Knowledge is (market) power? : analysis of the effects of 'free' goods in online platforms on competition

(1)

Knowledge is (Market) Power?

Analysis of the eﬀects of ’free’ goods in online platforms on competition

6 August 2017

Sophie Wiedijk Supervisor: prof. dr. Maarten Pieter Schinkel

10656502 Faculty of Economics and Business

(2)

Abstract

As the digital economy gains in importance, some are concerned that dominant platforms can use their ability to collect, store, and analyse big data to create entry barriers in a manner that is harmful to consumers. However, economists do not agree on whether this is actually the case. In this thesis, the costs, benefits, and antitrust implications of the data-driven market power of online intermediaries is investigated. A model of platform competition is constructed in which an incumbent platform is able to oﬀer consumers personalised recommendations by using data it has collected about their online behaviour. Because consumers experience choice overload, this increases demand for the platform. However, consumers with a preference for privacy see their utility decrease as more data is collected, thus constraining the incumbent’s revenues from data collection. Then, when consumers are perfectly informed about the utility costs of disclosure, both consumer surplus and platform profits increase. Consumers that do not care about privacy benefit from the decrease in transaction fees necessary to get privacy-concerned consumers on board. When a potential entrant is introduced, it is shown that the ability to credibly commit to refrain from data collection increases consumer welfare when entry is accommodated. The reverse is true when the incumbent deters entry, since this is associated with an ineﬃciently low amount of data collection. Making it compulsory for platforms to share their data is then shown to reduce welfare, whereas prohibiting third-party data trade increases competition and consumer welfare by providing individuals with the choice to search through the platform that matches their privacy preferences.

Statement of Originality

This document is written by Sophie Wiedijk, who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

Chapter I

Introduction

In May 2017, the European Commission fined Facebook €110 million (Scott, 2017) . The reason for the fine was that the social network had given misleading statements in the 2014 assessment of its merger with online messaging service WhatsApp. At the time, the company had promised not to combine its data with that of WhatsApp, after which the merger was cleared. In August 2016, this promise was broken.

This is not the first broken promise by the tech giant. In 2009, Facebook’s CEO Mark Zuckerberg told users that the social network’s privacy settings had been made made easier to understand. Later, it turned out that the new settings implied that a significantly larger amount of data was collected about users of the network. Zuckerberg apologised. In 2011, he had to apologise again, this time to the Federal Trade Commission (FTC) (Martijn & Tokmetzis, 2016). Facebook had promised it would no longer share user data with advertisers. In reality, it continued to do so. Users were promised that if they deleted their accounts, any photos or videos they had shared in the past could no longer be viewed. This was a lie. Then, the company tried to console users of its social network by promising that certain information would be kept private. This also turned out not to be true (Martijn & Tokmetzis, 2016). At the time, Facebook was able to reach a settlement with the FTC by promising to no longer deceive consumers by failing to keep its promises (Federal Trade Commission, 2011). The company did not have to pay any fine, because it successfully argued that since consumers did not pay for the using the service, they could not be harmed by Facebook’s conduct (Newman, 2015). Yet this May, the firm was fined, even though the social network (as is stated when logging in to its website) ’is free and it always will be’. European Commissioner for Competition Margrethe Vestager defended the imposition of the fine by arguing that consumers did in fact pay for using the service, using their data as currency (European Commission, 2016).

The use of consumer data for commercial purposes has generated gains for both firms and con-sumers. However, many of the industries in which big data play an important role are characterised by high concentration, often to the extent that a single firm is dominant. Data-driven giants such as Facebook and Google1 _{are able to provide services for a price of zero because these firms have the}

characteristics of intermediaries or platforms. This means that users of the services they provide are subsidised by advertising revenues on the other side of the market. Thus, although consumers do not pay these firms with money, they do provide these platforms with profits by making the service more attractive to advertisers. Online platforms have recently started to raise concerns, including issues with regard to transparency, the manner in which platforms use the information they collect, the relationship between platforms and suppliers, and constraints on both individuals and businesses to move between diﬀerent platforms (European Commission, 2016a).

Thus, it might be the case that these dominant zero-price firms have both the ability and incentive

1_{Actually, the parent company of all Google services has been named Alphabet Inc. since 2015. In this thesis, the}

(7)

to abuse their dominant position, albeit not by raising their prices. Establishing consumer harm must then be done without relying on price theory, which can be challenging in the neoclassical framework used by most competition authorities. Furthermore, economists disagree about the extent to which antitrust enforcement is necessary in these online industries. Opponents argue that online market regulate themselves, that issues that relate to privacy should not be addressed by competition policy, and that entry barriers are low in these markets (Sokol and Comerford, 2016; Lambert and Tucker, 2015). On the other hand, proponents of antitrust scrutiny in online platform industries warn that the increasing data-induced market power of online platforms can harm consumers, stifle innovation, and that these intermediaries can use their data to exclude competitors from the market (Stucke and Grunes, 2015a). In addition to this, many of them warn that competition authorities should act now. The reason for this can be illustrated by the Collingridge dilemma, which is the idea that by the time a novel technology’s undesirable consequences are discovered, the technology is often so deeply entrenched into the economy and into society as a whole, that its control is extremely diﬃcult (Gal & Rubinfeld, 2016). Indeed, although individuals are becoming increasingly aware of the costs that large-scale data collection might entail, such sentiments have not reduced demand for these free online services. In order to gain insight into this subject and its implications for competition policy, in this thesis an attempt will be made to answer the following question:

Under which conditions do online platforms that offer ’free’ services harm competition? To this end, this thesis will commence with an extensive literature review in which the char-acteristics of big data and online platform markets will be discussed. Using these insights, as well as findings from previous research, an assessment will be made of whether the dynamics of online platform markets are likely to lead to dominance by one or a few firms and if so, if this would be harmful to consumers. Next, a model will be developed in which data collection by a monopoly platform has both costs and benefits for consumers. A potential entrant will be introduced that does not collect any data in order to assess what the effects of data collection, as well as the ability to commit to refrain from collecting data, are on competition and welfare. After this, the implica-tions of the equilibria that have been derived in the model will be discussed in order to arrive at recommendations for competition policy. A number of possible extensions to the model will also be elaborated on, both in terms of their effects of the model’s equilibria and as a suggestion for further research. The final chapter concludes.

