maven and ScholarSite attract more instructors (some of whom bring their own followers), they will attract more participants who will start to come to their sites to discover courses. And with more people looking on their sites to discover courses, they will attract more instructors. In short, they will both benefit from network effects. In this regard, what Maven and ScholarSite are building is similar to Udemy and Teachable, which also enable the discoverability of online courses. The key difference there is that Maven and ScholarSite are designed for cohort-based teaching just like traditional executive education, which is likely a much more effective way to learn in most cases.

Concluding thoughts

The emergence of new platforms that enable scholars to create courses online leveraging their research expertise is great news for instructors like us. They raise the alluring possibility of a “1000 true fans” future, in which academics and other creators working in very niche areas will be able to monetize their unique skills and knowledge, while at the same time bringing together participants from all over the world who can also learn from each other.

We are excited to be part of this transformation with our own course and we hope some of you will join us on this journey.

At a high level, there are two ways for merchants to build competitive advantage and withstand the threat of commoditization. They can go broad by employing a roll-up strategy a la Thrasio– we discussed this strategy here. Or they can go deep, by offering a highly specialized product or service and leveraging the platform’s economies of scale to grow.

The latter option is the topic of this post, and it is relevant to the vast majority of platform participants that cannot afford the more expensive go-broad strategy – and don’t want to sell out to roll-up firms like Thrasio.

Going deep

Specializing in a product or service (in other words, having a clear and narrow focus) has always been a good strategy for achieving competitive advantage. Large digital platforms have taken the logic of specialization and have turbo-charged it when applied to the merchants that operate on them (sellers, app developers, content creators, etc.). Indeed, digital platforms, via their algorithms and rating systems, usually make it much easy for consumers to find their ideal product. And by breaking down geographic barriers and accumulating very large user bases, they greatly increase a seller’s reach. Taken together, those qualities mean that the returns to becoming the highest-quality or the lowest-cost provider in a narrowly defined product or service category are significantly higher for sellers who conduct business on platforms than for those outside platforms.

In other words, platforms reward hyper-specialization through a self-reinforcing cycle. The more a product aligns with what buyers (customers, viewers) are searching for, the higher its ratings will be, increasing the chances that the platform’s algorithms will drive target customers to it. That means more customers will buy and rate the product, further heightening its advantage. Sure, to some extent, this cycle also operates outside platforms (through search results and possibly ratings on various sites), but centralized platforms make it a lot more powerful. This explains why large platforms have enabled extremely narrow product/service/content niches to emerge and thrive, that previously might not have been viable.

Examples

Take Anker, a Chinese company specializing in computer and mobile-phone peripherals such as chargers and power banks. With a market value of about 7 billion (this week), it is one of the most successful third-party sellers on Amazon.com. “Amazon reviews are the single most important input to our new product development process,” founder Steven Yang has said. Initially focused on Amazon, Anker now sells through many channels, including offline ones such as Best Buy, Target, and Walmart.

Examples in other contexts include unique and hyperspecialized content providers on YouTube (such as 5-Minute Crafts, Dude Perfect, and MrBeast), Facebook (Bored Panda), and Instagram (9GAG). The more views and likes these channels obtain, the higher their revenues, which they can invest in producing more and better content, leading to ever-larger audiences.

Contrast these content providers with traditional media companies (e.g. the BBC or the New York Times), which have tended to be generalists, publishing many different types of content (daily news, opinion pieces, in-depth investigative reports, sports, finance etc.). As these companies are increasingly forced to play on social media MSPs (Facebook, Twitter, YouTube), their traditional strategy does not appear to be sustainable in the face of new types of publishers that specialize in one type of content only. Consequently, traditional media companies should arguably narrow their focus to the types of content they have a comparative advantage in. For instance, the New York Times should probably double down on in-depth investigative reports and opinion pieces, rather than continuing to spend resources on covering sports or finance. Likewise, local newspapers may do best focusing on hyper-local issues and events, rather than national or international issues, for which there are many other, better sources, easily accessible through Google or social media.

Concluding thoughts

Hyper-specialization is part of a more general trend that has arisen due to the power and reach of the Internet. It is the reason behind Kevin Kelly’s 1000 true fans concept, that every kind of quirky creator becomes viable because all they need is 1000 true fans somewhere in the world. Thanks to the Internet, it is possible for these fans, wherever they are, to discover the creator and pay for their creations. In Kevin Kelly’s words:

“If you lived in any of the 2 million small towns on Earth you might be the only one in your town to crave death metal music, or get turned on by whispering, or want a left-handed fishing reel. Before the web you’d never be able to satisfy that desire. You’d be alone in your fascination. But now satisfaction is only one click away. Whatever your interests as a creator are, your 1,000 true fans are one click from you. As far as I can tell there is nothing — no product, no idea, no desire — without a fan base on the internet. Every thing made, or thought of, can interest at least one person in a million — it’s a low bar. Yet if even only one out of million people were interested, that’s potentially 7,000 people on the planet. That means that any 1-in-a-million appeal can find 1,000 true fans. The trick is to practically find those fans, or more accurately, to have them find you.”

The point we are adding in this post is that platforms that aggregate users and help with discovery (Amazon, eBay, Esty, Facebook, Instagram, Medium, Only Fans, Patreon, Tiktok, Twitch, YouTube) greatly amplify this process, and as a result increase the rewards to hyper-specialization for their participants.

The risk of commoditization

Large marketplaces, such as the ones run by Alibaba, Amazon and Apple, have made it much easier for merchants to be discovered and reach new customers. They bring in customers who are specifically on the site looking for suitable products or apps. Indeed, the majority of shoppers in the U.S. start their product search on Amazon (close to 75% according to this recent report).  

But as merchants on Alibaba and Amazon, app developers on Apple’s iOS and Google’s Android, restaurants on DoorDash, Grubhub, Uber Eats and Seamless, hotels on Booking.com and Expedia, small businesses on Tencent’s WeChat and others have discovered, participation on large marketplaces can be a double-edged sword. Along with the increased chance of being discovered comes the increased risk of being held-up and/or commoditized. This risk can take several forms:

• Marketplaces can extract higher fees over time. Take Property Guru for instance, the leading marketplace for real-estate properties in South East Asia and the dominant player in Singapore, where it is almost exclusively used for property searches. In 2019, it increased the listing fees charged to agents in Singapore by 10 to 25%, to the dismay of those agents.

• Marketplaces can make advertising (paying for prominence on the marketplace) more and more important for merchants to maintain visibility (as Amazon has done over time).

• The environment on large marketplaces may become increasingly competitive. This can be the result of both natural growth (more users and more merchants are attracted to participate), and of marketplace design changes (e.g. the recommendation algorithm places a higher weight on offerings with lower prices or those that pay to be promoted).

• Some marketplace owners may vertically integrate into some of their merchants’ businesses after observing which products and designs sell better. This practice has been employed by Amazon, Apple and Microsoft, among others. We wrote about the merits of allowing platforms to do this in an earlier post. Regardless of those merits, from a merchants’ perspective, this is obviously bad news, especially if the marketplace is using their private data to inform its decision as to whether to compete against them.  

