Algorithmic Network Effects

Digital platforms build on direct and indirect network effects, as previous platforms, but they also rely on a new type of network effects: algorithmic network effects. Big data is necessary for the provision of some popular digital service, the best example of which is search. Google relies on the massive amount of previous searches to respond to a specific search.26 In the words of the European Commission: “because a general search service uses search data to refine the relevance of its general search results pages, it needs to receive a certain volume of queries in order to compete viably. The greater the number of queries a general search service receives, the quicker it is able to detect a change in user behaviour patterns and update and improve its relevance.”27

Algorithmic network effects mean that a service improves as it obtains more data from users: the more the service is used, the more data is captured; and the more data is captured and processed, the better the service becomes.28 For this reason, data has been considered “the new oil.”29 Amassing data, if properly managed, can provide a powerful competitive advantage. Since Google has the largest pool of search history on the planet, it has a competitive advantage. Bing’s technology might be as good as Google’s, and it is only “a click away,” but being way behind Google’s pool of data, it cannot replicate the quality of Google’s searches. This is again a network effect, and is so powerful that tips the market in favor of the largest player.

Algorithmic network effects play a fundamental role for digital platforms as they are key to overcoming the most relevant risk of any popular network: congestion. As the number of users of a service or an infrastructure increases, it can become so crowded as to hinder or even prevent its use. Attention is usually focused on positive network effects, but negative network effects can be just as relevant and can counterbalance the benefits derived from pooling large pools of users. Congestion is the most relevant negative network effect. In case of congestion, a new user does not add value, but actually detracts value, as it further prevents the use of the service for others. We are all familiar with road congestion, as well as the usual policies to fight it: increase capacity, raise prices for the use of the service to reduce demand, and so on.

Technology has traditionally helped to reduce the problem of congestion in network industries. Mechanical automatic switching was introduced in the 1920s, solving the problem of congestion in early telephone networks. Digital switching, which made use of computers, further empowered telephone networks to switch millions of telephone calls.

Digital platforms have always faced the risk of congestion. It is not only that many start-ups could not always cope with exploding traffic by installing more and more servers. In a more fundamental way, the larger the pool of users, the more difficult it is to ensure a coherent and fruitful interaction between them. In a marketplace like eBay, having more sellers is key to success, but having millions of sellers and even more items for sale makes it complicated to ensure buyers can find the items they are looking for (and even more difficult to trigger the availability of things they might desire, but that they are not actively searching). In social networks such as Facebook, as the amount of users reaches the billions and each user is uploading more and more information, it is important to select which information is displayed to the viewers browsing Facebook. The same applies to videos proposed by YouTube, to apartments shown by Airbnb, and to drivers matched by Uber.

Sorting presents diseconomies of scale. Sorting becomes more difficult and expensive as the number of items to be sorted increases. “The unit costs of sorting, instead of falling, rises.”30 Automation has always been a key to the performance of digital platforms. Platforms use algorithms to automatize the matching decisions that allow interaction between the different sides in the platform. The Oxford Dictionary defines an algorithm as “a process or set of rules to be followed in calculations [...] especially by a computer.” Algorithms are increasingly using machine learning technology. They are not a set of fixed rigid commands predefining the links between the parties in the platform; on the contrary, algorithms peruse through the stored data in order to predict the most useful link between the different sides in the platform.

Algorithms improve upon themselves with each interaction.31 Google, Facebook, and Airbnb identify users’ reactions, for instance in the form of clicks in one of the search results, clicks in one of the videos proposed, or ratings by riders after using Uber. These reactions are incorporated into the algorithm in order to inform future services. In this way, algorithms are in a position to predict which information will be of most interest for future users. They predict the best result of a search, such as which YouTube video will be of most interest for the viewer or which apartments an Airbnb guest will prefer. They predict which driver will arrive soonest to pick up an Uber rider and provide the best service. More data makes predictions better.

Algorithmic network effects are a precondition for and reinforce direct and indirect network effects. Platforms will only succeed if they are able to efficiently intermediate between users. Large pools of users will only be attracted to the platform if the intermediation service is provided smoothly, adding value to the users. Only the most sophisticated algorithms are in the position to overcome the challenges of ensuring a productive interaction between billions of users. At the same time, the more users that are attracted to the platform, the better the algorithms can work and the better the service will get. The larger the pool of data and the better the algorithm that makes sense of it, the better the intermediation service provided by the digital platform can get.

Regulators have identified how algorithmic network effects create market power. The European Commission concluded that Google’s dominance in the search markets relied on barriers to entry and expansion, such as the large investment required to create the search engine, but “also needs to receive a certain volume of queries in order to improve the relevance of its results for uncommon (‘tail’) queries. [...] The greater the volume of data a general search service possesses for rare tail queries, the more users will perceive it as providing more relevant results for all types of queries.”32

The US Federal Trade Commission has identified the role algorithmic network effects in Facebook’s business model: “Advertisers pay billions—nearly $70 billion in 2019— to display their ads to specific ‘audiences’)...] created by Facebook using proprietary algorithms that analyze the vast quantity of user data the company collects regarding its users. This allows advertisers to target different campaigns and messages to different groups of users.”33

European antitrust authorities have analyzed the role of data in potential anticompetitive practices. In 2016, French and German authorities published a report entitled Competition Law and Data.34 In 2019, the Italian antitrust authority published a joint report with the national telecom and data protection authorities.35 The Bundeskartellamt adopted an innovative decision in February 201936 declaring that Facebook had abused its dominant position in the market of social networks for private users; that is, the zero-pricing service provided by Facebook to users. The authority concluded that Facebook imposed exploitative business terms in relation to data, particularly imposing the combination of data extracted from other corporate services (like WhatsApp) with data extracted from the Facebook service, in order to display targeted ads in Facebook. No fine was imposed, but the German antitrust authority imposed a data separation remedy. Data from different sources can now only be combined with the explicit consent of the user.

 
Source
< Prev   CONTENTS   Source   Next >