(8)

Chapter II

Literature Review

In this chapter, a number of insights obtained from previous research on competition policy and its role in dealing with big data are discussed in detail. An overview is provided of the theories, debates, and economic models that have been developed with regard to this topic. This will place this analysis in perspective and will identify which of the characteristics of online platform markets need to be investigated to answer the research question.

To this end, the chapter commences with a disquisition regarding the theoretical background of the research topic. Here, it will be assessed whether online platform industries are likely to have a highly concentrated market structure, and if so, whether this is harmful to consumers’ well-being. Thereafter, a number of models in which topics such as data collection, privacy, online advertising, and online platform competition are analysed are reviewed. In the final section, it will be explained how the model that is developed in the next chapter contributes to existing research on this topic. Special attention is given to the manner in which this model includes certain dynamics that have not been analysed until now, but which could have important implications for competition policy in data-driven platform industries.

1 Theoretical Background

In this section, the most important concepts and theories that explain the dynamics of online platform markets are presented. It is assessed whether these dynamics are likely to lead to antitrust concerns by reviewing and weighing the arguments that have been put forward both in favour and against antitrust involvement in online platform markets.

Before the competitive implications of a data-driven economy can be discussed, ’big data’ needs to be defined. Therefore, the first section oﬀers a description of the characteristics of big data. Because the ’free’ services under investigation are commonly oﬀered in multi-sided markets, the next section will provide an explanation of the competitive dynamics observed in such markets according to economic theory. After this, it is assessed whether the combination of big data and multi-sided markets lead to industries that are characterised by high concentration. Then, it is evaluated whether online platforms with a dominant position have the ability and incentive to abuse said position. The final section concludes.

1.1 Big data

There is no single definition for big data (De Mauro et al., 2014). Here, the following definition will be used: big data has a combination of volume, variety, velocity, and value that implies that the data cannot be analysed by traditional tools (Schroek et al., 2012). Thus, big data refers to the large datasets that can be analysed to reveal patterns, trends, associations, and other information

(9)

that may provide firms with a competitive advantage (European Commission, 2016a). As becomes apparent from the definition used, big data is commonly characterised by ’four V’s’: the volume of data, its variety, the velocity of data collection, and the value of data. Each ’V’ will briefly be discussed below.

Volume Volume refers to the magnitude of data and is often regarded to be its primary attribute (Russom, 2011). Because datasets are deemed to be large enough to constitute big data when they are so large that they exceed the processing capacity of conventional database systems, the volume that is large enough for data to be labeled ’big data’ is relative (Wilder-James, 2012). Due to increasing storage capacities, what may be referred to as big data today may thus not be large enough to fit the definition in the future (Gandomi & Haider, 2014). The volume of data is ever-increasing and is likely to continue to grow; the amount of data doubles every two to three years (Mayer-Schonberger & Cukier, 2013). Part of the reason for this increase in data collection is the decrease in the costs of collecting, storing, processing, and analysing data (OECD, 2014). Another explanation can be found in the rise of internet access, smartphones, e-commerce, and social networks, all of which imply that individuals increasingly divulge personal information (Stucke & Grunes, 2016). With the rise of the ’Internet of Things’2_{, objects will collect even more data about}

individuals (Stucke & Grunes, 2016).

Variety Another important characteristic of big data is its variety, or the heterogeneity of a dataset (Gandomi & Haider, 2014). There are various categories of data, and data are collected on many diﬀerent types types of variables (Katal et al., 2013). This variety also makes it more diﬃcult to analyse data by use of traditional data processing technologies (European Commission, 2016a). Data’s value can increase by means of ’data fusion’, which is the process of aggregating various data sources such that new insights can be obtained from it (Stucke & Grunes, 2016).

Velocity The concept of velocity refers to the speed at which data are generated, as well as the speed at which it is required to be analysed and acted upon (Gandomi & Haider, 2014). The velocity at which data are generated, acquired, processed, and analysed has increased, sometimes to such an extent that it is approaching real time. This implies that the potential for big data to have an immediate impact on decisions is increasing (European Commission, 2016a). The velocity of big data also refers to data’s time value. Depending on the type of data, many sorts of stored information decrease in value over time (Stucke & Grunes, 2016).

Value The volume, variety, and velocity of data have increased because of big data’s value (OECD, 2014). The value of big data comes from its analysis. Big data would have less value if it could not be used to extract insights that can be used to understand, influence, or control the data objects of these

2_{The Internet of Things is the manner in which computers and sensors, but increasingly also other objects such as}

cars, cameras, and even fridges, which are connected to the internet interact with each other and process data (Stucke & Grunes, 2016).

(10)

insights (OECD, 2014). At the same time, machine learning relies on the use of a large volume and variety of data. Algorithms’ capacity learn increases with the amount of relevant data they process (EDPS, 2014). To summarise, big data’s value derives from the other three ’V’s’. A larger volume of data allows corporations to find correlations in large, unstructured datasets, possibly providing more useful insights than smaller, but cleaner datasets (Stucke & Grunes, 2016). Big data’s value is also positively related to its variety, as can be seen from the practice of data fusion. Finally, a firm that is able to be the first to collect, analyse, and store certain data may have an advantage over others by being able to respond to changes in real time, and younger data are generally more valuable, implying that data’s value also stems from its velocity (Grunes & Stucke, 2016).

1.2 Multi-sided markets

In multi-sided markets, two or more distinct groups of economic agents interact, and this interaction is facilitated by intermediaries or platforms (Armstrong, 2006). The existence of such intermediated markets stems from the presence of direct and indirect network effects. Direct network effects arise when the utility that a user of one side of the platform receives is a function of the number of other users of that side of the platform. Indirect network effects are present if those using one side of the platform impose an externality on users on the other side (Evans and Schmalensee, 2012). Network effects can be both positive or negative.

In case of (net) positive indirect network eﬀects, the benefit obtained on one side of the market will depend on how well the platform is able to attract users to the other sides (Armstrong, 2006). This then implies that there is a so-called “chicken-and-egg problem” in multi-sided markets. Namely, if users are not expected to “get on board” on one side of the market, there are no indirect network externalities for the other sides to be obtained and thus no incentives to join the platform (Rochet & Tirole, 2003).