• Marketplaces tend to weaken merchants’ relationship with buyers (their customers). This occurs naturally because buyers come to (and typically register with) the marketplace first. Sometimes marketplaces may go further, only sharing buyer data partially (if at all) with merchants, as seems to be the case on food delivery marketplaces.

To some extent, marketplaces may purposefully try to commoditize merchants in order to get them to compete harder on price (and sometimes quality), which in turn makes the marketplace as a whole more appealing to buyers/customers, therefore attracting ever more of them. It is, however, important to realize that marketplaces also create self-reinforcing cycles that go in the opposite direction, i.e. help the (few) merchants who know how to navigate the competitive waters of large marketplaces. Namely, the more a merchant’s product aligns with what consumers are searching for, the higher its ratings will be, increasing the chances that the platform’s algorithms will drive target customers to it. That means more people will buy and rate the product, further increasing its competitive advantage.

That being said, most merchants face an increasingly competitive environment on marketplaces, which raises the question of what such merchants can do to escape the commoditization trap. Below we propose a few strategies to consider.

Develop direct channel making full use of new online tools

Even if it is impossible to avoid conducting business on key marketplaces, merchants should always seek to build direct relationships with their customers via their own channels (e.g. a branded website and/or app). These direct channels can help limit the merchants’ dependence on marketplaces and protect them against the changing whims of the marketplace owners. The good news is that given the widespread availability of business-in-a-box solutions such as Shopify, BigCommerce, Magento, WooCommerce, Mailchimp, Square, Appy Pie, and Wix, creating a fully functional online storefront is increasingly easy and affordable.

The key difference between relying on such providers of software tools and relying on a marketplace is that the former do not control the merchants’ relationships with their customers. For example, Shopify provides all the digital tools and infrastructure a brand needs to sell online, typically without consumers realizing that the brand’s store is powered by Shopify. Part of the reason Shopify is so appealing to online merchants (it has more than one million as customers) is that unlike Amazon, Shopify is not a marketplace connecting them with consumers and therefore does not commoditize them. In the words of Tobias Lütke, Shopify’s founder and CEO, “Amazon is trying to build an empire. Shopify is trying to arm the rebels.”

Similar business-in-a-box solutions are available in other verticals. Consider the restaurant sector. DoorDash, Grubhub, and Uber Eats enable consumers to place takeout orders and arrange for delivery in the U.S. Because restaurants have become increasingly reliant on these marketplaces in the past five years, they now charge a 15% to 30% commission. In contrast, ChowNow and Olo, two fast-growing start-ups that provide back-end technology to enable restaurants to sell directly online, allow the restaurants to keep full control of their customer relationships. Restaurants choose delivery and sales channels, and pay these providers subscription fees for powering their websites and mobile apps, handling orders and payments, running the restaurants’ loyalty programs, and helping with marketing services.  

In recent years the “Shopify for X” approach has been applied in more and more sectors. Dumpling lets people start their own personal shopping businesses so that they can reduce or eliminate their dependence on platforms like Instacart. NearSt lets brick-and-mortar retailers in the United Kingdom make their inventories searchable on Google without listing them on Amazon. And by now there are many startups that enable online merchants to fulfill orders (Shipbob, MasonHub, Airhouse, Fabric and so on) so they needn’t be reliant on Amazon to offer world-class fulfillment.

The downside to such business-in-a-box solutions, of course, is that merchants must figure out how to get customers to their sites. The solutions providers can help to some extent through partnerships. For instance, Shopify entered into a partnership with Facebook in May of 2020 to allow its merchants to create storefronts on Facebook and Instagram. It partnered with Walmart that June to allow its merchants to easily become third-party sellers on Walmart’s e-commerce marketplace.

Multihome

Whenever possible, merchants should multihome, i.e. list on multiple competing marketplaces. This can allow merchants to become less reliant on any one of them. Merchants can try to steer their customers towards their preferred platform, like they do when they put up signs “We prefer MasterCard” or “We prefer Visa”. Alternatively, if one marketplace tries to push unfavorable terms, merchants can delist from it and thereby steer customers to transact through other rival marketplaces. The threat to steer business (or delist altogether) can keep marketplaces in check, provided the merchant is strategically important, or can coordinate with many other merchants, or can do something newsworthy to highlight their plight.

Use marketplaces as showrooms

Few online merchants can completely wean themselves off the dependence on large marketplaces given the huge numbers of consumers these marketplaces attract. But by taking advantage of the business-in-a-box solutions discussed above, merchants can use marketplaces mainly as funnels for obtaining new customers, while serving repeat customers directly. This approach gives merchants more control over their customer relationships, including customer data, enabling them to better tailor their offerings and differentiate themselves. Essentially, it lets them use the marketplaces as showrooms.

One tactic for doing so is to offer deals and directions to the seller’s own channel when filling orders through a marketplace. For example, restaurants can drop coupons into the bags picked up by food-delivery platforms, steering customers to their websites and offering discounts on the next direct order.

Merchants should also limit their offerings on marketplaces and offer a broader variety of products, services, and loyalty rewards in their direct channels. Some do so with Amazon: They use Amazon’s marketplace to obtain their first order with a customer, and when fulfilling it they include a coupon aimed at attracting the consumer to their own channel for repeat orders, sales of other products, and subscriptions.

These principles have become all too relevant to content creators on OnlyFans since last week, when the platform announced it would ban sexually explicit content starting October. This policy change threatens to wipe out the revenues of many creators who are largely responsible for the platform’s quick growth over the past few years. These creators now probably wish they had started investing in their own channel and “showrooming” OnlyFans a lot earlier.

Of course, the most obvious way to get consumers to buy via the direct channel is to charge lower prices there. Some marketplace contracts forbid that. In the U.S., Amazon has had a policy of prohibiting merchants from undercutting the prices offered on Amazon’s marketplace when selling through any other channel including directly via their own websites. While that policy was removed from its Business Solutions Agreement in March 2019, according to one lawsuit, a similar policy resurfaced shortly afterwards under Amazon’s Fair Pricing Policy. To the extent the policy applies, the practice of offering discounts via coupons for repeat orders can help avoid it.

Booking.com and Expedia have the same type of contracts with hotels, preventing them from undercutting the prices they set on these platforms when selling via other channels (including directly on their own websites). Many hotel chains have responded by offering customers loyalty rewards and additional perks, such as the ability to choose specific rooms and to access upgrades, when they book through the hotel’s own channel. Moreover, by enrolling repeat guests in a loyalty program, hotels can offer them lower prices since the rates do not need to appear on the hotel’s public website.