Furthermore, the proportion of users on each side of the market is relevant. If one side of the market is too small relative to the other, a negative feedback loop may result, in which users that leave the ’overcrowded’ side of the platform make users on the too-empty side more likely to leave as well (European Commission, 2016a). The solution to these issues is to either subsidise one side of the market or to invest in one side to improve its quality, or both. Both solutions are intended to compel users to join by making the platform more attractive (Evans, 2003).

The fact that platforms are likely to cross-subsidise groups of users then implies that platform operators do not choose a price level for their service, but rather choose a price structure, taking into account the presence of any network effects. Since subsidising one side of the market means that the platform operator is able to obtain less of that side’s surplus, the choice of which side to subsidise is an important one. Given the presence of indirect network effects, it will then generally be optimal to subsidise the side that exerts the largest network externality on the other. Thus, platforms often treat one side of the market as a profit centre and the other as a loss leader (Rochet & Tirole, 2003). Three factors are of importance in the determination of the price structure offered to different groups of platform users (Armstrong, 2006).

(11)

The first factor is the relative size of indirect network externalities. As was indicated before, the group that exerts the largest positive network externality on others will tend to be subsidised. Furthermore, unless they have the effect of tipping the industry structure to monopoly, positive indirect network externalities have the effect of intensifying competition and reducing the profits obtained by platforms. This is the case because being able to compete on one side of the market requires sufficiently good performance on the other sides. This leads to downward pressure on prices on all sides of the market (Armstrong, 2006).

Second, platform operators can choose to charge a fixed fee for using the platform, but may also be able to charge per transaction. If the latter is feasible, this implies that the price of using one side of the platform is effectively a function of performance on the other sides (Armstrong, 2006). Per-transaction charges then reduce the impact of indirect network effects. This is the case because it becomes less important to get the other side ’on board’ when trying to attract one group of agents when these agents only have to pay for the benefit of interacting with other sides when such an interaction actually occurs (Armstrong, 2006). Since indirect network externalities were shown to put downward pressure on prices in a competitive environment, setting a per-transaction fee will have the effect of increasing platform profits. It has to be noted that, in case of a monopoly platform, there is no difference in outcomes between the two methods of payment (Armstrong, 2006).

Finally, the price structure set by a platform is aﬀected by whether users single-home or multi-home. If users single-home, this means that they only use one platform, whereas multi-homing implies that users may use several platforms for the same purpose at the same time. In platform markets with two sides, it may be the case that one side multi-homes, whereas the other single-homes. In this situation, competitive bottlenecks may exist. Namely, if an agent on the multi-homing side wants to interact with an agent on the single-homing side of the market, it can only do so through the platform chosen by the single-homing agent. This gives the operator of said platform monopoly power over access to agents on the single-homing side. This can in turn be expected to raise prices on the multi-homing side. Competition over the single-homing side can be expected to lead to low prices for agents on that side, possibly as low as zero (Armstrong, 2006).

The divide-and-conquer strategies typical in platform markets imply that the ex-post market structure is likely to be quite concentrated. However, competition for the market ensures that prices are disciplined somewhat if the prevailing market conditions are conducive to competition (Jullien, 2005). It is often the case that, due to network eﬀects, dominant platforms remain dominant, since users expect suﬃcient others to be present on all sides of the market and will thus be less reluctant to remain on or join the dominant platform. However, when confronted with a higher-quality competitor, even dominant platforms might lose market share (Halaburda, Jullien, & Yehezkel, 2016).

With regard to online platforms, the dynamics just discussed remain relevant. However, a number of additional characteristics are shared by most online platforms. This includes the ability to collect, process and use big data to improve the services oﬀered to users on all sides of the market. This ability to aggregate data gives the operators of online platforms an informational

(12)

advantage over separate users, thus causing information asymmetry. The data aggregation capacity of online platforms thus implies there are economies of scope present when platforms are able to obtain data of a large variety. Namely, the value of an aggregated dataset is larger than the sum of benefits of smaller datasets, because less correlations and complex patterns can be discovered from the latter (European Commission, 2016a; Martens, 2016). Furthermore, online platforms’ ability to facilitate transactions (as in more traditional multi-sided markets) also relies on the use of information technology (European Commission, 2016a). By using an online service provided by a platform, users provide platform operators with large amounts of data. This facilitates platforms’ ability to match all sides of the market they operate in and to improve their service (European Commission, 2016a).

To see that many online firms share the characteristics of multi-sided markets, Google’s search engine can serve as an example. Here, the search engine is the platform, which can be argued to have three sides. On one side, users query the engine for information. On another, content providers provide this information and compete for users’ attention. Content providers are indexed by Google based on their relevance3_{. On the third side of the market, advertisers finance the platform by}

paying each time a user clicks on an advertisement (Rieder & Sire, 2014). Here, it becomes apparent that Google solves the ’chicken-and-egg problem’ in the way proposed by Evans (2003). Namely, it subsidises both search engine users and content providers to get advertisers on board of its platform. Further, it invests in the advertising side by refining behavioural targeting through extensive data collection and by providing a large set of tools and services to advertisers. It also invests in the other two sides by continually improving its indexing and search technologies (Rieder & Sire, 2014).

1.3 Dominance

Many online platforms industries, such as the markets for online search, e-commerce, and social media, are characterised by high concentration. For instance, in the online-search market in Europe, Google Search handles approximately 90% of search queries (European Commission, 2016). The characteristics of big data, combined with the dynamics observed in multi-sided markets, may imply that dominance is indeed the most probable outcome in these industries. Many online firms’ business models use data as their main input (Grunes & Stucke, 2015b). Because of this, it is possible that data could provide platforms with an opportunity to obtain and sustain a competitive advantage by use of data-driven strategies. This also implies that it might be in firms’ interest to limit rivals’ access to the data they have accumulated and to engage in strategic acquisitions intended to obtain certain datasets (Grunes & Stucke, 2015b).

However, there is much debate in the academic literature regarding whether or not data can give rise to entry barriers and can thus be used to obtain a competitive advantage (Sokol and Comerford, 2016), or whether the dynamics of online platform markets should give rise to a highly concentrated market structure at all. In this section, it will therefore be discussed whether and how the outcome

3_{There may be other factors that influence the search ranking oﬀered by Google. The possible bias resulting from}

(13)

of firms’ business models in online platform markets is likely to be a dominant position for one or a small number of companies.