Conclusion

While many merchants have become increasingly reliant on large marketplaces to attract consumers, merchants needn’t become completely beholden to them. There is an increasing set of backend tools that merchants can use to become less dependent on marketplaces. By employing these tools and the strategies of multihoming and showrooming, merchants can gain more leverage over existing marketplaces to develop their own brands and take back the relationship with their customers. Another option that is becoming increasingly common is to sell to marketplaces aggregators like Thrasio (discussed in our previous post) and let them develop the brands over the long haul.

Some examples

One of the first and most successful adopters of this strategy is Thrasio. Founded in July 2018, it achieved unicorn status in just two years—a record. It has raised vast amounts of venture capital funds and acquired over 125 third-party merchants on Amazon to date. After acquisition, Thrasio aims to grow the sales of each merchant by leveraging its expertise in operations, marketing, and pricing in several ways:

• Further improving the merchants’ products, often by making better use of customer feedback to optimize the product offered.

• Improving the merchants’ product listing page (e.g. photos and text).

• Knowing how and when to make use of Amazon advertising to boost product sales.

• Optimizing the pricing algorithm used.

• Keeping up-to-date with Amazon’s rules and regulations, handling bad (fake) reviews and managing Amazon’s response to them.

• Improving logistics by negotiating better rates with third-party shippers and by sharing fulfillment best practices for handling customers directly.

• Cross marketing the more than 14,000 products it offers (from pet-odor eliminators to socks and kitchenware). For instance, this can be done via promotional offers included with shipped items.

Thrasio pays merchants a combination of upfront payment as well as additional earnout payments within 12-24 months of the sale based on the performance of the business. The role of the earnout payments is presumably to make selling to Thrasio more attractive for higher-growth businesses.

Many other firms have adopted the rollup strategy on Amazon (at least 69 to date). Some of the best-known ones are Perch, Boosted Commerce and Heyday, all of which have raised significant venture funding. Amazon aggregators are not constrained to North America: factory14 is applying this in Europe, Rainforest is doing so in Asia, and similarly there are others in Latin America and India. And some have also started applying the rollup strategy in other ecosystems: for instance, OpenStore, Hedgehog and Pattern are doing it with Shopify merchants.

A version of the roll-up strategy is also employed by Wave.tv, a rapidly growing start-up that publishes offbeat sports content on social media platforms, including Instagram, Facebook, Snapchat, TikTok, and YouTube. Wave.tv acquires or licenses content from many sources and leverages its technology and data analytics to repackage it under more than 15 brands, including BenchMob, Haymakers, and Buckets, increasing the content’s reach on the many platforms on which it operates.

The arbitrage opportunity exploited by aggregators of platform merchants

Fundamentally, merchant aggregators are exploiting an arbitrage opportunity. And there are three underlying factors that create this opportunity.

First, the aggregators add value via economies of scope (i.e. via synergies) and scale (i.e. via volume), as illustrated by the Thrasio example described above. At the most fundamental level, the opportunity for such value creation comes from the explosion of digital platforms and tools now available for online merchants and digital creators to reach their customers. These platforms and tools are easy to use but hard to master, so many merchants and content creators operate below their potential scale: aggregating them creates scale and scope efficiencies.

Second, aggregators can raise large amounts of capital at cheaper rates than individual merchants could achieve on their own. The reason is that these aggregators have access to venture capital, which at least currently, is plentiful. Small individual merchants are not at venture scale and so cannot access this funding, unless they are aggregated together into a portfolio.

And third, and more controversially, some investors might be drawn by the fast top-line growth in revenue that can be shown by aggregators buying more and more merchants. Of course, this does not represent real growth in the underlying businesses, but it might be hard for investors to perfectly disentangle real growth due to synergies (first point above) from growth that is “manufactured” via ongoing acquisitions.

So how sustainable are the arbitrage opportunities exploited by merchant aggregators?

It seems pretty clear that the capital arbitrage opportunity (the second and third factors above) on Amazon will be wiped out pretty quickly by competition. Not surprisingly, the prices aggregators have to pay to acquire merchants have already been bid up significantly (by now, some are even giving away Teslas to convince merchants to sell). Furthermore, there is a sense in which this is a zero-sum game on Amazon: since many of the synergies revolve around improving marketing and sales, if all the good merchants on Amazon improve their marketing and sales, there is no sustainable competitive advantage being created by the aggregators.

On the other hand, there are a couple of factors that point towards a positive-sum game and bode well for the future of the aggregator business model. First, Amazon is massive and still growing fast, with one estimate being that it has 5 million merchants, with more than 1 million added in 2020. And second, there is an even larger opportunity to apply the roll-up strategy across ecommerce more broadly (e.g. eBay, Etsy, Shopify, BigCommerce, Taobao, etc.). This need not be limited to an individual platform. Una Brands is applying the roll-up strategy across the multitude of marketplaces and other providers that merchants use in Asia-Pacific. And further synergies arise when aggregators can help merchants that are only selling in one channel access other channels (e.g. taking a good Amazon seller and leveraging it on eBay as well as in its direct channel via Shopify).

Aggregating merchants on digital marketplaces vs. direct-to-consumer merchants

While the bulk of these aggregators have been focused on Amazon, as noted earlier, more recently several roll-ups have emerged for merchants that are built on Shopify. This raises the interesting question of whether the aggregation strategy works better with merchants on digital marketplaces like Amazon’s, or with online merchants that sell directly (powered by the likes of Shopify and BigCommerce)?

The advantage of doing this on Amazon is the massive reach that sellers can obtain if they manage to rise to the top (e.g. if they can win Amazon’s Buy Box). This makes it possible for an Amazon aggregator to scale up individual businesses 10X or perhaps even 100X if they manage to get things right. The reason is that digital marketplaces create powerful self-reinforcing cycles around desirable products. The more a product aligns with what consumers are searching for, the higher its ratings will be, increasing the chances that the platform’s algorithms will drive target customers to it. That means more people will buy and rate the product, further heightening its advantage.

By comparison, the upside from aggregating individual merchants that sell directly (so have to find their own customers) will tend to be more limited, at least in the short run. On the other hand, there may be more scope in the long run to develop unique and defensible brands based on Shopify merchants, since Shopify merchants fully own and control the relationship with their customers. Not only does this mean the merchants being aggregated are more defensible, but it allows for much better cross-selling opportunities. As Keith Rabois, a prominent VC and founder of Shopify aggregator OpenStore notes, “We want to have control about how we stitch these together … And obviously Amazon is not going to just let us do whatever we want.”

This suggests that some merchants may be inherently more suitable for roll-ups on top of digital marketplaces (e.g. sellers of fast-growing products, but perhaps with shorter lifecycles), while others are more suitable for roll-ups that sell directly via their own websites (e.g. sellers of products that can create defensible brands). And of course, this in turn makes it possible for aggregators to specialize accordingly.

The future of the aggregator strategy

The idea of rolling up small businesses and scaling them up is hardly new. Tiny has been doing this since 2014, successfully aggregating small technology-based companies. And Berlin Brands Group started aggregating D2C companies well before Thrasio was founded. However, the scale at which roll-ups can be done and the potential payoff from doing so have been turbo-charged by the emergence of large digital platforms and providers of software tools that support millions of merchants. So how far can this aggregator strategy go?