To this end, the section will commence with a description how network eﬀects and the charac-teristics of big data give rise to economies of scale and scope in online platform markets. Next, it will be assessed whether switching costs are high in these markets, as well as whether strategies exist that may be used by online platforms which may lead to lock-in. After that, a number of other exclusionary strategies in online industries will be discussed. Finally, some light will be shed on the ongoing debate of whether big data can be used to raise entry barriers in online industries.

1.3.1 Economies of Scale and Scope

As has been discussed earlier, the business models of many online platforms are based on collecting, analysing, and storing user data in order to improve the quality of the oﬀered service to users, while simultaneously providing advertisers with a means to target their advertisements to the most relevant audience. Big data can be used to resolve the matching challenge that results from the ever-increasing amount of users on all sides of the market and of the volume of data present online. Furthermore, having access to big data can aid firms in the optimisation of business processes, help them to identify trends, and makes it possible to improve self-learning algorithms (European Commission, 2016). It thus appears to be the case that data are an important productive input for online platforms that could lead to economic benefits. However, big data may give rise to economies of scale and scope in online platform markets, which could lead to high concentration, for a number of reasons that will be discussed below.

First, the marginal cost of distributing digital products has decreased with the ever-shrinking cost of computing power, storage capacity, and transmission bandwidth. However, the initial investments necessary to install this infrastructure and storage capacity are high (Hoofnagle & Whittington, 2013). Substantial spending on research and development is needed to develop algorithms and to improve machine learning. Furthermore, since the velocity of data is high, services have to be adjusted to respond to changes in user preferences in real-time and to constantly improve the quality of user and advertiser tools (Graef, 2015; European Commission, 2016). The costs of setting up such tools are typically fixed. For instance, in 2014 alone, Google has spent close to $11 billion on real estate purchases, production equipment, and data centre construction and has spent a further $10.5 billion on research and development, in order to be able to serve over a billion search queries each day (European Commission, 2016). Combined with low marginal costs4_{, this implies that economies}

of scale are created that may constitute an entry barrier for new or smaller firms (Graef, 2015). Secondly, in addition to economies of scale, platforms with access to a larger variety of data may also be able to sustain their dominant position through economies of scope. Economies of scope arise in online platform markets because the value of an aggregated dataset exceeds the the sum of values of the separate datasets it is made up of (European Commission, 2016). This is the case

4_{The cost of distributing a service to one additional user when all the aforementioned fixed costs have been incurred}

(14)

because more correlations and trends, and thus more insights can be obtained from a larger, more varied dataset.

Rieder (2014, 2015) argues that data analytics lead to concentric diversification, which is es-sentially the same as concentration through economies of scope. Namely, firms can expand data collection and analysis to adjacent data areas. In this way, their joint analysis of existing and ad-ditional data gives these firms an advantage over firms that analyse data in one separate area only (Martens, 2016).

One of the reasons why high concentration is likely to be observed in these markets is that the collection and analysis of data creates an information asymmetry, since a platform operator has more information than individual platform users on each side. This puts platforms in a dominant position, especially if the platform itself also oﬀers a service that competes with others on one side of the market, as is the case for Google’s AdSense (Martens, 2016). The informational advantage of platforms does not only stem from the collection of raw data, but mainly from the analysis of that data and the resulting insights that can be obtained. This implies that economies of scope in data collection and analysis (i.e. the fact that algorithms improve when more data are available) indeed play an important role in these markets (Martens, 2016). Additionally, achieving scale in data is important for providing users with relevant responses or recommendations when those users are looking for infrequently asked for information or products. For instance, scale is important for search answers since this gives them the ability to provide users with relevant responses to tail queries (Graef, 2015). These findings are supported by basic statistics: if data are randomly drawn from some distribution, the Central Limit Theorem implies that the accuracy of estimates of the distribution’s mean increases with the square root of the number of items in the database (Gal & Rubinfeld, 2016).

Further, in online platform markets, network eﬀects imply that increased participation on one side of the market makes it more attractive for customers on the other side(s) of the market to join a platform as well. This gives rise to two types of feedback loop, which may tip the market to monopoly, or in the least lead to a more concentrated market structure (Evans & Schmalensee, 2012).

The “user feedback loop” is the process that stems from the fact that those platforms that are currently serving most users are able to collect most data. Using these data, the platforms can improve their algorithms, which are ’self-learning’, meaning that algorithms can improve themselves by discerning patterns or trends in the data that would not be observable by humans. The im-proved algorithms can then be used to increase the quality of the platforms services by providing recommendations that are in close accordance with users’ preferences and by matching users on all sides of the market with increasing accuracy (European Commission, 2016). The improved quality of the platform’s services then attracts more customers to the platform, which in turn means that the platform can obtain even more data (European Commission, 2016).

Users of online platforms can further benefit from other users in a number of direct or indirect ways. For instance, on social media platforms such as Facebook, users benefit from each other’s

(15)

presence on a platform directly. However, users also indirectly obtain benefits through other users on their side of the market. This is the case because a platform with more users will attract more content providers to the platform. The presence of more other users will also provide the platform with higher advertising revenues that can be invested in improving the infrastructure that can be used to improve the platform’s service (European Commission, 2016). Users of online platforms are generally unaware that carrying out transactions on a specific platform enables it to improve the relevance of its results or recommendations for other users, thus giving rise to an indirect network externality (Argenton and Prüfer, 2012).

Similarly, the “monetisation feedback loop” refers to the idea that when a platform gains more users and their data, its ability to target and therefore sell advertisements improves. This means that it becomes easier to monetise the platform and obtain higher revenues, to be used to improve the quality of the service, thereby attracting yet more users (Sokol & Comerford, 2016).

An example can show how this doubly benefits the platform operator. The example again comes from Google. Google’s philosophy holds that, since users are likely to experience nuisance when shown too many ads, it is better to show a small number of relevant ads. Firms that want to advertise on Google search bid on search terms, or to be more precise, bid their willingness to pay per time a user clicks on their ad (Levy, 2009). All ad slots are auctioned oﬀ simultaneously. Advertisers’ may have an inclination to lower their bids in order to avoid the winner’s curse of paying much more than the advertiser just below them on the page. Google’s chief economist, Hal Varian, has therefore ensured the payment system resembles a second-price auction; the winner of each auction pays the amount of the second-highest oﬀer plus one cent, thus encouraging higher bids (Levy, 2009). However, not only the auction determines who will ultimately be listed on top of a result page. Namely, the ultimate outcome is also determined by the quality score. This is a metric that is intended to ascertain that the ads shown on each result page are relevant matches to what users query (Levy, 2009). This process is called ’query-results-ads matching’ (QRAM) (Rieder & Sire, 2014). In light of this example, it is easy to observe that a platform that can provide users with more relevant recommendations will be more attractive by limiting advertising nuisance and instead showing results that users are actually interested in. Furthermore, the platform will become more attractive to advertisers because of this, that are likely to place higher bids to obtain an ad slot. Namely, over and above the larger amount of users seeing and possibly clicking on their ad, the increased relevance of the advertisements and advertisers’ improved ability to engage in behavioural targeting implies that conversion rates5 _{are higher and that these firms will thus be willing to pay}

more to obtain an ad slot.

It should be noted that at some volume of collected data, returns to scale may start to diminish. Namely, after a large amount of data has been collected, the value of additional information will decline in the total amount of data that has been collected. The strength of the aforementioned benefits of data collection then depends on the volume of data at which the returns from additional data will begin to diminish (Graef, 2015). If diminishing returns only set in at a large volume of data,

(16)

having access to big data can give rise to a competitive advantage (Graef, 2015). Furthermore, due to product differentiation and the fact that users often multi-home, monopoly is not often observed, yet high concentration is (Evans & Schmalensee, 2012). For potential entrants, the aforementioned feedback loops make it difficult to attain the critical mass of users needed to get the platform ’off the ground’. Even if critical mass is reached, entrant platforms will still have a hard time competing with incumbent platforms that already have a large number of customers on all sides, because the network externalities are smaller when a lower number of transactions are made on the entrant platform (Evans & Schmalensee, 2012).

1.3.2 Switching Costs and Lock-in

Carl Shapiro and Hal Varian (1999) contend that switching costs and lock-in are common in infor-mation markets. First-mover advantages can be powerful and may be sustained for a long time in markets with lock-in, especially if scale economies are substantial. The fact that online platforms track users and use their data to adjust their services to users’ individual preferences implies that users become locked in to a service. This is the case because a platform that does not have a user’s data stored cannot provide the same quality of service to this user as an otherwise identical platform that does have these data because the user has already been using the platform. The user will then be less likely to switch to the former platform if she cannot easily take her data with her.

However, evidence of multi-homing in some online industries could mean that switching costs are not actually that high in online platform markets. Because of this, dominant platforms may utilise strategies that raise switching costs for those already using their platform, possibly leading to more pronounced lock-in eﬀects. A number of the strategies that are currently observed in online markets will be elaborated on below.

1.3.3 Exclusionary Strategies

The fact that online platforms’ business models are based on collecting and monetising user data may give those platforms that have access to large amounts of data both the ability and incentive to eliminate potential rivals. As this occurs, smaller rivals’ access to necessary data may be partially foreclosed or altogether eliminated, leading to a reduction in rivals’ incentives to innovate and compete with dominant firms. Here, a number of strategies that can be employed by dominant platforms to exclude competitors will be discussed.

Exclusive dealing Dominant platforms may sign exclusive dealing contracts in order to foreclose the possibly essential input that data forms for competitors. For instance, Doganoglu and Wright (2010) find that if network effects are present, whereas scale economies are not, it will be profitable for an incumbent platform to exclude a more efficient potential entrant by offering exclusive dealing contracts to one side of the market before entry would occur and simultaneously charging higher prices on the other side of the market.

(17)

Nowcasting A dominant firm with access to big data may utilise the velocity of data to identify potential competitive threats using a practice called “nowcasting”. Nowcasting refers to using data to discern trends well before others can, resulting in platforms’ ability to ’predict what is happening as it occurs’ (Stucke & Grunes, 2015a; 2016). This means that dominant platforms can spot competitive threats before they become visible to regulators and others. Without access to the same amount of data dominant platforms do, it may be quite diﬃcult for competition authorities to prove that competition would likely be harmed by exclusionary conduct or by a strategic acquisition (Stucke & Grunes, 2015a). Additionally, those entrants that are eliminated or acquired are likely to be the most disruptive innovators on the market, since dominant platforms perceive them as a threat. This then implies that a potential source of innovation is removed, with adverse consequences for competition (Sokol & Comerford, 2016).

Offering ’free’ services Offering ’free’ goods or services can also be an exclusionary strategy, because it will be difficult for a potential entrant to incur the large investments needed to set up an online platform when it is not certain whether the new platform will be able to recoup these costs (e.g. through advertising revenues) in the future (Evans & Schmalensee, 2012).

Increasing switching costs Online platforms can use a number of strategies to increase users’ switching costs and thus make it more likely for users of their service to get locked in.

First, most firms disable each other’s data collection software, making it more diﬃcult for users to change services because other providers can personalise their services to a lesser extent (Gal & Rubinfeld, 2016). Furthermore, platforms may contractually limit users’ data portability, thus limiting the user’s ability to export personal data from one application to another with reasonable ease. This is a way in which firms may generate lock-in eﬀects. For instance, Facebook has blocked a Google extension that would have made it possible for Facebook users to export their data to Google’s social media platform Google+ (Gal & Rubinfeld, 2016).

Online platforms may also increase the number of services they oﬀer, thus creating an ’ecosystem’ of functionalities rather than just oﬀering a simple service (European Commission, 2016). An exam-ple of this comes from Google, which has expanded its services from solely being a search engine to providing many free products and services, including Gmail, smartphone operation system Android, Google maps, Google Scholar, YouTube, and many more6_{. In this manner, Google is able to increase}

its user base and keep users locked-in by becoming an interconnected ’super platform’ (Martijn & Tokmetzis, 2016). Again, the widespread use of these services makes it possible for Google to collect more data on users and thus adjust those services to each user’s individual preferences, making the lock-in eﬀect more pronounced.