Right now, the capital that is flowing into aggregator businesses on Amazon feels like a gold rush. And the latecomers are unlikely to find much gold left. Prices to acquire high quality merchants are being bid up. And the potential gains from roll-ups will diminish as more and more merchants are aggregated across each market segment. Indeed, each aggregated merchant will be competing with a greater number of other similar (aggregated) rivals and the quality of the merchants that are still left to be acquired will decrease rapidly.

By now, the biggest potential gains from roll-ups likely lie outside Amazon. As we noted earlier, there are many other digital platforms to scour (e.g. eBay, Etsy, Faire, Taobao) and there is a lot more room to run in D2C (e.g. Shopify-powered merchants). Moreover, aggregators will increasingly look to build up brands of their own in order to create some defensibility, i.e. follow the approach of Unilever or Procter & Gamble. Perch has already taken some steps in this direction.

Finally, the aggregator strategy can also be applied to areas outside of ecommerce. Social media platforms (e.g. YouTube, Instagram, Snapchat, TikTok) provide a clear opportunity, as we have seen with Wave.tv. Other opportunities include aggregating small SaaS companies, podcasts (to some extent Spotify is already doing this), mobile apps, or Substack newsletters like ours (an aggregator could pool together multiple creators and help grow, monetize and cross-sell their audiences).

In short, we think there is a lot more gold to be discovered by aggregators, but perhaps not where most investors are currently looking.

Many successful tech firms are content aggregators that rely on subscription-based business models: Calm, Masterclass, Netflix, Spotify, etc. As a result, in recent years many startups and their investors have become enamored by subscription pricing and are trying to make it work, even when sometimes it is not the best fit.

In this post we discuss the pros and cons of subscription pricing for aggregators of content, which determine when it should and should not be used.

Key benefits of pure subscription pricing

First, pure subscription pricing is simple and convenient for users. Customers pay a fixed subscription every month (or every year) in exchange for unlimited access to all available content. This affords them the peace of mind of not having to worry about which and how much content they can afford to consume.

Second, subscription pricing is effectively a form of bundling: consumers gain access to all products at a single price. This works extremely well when it reduces the variance in consumers’ willingness-to-pay across the various products. To illustrate how this works, here is a simple numerical example in which charging a single subscription price for a bundle is optimal.

Example 1

It can be easily verified that in this situation, when products 1-4 are priced independently, the optimal prices are $60 for product 1, $65 for product 2, $70 for product 3 and $75 for product 4, so each product only gets sold to the consumer that values it the highest, and the firm’s total profit is $270. Meanwhile, when products 1-4 are sold as a bundle, the optimal price for the bundle is $90, so consumers 1-4 get all products, and the firm’s total profits is $360, which is 33% higher. And consumers are better off too: consumers 2-4 are left with a positive surplus rather than nothing. This demand scenario captures the “magic” of bundling: because consumers 1-4 have very similar total willingness-to-pay but differ in their favorite product, bundling reduces the variance in willingness-to-pay across different consumer segments and therefore allows the firm to extract a lot more value from consumers while also making consumers better off.  

This is an argument that has long been understood by economists, and has more recently been popularized for entrepreneurs and investors in various blog posts such as this. While many of the examples are based on just two products that are bundled together, economists have also shown the same logic works well when a large number of different products are bundled together, as is the case for streaming services.

The bundling mechanism illustrated by our numerical example above is closely related to Shihir Mehrotra’s arguments here and here. As he explains it, bundling of content increases access to casual fans, which means consumers also benefit: with bundling, consumer segments 1-4 get access to all products, instead of one each. The logic of this example arguably applies reasonably well in the case of cable, Calm, Netflix, and Spotify. Each of these services provides access to a variety of content and different users may have very different favorite channels, meditation tracks, movies or songs. However, different users may be quite similar in their total willingness-to-pay for the entire bundle of content.

Key downsides of pure subscription pricing

The main problem with pure subscription pricing for a bundle of content is that revenue per user is fixed (by definition), irrespective of how many products each user consumes. This is particularly problematic when there is a fixed marginal cost per product. In those cases, revenue will not be able to cover cost for the most enthusiastic customers of the service, which means the business model will likely fail. MoviePass is the most spectacular example of inexplicably failing to grasp this: in exchange for a $10 monthly subscription, users could go watch a movie every day in a theater of their choosing (the ticket was bought by MoviePass at retail price), which predictably led to the company’s demise in 2019. And the same issue also plagued ClassPass early on when it was offering access to an unlimited number of fitness classes in exchange for a fixed monthly fee (the unlimited plan was later eliminated).

And even if marginal costs are zero or proportional to revenue (e.g. proportional licensing fees for Netflix and Spotify), subscription pricing is sub-optimal and leaves a lot of money on the table when user demand for variety is positively correlated with total willingness-to-pay. This is because subscription pricing (bundling) throws away the ability to charge more to users that consume more products. To illustrate this, consider the following demand scenario.

Example 2

Unlike in our first example above, here it clearly makes no sense to bundle. If the products were all bundled, the optimal price for the bundle would be either $20 or $30, resulting in total profits of $60 either way. Instead, each product should be sold independently (a la carte) at $10, resulting in a total profit of $100. Any bundling (of two or more products) would do worse than this. The key is that here a consumer that is interested in more products has a higher total willingness-to-pay, so there is perfect correlation between demand for variety and willingness-to-pay, which means selling a bundle of content under a subscription is inefficient.

To further emphasize the importance of the positive correlation between demand for content variety and total willingness-to-pay, suppose instead the configuration was like this:

Example 3

Here, the correlation between demand for variety and total willingness-to-pay is gone (the consumers that are interested in fewer products value each of them more), so bundling actually makes sense once again: $40 for all 4 products, yielding $160 in total profits. Independent pricing would make at most $105.

Obviously, real-world scenarios are more complicated than these simple numerical examples and possibly combine aspects of each. For instance, some Spotify users may love listening to the same songs or playlists over and over again, whereas others may prefer to always explore and discover new artists and songs. And at the same time, it is also possible that total willingness-to-pay for the service is positively (though not perfectly) correlated with the number of songs a user is interested in. This positive correlation seems even more plausible in the case of Netflix: users who watch more movies and shows are likely willing to pay more for the service than users who watch less content. Yet Netflix maintains a pure subscription model, likely because it believes the simplicity and convenience benefits mentioned above are more important.

The second (and related) big issue with offering a bundle of content under a pure subscription price is that it cannot sufficiently reward or incentivize superstar creators.

To see this clearly, consider the following scenario.

Example 4

This is just like our first numerical example above, except we added “superstar” product 5, which everyone values at $70. It is easily seen that product 5 should not be sold as part of any bundle, but rather priced independently at $70 (products 1-4 should still be bundled for $90). The reason is that if product 5 were sold as part of a bundle, consumer 5 would not get access to it. Indeed, to ensure consumer 5 can access product 5 under bundling (subscription) would require selling the bundle for $110, which would generate $550 of profit, as opposed to the optimal bundle price of $160, which generates $640 in profit.