This ability of some online platforms to connect users in ecosystems of services is also a manner in which platforms drive demand between various sides of the market. For instance, if two platforms have developed mobile operating systems, but one of the two has a larger initial user base, this means

6_{On https://www.google.com/about/products/, the astounding number of products and services oﬀered by Google}

(18)

developers are more likely to design applications that run on the operating system of that firm. This then attracts more users to the platform, leading to a positive feedback loop resulting in growth that exceeds the linear growth that would result from traditional economies of scale (European Commission, 2016).

Another example of a business strategy that may serve to increase switching costs is used by social media platforms. In addition to the social networking services they provide, the largest social media platforms also oﬀer a service called social log-in. Through social log-in, individuals who have to register in order to use an online service have the option to authenticate themselves through their social media account with a single click. This is often far more convenient than entering one’s personal information manually. However, it also permits the social medium that is used for authentication to track the individual on this service, rather than only being able to track users on their own platform (European Commission, 2016). Furthermore, it makes it more attractive for individuals to use other tools that are connected to the social media platform. This can increase individuals’ switching costs, thus making it appear that the service, rather than being labelled social log-in, would be more appropriate to refer to as ’social lock-in’.

Bias in search rankings Finally, search engines may be able to exclude rivals by introducing bias in their search ranking. For instance, Google is regularly accused of manually downgrading the ranking of firms that compete with one of its services (Rieder and Sire, 2014). Some argue that content shown on Google is adapted so as to please advertisers, but in light of Google’s dominant position in both search and advertising, this appears to be unlikely. However, Google may treat various types of ’organic search’ links according to their potential for generating revenue. Google may, for example, rank links to other Google properties higher to get more traffic for its own services. Similarly, ’ad network’ links, which lead to websites serving ads through Google serving tools AdSense or DoubleClick may also be ranked higher to increase clicks and with it Google’s revenues. However, there is another reason to favour these links as well as ’ad-resistant’ links, which redirect the user to websites that do not advertise at all and are not likely to ever do so. Namely, if links to websites that increase revenues for Google itself, for firms that advertise using Google’s services, or for no one at all, firms that could advertise their product or service using Google’s products are pushed down in the search ranking and see their revenues decrease. This may then compel such firms to use AdWords or AdSense, increasing Google’s dominant position in ad serving (Rieder and Sire, 2014). Such a strategy thus makes it quite difficult to compete with Google for both other search engines and for other firms that offer advertising tools.

1.3.4 Can Big Data Form a Barrier to Entry?

The internet generates enormous amounts of data. The current-day preference for sharing and transparency, in combination with the fact that users appear to be willing to share their personal information in return for a small benefit increases the amount of data that can be collected. Because of this, enabled by technological progress, many firms have specialised themselves in collecting,

(19)

analysing, and storing data. Information on users’ preferences is useful for any firm, but as can be seen from the discussion above, it has special importance for online platforms, whose business model rely on the acquisition and subsequent monetisation of user data (Graef, 2015). Data can either be used as an input to production or it can be sold to other firms (Gal & Rubinfeld, 2016).

Raw data are generally not quite valuable when no analysis is performed to turn it into prescrip-tive or predicprescrip-tive knowledge. The information that can be obtained from raw data has increased through data science techniques that are continuously becoming more advanced. Natural-language processing, pattern recognition, machine learning and statistics can be used to mine information from the raw data, creating a virtuous circle between incentives to collect more data and advances in data synthesis and analysis (Gal & Rubinfeld, 2016).

There are various types of information that platforms can obtain from user data. A first type is referred to as volunteered data, which is data provided by users themselves. This includes profile information, friends lists and photos from social media, contacts that users have shared online, and search queries typed in by users online (Graef, 2015). A second category is labelled observed data. This is data created by platforms through their analysis of user behaviour, obtained by using cookies and trackers (Graef, 2015). A final type of data is called inferred data, also known as metadata. This is ’data on data’, derived from firms’ analysis of the first two categories of information (Graef, 2015).

The wealth of information available online creates a scarcity of attention, for which online plat-forms compete. The provision of free services in exchange for data is one of the strategies through which platforms attempt to obtain as much attention, and thus as much data as possible (Gal & Ru-binfeld, 2016). It can thus be argued that platforms’ business models have moved away from focusing on network eﬀects alone to a focus on data collection and processing abilities (Martens, 2016). In light of the arguments put forward above, the competitive strength of many online platforms could increasingly be determined by the volume and quality of data they hold.

However, there is no consensus yet regarding whether data could form a barrier to entry. Some argue that because data has become such an important input into decision-making and process optimisation, access to it has become an important strategic asset, possibly giving rise to a compet-itive advantage (Gal & Rubinfeld, 2016). Those who believe this to be true further argue that the value of data is evidenced by the number of data-driven acquisitions of data sources and the huge investments incurred in order to gather data and to improve machine-learning capabilities (Gal & Rubinfeld, 2016).

However, others argue that the economic characteristics of big data protect online markets against competitive harm. The question at the hart of ongoing debate in the academic literature on big data and its relation to antitrust is thus whether the eﬀort and cost that an entrant is required to incur in order to collect a suﬃcient volume and variety of data to be able to compete on equal footing with incumbent platforms is large enough to give rise to a barrier to entry in online markets (Graef, 2015).

(20)

can indeed give rise to barriers to entry. To this end, a number of arguments provided by those who believe that data cannot give rise to competition concerns will be elaborated on, as well as the counterarguments given by those on the other side of the debate, in order to arrive at a verdict regarding the competitive nature of big data.

Disruptive Innovators? First, those who do not believe that antitrust involvement is required in data-driven markets argue that these markets typically have low entry barriers. Evidence for this can be found from a number of innovative entrants that have rapidly emerged and that have displaced dominant firms with access to far greater datasets than themselves (Tucker & Welford, 2014). Most online services, it is argued, do not require much user data as a starting point if an innovative product is oﬀered that addresses consumer needs. Users that are attracted by the service and start using it will then provide the entrant with the data that is necessary for continued quality improvements (Tucker & Welford, 2014).

However, others argue that even if it is possible for today’s dominant online platforms to be replaced by a disruptive innovator, dominant firms can still stifle innovation in the meantime. Smaller firms are often not able to compete on equal footing with their larger counterparts since the former lack the volume of data the latter has access to. This means that the smaller firm will not be able to oﬀer a product of comparable quality. As the data and quality gaps widen, the competitive constraint imposed on the dominant firm will diminish, which may reduce incentives for quality improvement and innovation (Sokol & Comerford, 2016). Consumers should not suﬀer harm because eventually, the dominant incumbent will be pushed from its throne (Stucke and Grunes, 2015a).

Furthermore, the practice of ’nowcasting’ discussed above may imply that with present-day proficiency in data analysis, disruptive innovators may not ever get the chance to achieve market share in the first place.

The Facility of Obtaining Data The second argument proposed by opponents of antitrust involvement in online markets is that data is ubiquitous, low-cost, and easy to collect (Tucker, 2013). When launching a novel platform, data can easily be collected from consumers. Furthermore, both data and the necessary tools for storing and analysing the data are argued to be readily available from numerous third parties. The fact that data can be acquired from third parties then also implies that a platform operator can obtain insights into consumer preferences before any user has interacted with the platform (Sokol & Comerford, 2016).

In response to this argument, Stucke and Grunes (2015a) state that firms currently spend con-siderable amounts of money and eﬀort to collect and analyse data to sustain their data-induced competitive advantage. If personal data were indeed as ubiquitous and low-cost as implied by the opponents of antitrust involvement in online markets, firms would have little incentive to incur the large investments needed to oﬀer free services from which data can be collected and analysed. Such firms have an incentive to protect the exclusivity of certain datasets, and will undertake actions to ensure that independent data sources can not, in fact, be easily obtained through licensing, pur-chase, or collection (Stucke and Grunes, 2015b). For example, many online platforms use intellectual

(21)

property law and trade secret law to protect the data they have in their possession.

Graef (2015) states that the fact that data can be bought from and sold to third parties such as data brokers also gives platforms with access to greater volumes of data access to an additional possible revenue stream (Graef, 2015). Thus, even if firms do not use the data they possess to improve their services, they may still be better able to monetise their platform if they have a larger amount of data stored.

Big data’s non-rivalrous, non-excludable nature Third, it is argued that data share the characteristics of public goods. Big data’s alleged non-rivalrous and non-exclusive nature would set it apart from other essential inputs. Data are often claimed to be non-rivalrous because the use of particular information by one firm does not decrease its value for another firm. Data are also non-excludable since it is diﬃcult to assign property rights to a piece of information (Martens, 2016). It is further argued that incumbent online platforms do not have de facto or de jure exclusivity over the data they hold, that there are no exclusivity clauses in the terms and conditions that users have to agree on, and that there are no structures that lock users into sharing their data with a single platform (Sokol & Comerford, 2016). For this reason, some economists have argued that it is not possible for data to form an entry barrier or to be used in strategies that would sustain dominance, foreclose rivals, or limit competition in another manner (Sokol & Comerford, 2016; Lambert and Tucker, 2015).

On the other side of the debate, authors have contended that the fact that data is ex ante non-rivalrous does not change the fact that its collection, analysis, and storage essentially transforms it into a private good (Gal & Rubinfeld, 2016).

Additionally, although some easily obtainable data can indeed be said to be non-rivalrous and non-excludable, such as public data and data that is useful as an aggregate, some data are specific and not widely available (Gal & Rubinfeld, 2016; Graef, 2015). As was previously discussed, there are manners in which platforms can limit rivals’ access to the specific data they possess, for which it may be diﬃcult to find a reasonable substitute (Graef, 2015). For instance, Facebook does not allow third parties to scrape content from their platform. Similarly, Google restricts data portability for advertising campaigns and requires exclusivity agreements to be signed if firms want to place search advertising on their platform (Graef, 2015).

In response to the claim that platforms can and do not have exclusivity over their data, various authors have stated the fact that platforms can (and do) make use of various laws to protect the data they have collected.

Under the Database Directive, a database is defined as “a collection of independent works, data, or other materials arranged in a systematic or methodical way and individually accessible by electronic or other means” (Database Directive, 1996). Databases enjoy dual protection according to this Directive. First, the structure of original databases is subject to copyright protection under Article 3(1) if the ’databases which, by reason of the selection or arrangement of their contents, constitute the author’s own intellectual creation. Second, the database as a whole may be protected by the sui generis database right created by Article 7(1) of the Directive if ’there has been qualitatively

(22)

and/or quantitatively a substantial investment in either the obtaining, verification or presentation of the contents’ (Database Directive, 1996). Since online platforms incur significant costs to obtain big data, their databases are likely to qualify for such protection (Graef, 2015).

Furthermore, data qualifies for trade secret protection under Article 39(2) of the Trade-Related Aspects of Intellectual Property Rights (TRIPS) agreement if three conditions are met. First, the information should be secret in the sense that it is not generally known or readily accessible to persons in the circles that normally deal with the type of information in question. Secondly, the information must have commercial value because of its secrecy. Finally, the rightful holder of the information must have taken reasonable steps to keep the information secret (World Trade Organisation, 1994). This then implies that if online platforms have signed non-disclosure or confidentiality agreements, their datasets may qualify for trade secret protection (Graef, 2015).

The fact that platforms can and do protect their data by making use if these laws must then imply that not all data are as low in value or as easily accessible as is argued by opponents of antitrust involvement in big data. Indeed, when access to data is exclusive and the collection of data is subject to economies of scale and/or network eﬀects, control over high volumes of data can give rise to a competitive advantage (Graef, 2015).

In addition to this, it is argued that if data were as low in value as claimed, there would be little explanation for platforms’ reluctance to give users the option to prevent collection of their personal data, as well as their lack of will to enable data portability (Gal & Rubinfeld, 2016).

It is also possible that temporal barriers to data collection arise when one firm has unique knowledge at a crucial point in time. When data are unique and can only be obtained through unique access points, situations may arise in which the data are not as easy to replicate. Such unique access points to data may be created as a goal in itself, or may derive from an otherwise-focused activity (Gal & Rubinfeld, 2016). Usually, platforms obtain user data through other productive activities. The data that results is diﬃcult to replicate by competitors that do not engage in this activity. This means that a two-level entry barrier can arise if potential competitors are not willing to perform the respective activity through which the data is collected (Gal & Rubinfeld, 2016). Even if rivals set out to engage in those activities that allow larger platforms to sustain their dominant position, is possible that entry barriers arise from unique, prohibitively expensive gateways for data collection (Gal & Rubinfeld, 2016). For instance, Google has provided internet connections to hard-to-reach areas in developing countries (Martijn & Tokmetzis, 2016). In this way, the platform can collect data on users that no other firm has access to.