So even if the firm were to pay the product 5 supplier the entire value created by product 5, under bundling it would only be able to pay $280, as opposed to the $350 that product 5 can yield by being sold independently.

The key point here is that offering a bundle of content under a pure subscription model makes it difficult to sufficiently reward “superstar” content (that many users want) and its creators. For example, this is why HBO and UFC fights are typically not included in the cable bundle and must be purchased separately. Similarly, this is also why online videogame stores associated with the latest consoles (PlayStation 5 Store, Nintendo Switch Games Store) leave the most popular games outside of their subscription plans. And this also explains why Netflix cannot afford to offer any recent hit movies (that it has not produced itself) through its service: the studios who produced those movies charge licensing fees that would quickly overwhelm the monthly subscription fee charged by Netflix to its users. For instance, right now in the US, Godzilla vs. Kong is available for $6 as a rental or $20 as a purchase on Amazon Video, whereas Netflix’s standard subscription price is $14/month. This comparison also shows how a mixed pricing model, with subscription (bundle) for a baseline library of content plus a la carte pricing for hit content (as is the case with Amazon Video), can be better than a pure subscription model (Netflix).

One may counter that firms can split the subscription revenue among their creators in proportion to the usage of their content, so more popular creators get paid more. This practice is used by Spotify and has also been implemented by some recent startups. For example, Playbook provides access to a bundle of workouts designed by “more than 100 trainers and athletes” in exchange for a single subscription fee of $15/month or $100/year. In an effort to incentivize popular trainers to bring their existing customers or social media followers to the platform, Playbook even lets every trainer keep 80% of the subscription revenue from users who sign up via the direct link provided by the trainer. These subscribers still get the benefit of accessing all available fitness content on Playbook. Still, it is important to realize this does not fully solve the problem because the subscription price remains the same for all consumers, and so superstars may not be able to extract the full value they offer, as illustrated in numerical examples 2 and 4 above.

This is particularly problematic when motivating creators to provide great content is very important, as is the case with many platform businesses. If the content already exists and it’s just a matter of distributing it through an additional channel, then not much is lost in terms of incentives by bundling and sharing revenues on the back-end in proportion to usage (e.g. Spotify). But if creators need to be constantly incentivized to produce new high-quality content, then it may be better to let them sell their content independently (a la carte) so that more popular creators can reap a bigger portion of the marginal value they create. This is why the pure subscription pricing model is not a good fit for platform businesses in the long run, given that they are all about giving control over pricing (and other decisions) to the supply side. If Playbook, which pitches itself as the “#1 platform for health & wellness creators” (and their fans) really wants to live up to its platform potential and attract superstar trainers, it will need to consider moving away from the pure subscription model by giving creators more control over their pricing.

Conclusion

If there is a large ongoing cost associated with additional content consumption, then bundling content under a pure subscription pricing model is doomed from day one. If there is a significant positive correlation between user demand for variety and total willingness-to-pay for the service, or if most of the content is supplied by “superstar” providers that need to be appropriately incentivized, then adopting a pure subscription model is also a bad idea.

On the other hand, if costs don’t overwhelm fixed subscription revenue per user and if few content creators have strong bargaining power, then firms can start with pure subscription pricing in order to make things as simple and as convenient as possible for users. Later on, they can move to a mixed pricing model of subscription plus a la carte prices for highly popular (or valuable) content. Indeed, as services that started off with pure subscription pricing add more content (e.g. Calm and Netflix), the revenue limitations of the subscription model will become clearer and more pressing, especially once subscriber growth slows down. Once the service runs out of room to grow by adding new subscribers at a fast-enough rate, it makes sense to unlock revenue growth opportunities per subscriber by at least adding some a la carte elements.

In this post, we emphasize a similar point with the example of Steezy, a fast-growing start-up providing online dance classes, but where the main additional value comes from providing the consumer side with a better experience.

In principle, one can find hundreds of thousands of free videos for helping people learn how to dance on YouTube, the ultimate marketplace for digital content. However, it is precisely that dizzying variety of content on YouTube that creates the opportunity for specialized resellers like Steezy.

Steezy is a reseller in the same way Netflix is: it either licenses or produces its own content, and then sells access to this content on a monthly or yearly subscription. Unlike YouTube, Steezy takes full control over how the content is produced and provided. For example, dancers producing content for Steezy have their dance moves professionally filmed in Steezy’s own studios to meet Steezy’s exacting standards. They sign over the rights to this content for Steezy to distribute. In contrast, dancers on YouTube have full control over their content creation (as long as it is not considered objectionable) and can remove or change their content whenever and however they want.

All this control allows Steezy to offer greater value to consumers. The content is high quality, standardized and neatly categorized, making it easy to find and access the specific content that best fits an individual consumer’s preferences (as opposed to endlessly browsing on YouTube). Steezy’s classes are optimized to teach subscribers short choreographies that they can later do on their own, as part of an integrated curriculum to learn a particular dance style. Subscribers can make use of viewing features that are designed specifically for learning dance, like looping a selected part of the video, switching between front or back view, or watching the video in mirror mode (so that the user can simply copy the instructor).

Together, all these aspects make the learning experience a lot better than what one could get on YouTube. Yes, YouTube offers orders of magnitude more variety of dance videos (for free), but as it turns out, the quality of the user experience matters more than content variety in this context. Of course, Steezy is working hard to increase the variety of different types of dances it offers lessons for, but it will never be able to match what is available on YouTube. Nor should it try to do so. Instead, its current vision is to become the “Disney of dance,” which clearly emphasizes its focus on providing a consistently good consumer experience.

The same arguments that explain why Steezy is best positioned as a reseller, should also apply to many other verticals. For example, they also explain how Calm and Headspace were able to build billion dollar businesses by focusing on online mediation and mindfulness, despite the existence of hundreds of thousands of free meditation and mindfulness videos on YouTube. Like Netflix and Steezy, they produce their own content (featuring superstars like LeBron James and Matthew McConnaughey for Calm) and offer it through their apps to users in exchange for monthly or annual subscriptions. For online fitness, another massive vertical on YouTube, there are many recent start-ups providing curated online video content via subscriptions, such as Fit!, which leverages select instructors to create and offer exercise programs targeted at specific use cases (golf, handstands, skiing, surfing, parkour, etc). Another start-up example in a different vertical is The Crafter’s Box. In exchange for a monthly subscription, users can learn a new craft each month: the company mails them a box of materials and provides an associated video in which an instructor shows them how to produce something. Like in the previous examples, The Crafter’s Box offers a much more convenient and curated experience (albeit for a much higher price) for people who want to try out new crafts, than what they could get by watching YouTube and purchasing supplies at their local craft shop.

Once such resellers have established themselves, they can still of course consider adding platform elements, using the methods we discussed in previous posts. In the case of Steezy, it has already started connecting its customers by creating a community of Steezy dancers that take part in weekly dance challenges among other activities. It could also contemplate opening the doors to third-parties, who want to sell products or services to its subscribers. This could be dance venues, or independent dance instructors who would like the chance to offer Steezy users one-on-one coaching or live online dance classes which complement Steezy’s core offering. By opening such a marketplace for third-party instructors, Steezy would have to give up some quality control and would risk hurting its brand, but at the same time it can expand to new formats or new types of dance in a much more cost-efficient way. And if some of these new third-party offerings turn out to be very popular and lucrative, it can always bring them in house later. This is of course the Amazon story: it started off as a reseller of books before expanding into reselling other products, then opened up a marketplace to cover virtually every imaginable product, and now it uses the marketplace to learn which third-party products perform the best so that it can sometimes offer them itself.

In conclusion, it is important to be able to recognize markets in which the quality of the user experience can trump the demand for content or supplier variety, so that specialized resellers (e.g. Steezy) can co-exist with horizontal marketplaces (e.g. YouTube). We believe there are many such opportunities for start-ups to leverage.

If you enjoyed reading this, sign up now so you will receive our posts directly in your inbox as they are published.

Subscribe now

In the meantime, tell your friends!

Examples include Amazon, Target, Walmart, Flipkart in India, JD.com in China, each of whom operates an online marketplace and also sells (or resells) products itself side-by-side with third party sellers on the same online marketplace. And many other platform companies also sell their own products on their marketplaces: Apple in its App Store, Google in its Play Store, Microsoft in Windows’ Apps, Intuit in its QuickBoks Apps, Salesforce in its App Exchange, etc.

Regulators around the world are increasingly proposing new antitrust laws that seek to ban large platforms from operating in dual mode. In this post we explain why such policies do not make sense.

The regulatory concerns

In recent years significant policy concerns have been raised about big-tech firms stifling competition. In late 2018, India passed a law that banned Amazon and Flipkart from selling products (on their respective marketplaces) of companies in which they have an equity stake, thus effectively banning the dual mode. And earlier this month new proposed antitrust laws have been unveiled in the United States, which among many other things, seem to be targeted at banning large platforms from continuing to operate in dual mode.

One proposed Act (the Ending Platform Monopolies Act) would make it unlawful for “covered platforms” to own or operate a business that presents a clear “conflict of interest”. Based on our reading of the Act, a covered platform is carefully designed so that it covers Alphabet/Google, Amazon, Apple, Facebook, and Microsoft, but no other firms. A conflict of interest arises when the “covered platform’s ownership or control of that line of business creates the incentive and ability for the covered platform to advantage the covered platform operator’s own products, services, or lines of business on the covered platform over those of a competing business…” Clearly this is aimed at stopping Amazon and Apple from operating in dual mode, given that they face a conflict of interest in selling their own competing products or apps in competition with third-parties on their respective marketplaces.

It is worth highlighting another proposed Act that is very much related. This is the American Choice and Innovation Online Act, which would prohibit many activities that involve self-preferencing or discrimination on covered platforms. The overall goal of this Act seems to be to stop any platform that faces a conflict of interest from acting on it.

Perhaps legislators proposing these bills are hedging their bets, but if the American Choice and Innovation Online Act is enacted, then there doesn’t seem to be much need for the more drastic Ending Platform Monopolies Act. Indeed, the concerns usually raised about platforms operating in dual mode are exactly the behaviors banned under the American Choice and Innovation Online Act.

Specifically, there are two main concerns that are commonly raised about platforms that operate in the dual mode:

1. That the platform obtains proprietary information on the third-party sellers’ products (e.g. detailed demand and pricing data, data on users' search behavior, what items they return and their reasons for doing so) via its marketplace, and then uses that information opportunistically to copy and compete on the more successful offerings, potentially leading to reduced incentives for third-party sellers to invest or innovate. For example, it has been claimed that some employees working as part of Amazon’s private labels group used data from its marketplace to inform their decisions about launching private label products, even though this is at odds with the company’s stated policies.

2. That the platform can steer consumers towards its own offerings (or affiliated products), rather than those offered by third-party sellers, by displaying its own offerings more prominently, a practice that has become known as “self-preferencing”. For example, Amazon can do this through its Buy Box, on which around 85% of all Amazon purchases are made. The Buy Box allocates a seller to the consumer when they click on “Buy now” or “Add to cart” according to a secret algorithm that Amazon controls, and oftentimes the allocated seller is Amazon itself.

Why banning the dual mode is bad policy

To begin with, it’s worth pointing out that if these concerns can be dealt with by behavioral regulations targeted at eliminating self-preferencing and the sharing of data across divisions that face a clear conflict of interest, then any remaining benefits of banning the dual mode are likely to be quite limited.

More importantly, the dual mode provides many clear efficiencies, which would be lost under such a heavy-handed, structural ban.

First, as noted in a previous post, some types of products may be more efficiently provided by the platform, while others may be more efficiently provided by third-party sellers. And there are clear economies of scale or scope from offering all products together on one site (i.e. in dual mode):

• One-stop shopping benefits to consumers (search and checkout) relative to having to search and purchase on two different websites (e.g. Amazon’s marketplace and Amazon as an ecommerce retailer).

• Cost savings to the platform from only having to build and manage one common website.

• Cost savings for users if the shipping of the platform’s and the third-parties’ products can sometimes be combined into single orders.

• The platform can make better recommendations between its own products and those of third-party sellers when they are all offered on the same site.

These economies of scale and scope should already make it clear that banning the dual mode across products does not make much sense.

Second, even when focusing on an individual product or product category, the dual mode allows the platform to create more competition within each product category. Indeed, by being able to step in and offer a product directly at a lower price, the platform ensures consumers will get more choice and lower prices on the marketplace in case established sellers have market power. In addition to creating price competition, the platform can also offer superior customer service or logistics, which puts pressure on third-party sellers to do likewise, thus benefiting consumers.

Third and relatedly, the platform can also help avoid stock-outs by increasing the supply of items that are in high-demand. It may even be willing to sacrifice margins to be the supplier of last resort given that it cares about customer experience on the marketplace much more than any individual third-party seller would (a consumer coming back to the marketplace may or may not buy from the same seller, but always benefits the platform).

Given the benefits of operating in the dual mode, it is not surprising that it is used by an increasing number of companies who open up marketplaces where they sell their own products alongside third-parties: Best Buy, Intuit, Salesforce, Shopify, Twilio, etc. In fact, there is now even a successful startup, Mirakl, which provides companies the necessary tools to create such marketplaces.

Concluding thoughts

It is striking to us that the benefits of the dual mode are clear and obvious, while the harms that can arise from it are not. The harms are based on behaviors (self-preferencing and accessing third-party data, as outlined above) that have been alleged but have yet to be properly empirically documented at scale. The fact that a dual-mode platform can tilt the playing field on its own marketplace in its favor vs. third-party sellers does not necessarily mean it will always do so. For instance, if Amazon was systematically steering consumers to less than ideal products, or blatantly crushing its third-party merchants, over time it would lose both merchants and consumers, to its own long-term detriment. Moreover, if self-preferencing and accessing third-party data can be established as significant problems, they can be targeted via competition law or new regulations that address the concerns directly (such as those contained in the proposed American Choice and Innovation Online Act, or in Europe, by the proposed Digital Markets Act) rather than imposing a full-scale breakup.

Lastly, given the obvious implications this discussion has for Amazon and Apple, we should report that neither of us has done any consulting work for Amazon or Apple, nor (unfortunately) does either of us hold Amazon or Apple stock, least readers suspect we are in their pockets. On the other hand, together with our co-author Tat-How Teh, we have written an academic paper on this topic which provides a much more in-depth analysis of the issues discussed here (for the more mathematically-inclined readers).

If you enjoyed reading this, sign up now so you will receive our posts directly in your inbox as they are published.

Subscribe now

In the meantime, tell your friends!

The company was initially launched in 2009 as a peer-to-peer online marketplace for swapping men’s dress shirts, more specifically “men’s long sleeve button-down shirts in the $70-$90 range”. It didn’t take ThredUp long to figure out that this was not the best clothing item to build a marketplace around. Within a year the company pivoted to a peer-to-peer marketplace for children’s clothing, where sellers were selling directly to buyers. This was a much better market opportunity given the obvious issue with children’s clothes: they are durable but children grow out of them quickly. ThredUp raised venture funding on this marketplace concept and achieved decent traction with it.

However, in 2012 ThredUp went back and raised another venture round in order to execute a more drastic pivot. It turned itself from a marketplace into something close to a reseller. Specifically, it built a warehouse and switched to acquiring the children’s clothes from sellers (on consignment), inspecting them, and then selling them to buyers. And in 2013 ThredUp expanded to women’s clothing, also under the consignment reseller model.

Our focus in this post is ThredUp’s marketplace-to-reseller pivot. This looks similar to Zappos’ early pivot (which we discussed in our previous post), but the fundamental driver was different. Zappos’ shift was about maximizing the quality of the experience for buyers. ThredUp did it in order to enhance the experience for sellers. In particular, ThredUp seized upon an important arbitrage opportunity: sellers of used clothes were never going to become “power sellers” trying to make a business out of that (like eBay or even Poshmark sellers). More than anything, ThredUp’s sellers were looking for an easy way to dispose of used clothes and hopefully make some money out of them. Convenience is what mattered most for these sellers and ThredUp figured out this was better achieved under its reseller model than under the marketplace model.

To be precise, ThredUp’s current model is close to the reseller end of the marketplace-reseller spectrum, but it is not 100% of the way there. Indeed, consignment means sellers do have some amount of control over how their items get sold to buyers. Specifically, ThredUp sends sellers a bag (or allows them to print a label to be used on any box), which sellers fill with the clothes they wish to sell to ThredUp. The company inspects the clothes, determines which items are eligible to sell (based on their quality and condition), and then sells them to buyers. The prices are set by ThredUp, but sellers do have the option to adjust those prices. Sellers get paid when their eligible items are sold. Sellers also have the option to have their unlisted items (those which are not eligible to sell) returned, in exchange for a fee. Otherwise, ThredUp disposes of them.

To illustrate the difference between the pre-pivot marketplace and post-pivot reseller business model in this context, it is instructive to look at the following table. These numbers are purely hypothetical.

The brilliance of ThredUp’s 2012 pivot was realizing that in return for the higher convenience of the reseller model, sellers were willing to accept a much lower but certain price ($50 in the example above), instead of the hypothetical price they might have hoped to obtain on the pre-pivot marketplace ($100), where they would have had to do the listing, selling and shipping themselves. That arbitrage opportunity had to offset the much higher operational costs associated with the reseller mode ($25 in the example above), as well as the fixed capital investment costs (building a warehouse), which of course become easier to cover at scale.

In the example above, the price to buyers is kept constant for simplicity ($100), but it is likely that in the reseller model ThredUp can also extract a bit more from buyers due to the superior experience, pricing knowledge and the possibility of mixing and matching clothes from multiple sellers. This makes the reseller mode even more attractive relative to the marketplace mode. However, it is important to emphasize that the biggest share of the gain comes from the price-convenience arbitrage on the seller side.

Concluding thoughts

The arbitrage logic illustrated by ThredUp is that when sellers are more interested in speed and convenience than in extracting the highest possible price for their items, it makes sense to be closer to the reseller end of the marketplace-reseller spectrum (all other things equal). This is analogous to services in a few other markets:

• Pawn shops

• Carmax for used cars

• Opendoor for houses

• Trade-in services for used smart phones and other electronics, such as Gazelle, Decluttr, Backflip

However, not all sellers in a given market are the same. This is one reason why the reseller mode can coexist with the marketplace mode in the same market. For instance, in the market for women’s fashion, consignment resellers like The RealReal and ThredUp can appeal to sellers who value convenience more than financial return or who do not have much expertise in selling. Meanwhile, marketplaces like depop and Poshmark can appeal to sellers who don’t mind spending time listing and selling their items, or who are professionals. And Vestiaire Collective offers both options. This is similar to the co-existence of Cars.com (marketplace) and Carvana (reseller) in the market for used cars, which we noted in our last post.

Finally, it is interesting to ask whether it ever makes sense to intentionally follow ThredUp’s (or Zappos’) path of starting as a marketplace and then switching to be a reseller. We suspect that their path and the pivots they made were not part of a masterplan when they first launched. If they knew what they know now, they would have probably chosen to start as resellers instead. This overcomes the chicken-and-egg problem associated with launching a new marketplace and delivers a consistently high-quality buyer and seller experience from the get-go. Of course, in practice, firms must sometimes experiment in order to figure out which model works best in their market (and indeed whether there is a viable model at all). It may make sense to start as a marketplace in order to find out at low cost which products or services are in highest demand, and then move towards the reseller mode in order to optimize the buyer and/or seller experience. This avoids the higher upfront capital expenditures associated with the reseller mode, which may make it harder to convince investors to fund the business at first.

The control continuum

As we explained in our post on what defines platform businesses, there is a continuum between pure marketplaces (e.g. Craigslist, eBay), which enable independent sellers to directly sell their products/services to buyers, and pure resellers (e.g. Zappos, 7-Eleven), which buy the products from sellers and then resell them to buyers.

The continuum is defined by how much control is taken by the intermediary over key aspects of the corresponding transactions: pricing, marketing, order fulfilment, customer service, etc. This continuum exists both for products

and for services

Department stores like Macy’s are in the middle because they sell some products (e.g. kitchenware, clothing) in their own name, while allowing independent brands to set up counters that they design and control for selling other products (e.g. cosmetics). Similarly, Amazon is in the middle (and increasingly moving towards the marketplace end of the spectrum based on number of products): it sells some products in its own name, alongside products sold on terms controlled by third-party sellers. Meanwhile, Uber is in the middle because it controls certain aspects of the driver-rider interactions (price and choice of whom to serve) while allowing drivers to control other aspects (choice of car, work schedule).

The marketplace-reseller continuum creates a fundamental tradeoff for companies. Moving towards the pure marketplace end of the spectrum makes businesses more scalable since they don’t have to incur the costs associated with taking ownership of and responsibility for products/services and their delivery. On the other hand, moving towards the reseller end of the spectrum incurs higher capital and operational costs, but the higher degree of control affords the ability to offer a better customer experience – for buyers, or sellers, or both.

Choosing the optimal position on the control continuum

Given this fundamental tradeoff between cost and control, there are several factors that can push the optimal choice of positioning on the marketplace-reseller continuum one way or the other.

1. Information

The first factor is, which party, the independent suppliers or the intermediary, has better information about how best to set the relevant decision variables like prices and other marketing activities. Given contracts are rarely perfect or complete, it matters who gets to make these decisions, and more value can be created if the control is given to the party with the better information. This is a key reason why all successful ride sharing companies (Didi, Grab, Lyft, Uber) set the prices of the rides they enable, whereas on freelancer marketplaces like Fiverr, Task Rabbit and Upwork, prices are set by the independent sellers. This is also why department stores have traditionally offered cosmetics through dedicated “counters” where displays are designed and controlled by individual brands (and sales staff are assigned exclusively to and trained by specific brands), whereas many other more standard products are offered on displays controlled by the department store and serviced by generalist sales personnel. It reflects that cosmetics brands have highly specialized knowledge about how best to market specific products to consumers, which is hard for stores and their sales staff to accumulate. This principle also guided Amazon’s expansion from an online book reseller to “the everything store”. Since it didn’t have expertise in selling much more than books initially, when it wanted to expand into electronics, it relied on third-party sellers who were better placed to decide what to stock, how to price and how best to market products.

2. Value of a better buyer/seller experience

As mentioned earlier, the higher degree of control associated with moving towards the reseller end of the continuum is associated with higher costs (so less scalability), but at the same time allows the intermediary to optimize the experience for buyers or sellers or both. This usually means faster, more consistent and reliable service. The key question is then just how much more valuable is that superior experience in the eyes of buyers and/or sellers. In some cases, it can create significant additional willingness-to-pay and customer loyalty, making the reseller model more appropriate.

As an example, it is useful to remember Zappos started off in 1999 as an online marketplace for shoes, where products were priced and fulfilled by third-party sellers/brands. However, after realizing that under this model it could not guarantee a consistent experience to its buyers, the company turned itself into a pure reseller by 2002, stocking inventory and taking full control of transactions with buyers. This allowed Zappos to offer its buyers unparalleled customer service, which eventually became the company’s key source of competitive advantage.

In other contexts, however, being closer to a pure marketplace does not entail a big downside in terms of the experience perceived by buyers and sellers, but a significant upside in terms of scalability. For instance, Etsy could certainly offer better services to both its buyers and sellers of arts and crafts, by moving towards the reseller model, but the added value in terms of convenience to both sides would be quite limited (unlike the case of Zappos). Meanwhile, the reseller model would be prohibitively costly in this case.

Many companies today try to strike a balance between scalability and very good customer experience by choosing a position somewhere in the middle of the continuum, which VCs and founders call “managed marketplaces”. We mentioned Uber above, which controls the price and matching of riders with drivers. As another example, marketplaces in the pre-owned luxury fashion industry like StyleTribute, TheRealReal, and Vestiaire Collective are typically highly managed. They authenticate the items being sold by third-party sellers, the way they are marketed, the way they are boxed and shipped, and the returns policy, so although they don’t set the price or take ownership of the items (they are normally sold on consignment), they handle almost everything else. 

3. Long-tail vs short-tail products

High-demand (short-tail) products are sold more efficiently by one large reseller than by many small sellers. The reseller can capitalize on economies of scale in purchasing, infrastructure investments (in warehouses and distribution centers, for example), delivery, customer support, and so on. These advantages, however, don’t apply so much to low-demand—or “long-tail”—products. This is why, in general, marketplaces work best for long-tail products or services – think eBay, Etsy, and TaskRabbit.

And this is why Amazon (who has the ability to act as either a marketplace or a reseller for any given product) mostly chooses the reseller mode for high-demand products (e.g. electronic accessories and gadgets, daily household supplies), while mostly acting as a marketplace for long-tail products (e.g. inflatable sloth floats or mobile phone jail cells), leaving them to independent sellers.

4. Cross-product spillovers

The fourth factor which determines an intermediary’s positioning between marketplace and reseller is the presence of demand interdependencies across products. If marketing, pricing, or other types of decisions for one product generate strong effects on other products (i.e. there are spillovers), the reseller mode becomes more attractive compared to the marketplace mode. This is because a reseller can more easily internalize these externalities when it is making its pricing and other marketing decisions. This can help explain why cable TV operators and Netflix predominately choose to operate in the reseller mode, extracting more value by bundling and cross marketing different channels or different shows/movies.

5. Overcoming the cold start problem

The biggest hurdle faced by any marketplace in its early days is overcoming the chicken-and-egg problem of attracting both buyers and sellers, when it does not have critical mass of either. A natural way to solve this chicken-and-egg problem is to start in reseller mode, buying products from sellers and reselling them to buyers, thus ensuring there will be an adequate range of supply from the start. Once enough buyers participate, the intermediary can switch to the marketplace mode. There are several examples of firms that have used this playbook. Amazon started as a reseller of books, before it had enough customers that it could invite third-party suppliers to sell their products. Similarly, Uber initially hired drivers to drive around specific areas in exchange for fixed hourly rates in order to generate enough demand, before it transitioned to drivers as independent contractors, paid per ride.

Concluding thoughts

At a high level, marketplaces are volume businesses: the goal is to enable as many transactions as possible. Meanwhile, resellers are arbitrage businesses: the goal is to find products or services that can be sold at a significantly higher price than what they can be bought or provided for. The arbitrage opportunity comes from one or several of the factors described above.

It is rare that all factors point in the same direction, i.e. either pure marketplace or pure reseller, so the optimal position is usually somewhere in the middle, leveraging the advantages of both modes. This explains why there are few if any pure, bare-bones marketplaces today (like Craigslist): all modern marketplaces are managed to a certain extent, because they found ways to profitably offer valued added services to one or both sides.

Finally, it is important to note that in some markets, companies may find success with different positions on the marketplace-reseller continuum, by appealing to different segments of buyers and/or sellers. In other words, there isn’t necessarily a unique “peak” position in a given market (e.g. eBay vs. Amazon vs. Zappos, Opendoor vs. Redfin vs. Zillow, Cars.com vs. Carvana). In particular, the optimal positioning for a startup entering the market should take into account where incumbents are located on the continuum – sometimes it may be better to focus on the space left by others.