Thus, dominant platforms may be able to exclude others from accessing essential data and thus sustain their competitive advantage (Graef, 2015). If the data is unique, diﬃcult to replicate, or data portability is limited, this will increase the strength of this entry barrier (Gal & Rubinfeld, 2016).

The velocity of data and its transience in value A fourth argument that has been put forth is the fact that data’s value is often short-lived. Datasets have to be updated frequently in order to serve and oﬀer advertising to users with changing preferences. The high velocity of such

(23)

data implies that control over it does not give rise to a durable entry barrier because its value considerably decreases over time, implying that any competitive advantage provided by access to data is temporary. (Graef, 2015). This means that entrants are unlikely to be disadvantaged to a large extent relative to incumbents with respect to data collection and analysis (Chiou & Tucker, 2014). Firms seeking to enter the market do not need to obtain a dataset equivalent in volume to the incumbent’s, but should rather design a strategy to obtain highly relevant and timely data (Shepp and Wambach, 2015).

Proponents of antitrust involvement in data markets contend that it is indeed the case that most data have only a short-lived relevance, implying that their value is transient (Graef, 2015). However, the transient nature of data also implies that indirect network eﬀects are strengthened. Namely, even if potential rivals are able to acquire relevant data from third parties such as data brokers, online platforms with an established user base will obtain ’fresh’ data automatically. Access to real-time data enables platforms to quickly adapt to changing conditions, thus making their service more attractive (Graef, 2015). At the same time, smaller platforms or entrants, which do not have a stable returning user base, will have to acquire data each time there is a need to update their database or have their algorithm take into account new developments (Graef, 2015).

Algorithms Finally, Sokol and Comerford (2016) argue that big data in itself does not provide much value. Therefore, being in the possession of a large volume of data does not secure competitive success. What is far more important is designing a superior algorithm, having engineering talent, oﬀering a high quality of service, speed of innovation, and attention to the needs of users (Sokol & Comerford, 2016).

Indeed, algorithms are used to return relevant results to queries by search engines, for product recommendations in e-commerce, and to show users content that matches their preferences on social media platforms. However, it is also true that stored data can be used to improve these algorithms’ proficiency in completing such tasks. This implies that even though having a well-functioning algo-rithm is an important condition for competitive success in online markets, access to greater volumes of data is also important for the extent to which algorithms can improve the services oﬀered by online platforms (Graef, 2015). For instance, to quote Tim O’Reilly, a chief Google scientist : “We don’t have better algorithms than anyone else. We just have more data” (Asay, 2010).

1.3.5 Are Online Platform Markets Likely to Lead to Dominance?

In conclusion, there are multiple reasons to believe that online platform markets are likely to lead to dominant positions by a small number of firms. Economies of scale and scope can raise barriers to entry if diseconomies set in at a large-enough scale of data. Furthermore, consumers may get locked-in to a service once it has reached domlocked-inance. Even if switchlocked-ing costs are low locked-initially, domlocked-inant platforms may use various strategies to raise them. A number of other exclusionary strategies can be used to maintain a dominant position as well. Finally, after reviewing the various arguments put forward in the debate whether data can form an entry barrier, it appears that those who think

(24)

they do are in the right. Thus, high concentration is a likely result of the dynamics of online platform markets. However, dominance is not a problem in itself. Only if high concentration will lead to consumer harm can there be said to be an abuse of dominance and thus a need for antitrust involvement. This subject is analysed next.

1.4 Consumer Harm

Online platforms oﬀer many benefits for consumers. Many of the services oﬀered online increase consumer convenience. Through rating systems and comparison tools, the information asymmetry between consumers and sellers of products is reduced. By buying goods online, consumers have more choice between various goods and sellers, and product information is easier to access (European Commission, 2016).

However, online platforms have also raised a number of concerns for consumers, most of which are related to the enormous amount of data dominant platforms possess and the purposes for which those data might be used (European Commission, 2016). Online platforms may collect, analyse, and store data related to order history, the users’ IP address7_{, browser type, the full clickstream}

from and to the platform website, the date and time of each click, product views, searches, location information, and session information with regard to users’ online behaviour, such as the length of time each page was visited, information about page interaction such as scrolling, clicks, and mouse-overs (and the speed or duration thereof), and the methods that were used to navigate away from the page (European Commission, 2016). As was discussed in the previous section, platforms may use the data they collect to improve the attractiveness of their service. However, platforms may also use individuals’ personal information in inappropriate ways (European Commission, 2016). In this section, these problems arising from online platforms’ data-driven strategies will be discussed, and it will be assessed whether any of them could amount to an abuse of dominance.

One way in which platforms may be able to abuse their market power is by using the data they possess to influence consumer choice through persuasive technology, which will be discussed first. After this, it will be addressed how the ways in which platforms use data may undermine the solidarity principle most types of insurance are based on. Platforms may also lead to an aggravation of discrimination. Possibly the most well-known concern that arises in relation to platforms’ use of big data is that it could lead to insuﬃcient protection of consumers’ privacy and the security of their data. This subject will be discussed in detail as well. Finally, it will be assessed in light of this discussion whether online platforms are likely to have the ability and incentive to abuse their dominant position.

7_{An IP address is an identifier of an (online) device. Thus, if one visits a website that is tracking users, what will}

be stored is this unique identifying number that makes the website operator able to recognise the computer, tablet, or smartphone the website was visited from when the user returns.

(25)

1.4.1 Persuasive Technology and Choice

It should be clear at this point that by using big data, platforms can analyse human behaviour in far greater detail than has ever been possible before. The young field of social physics uses new methods for data analysis, implying that human behaviour will not only be perfectly measurable, but also predictable and steerable by those who possess suﬃcient amounts of data (Martijn & Tokmetzis, 2016). This means, among other things, that platforms with access to large amounts of data may be able to influence human behaviour through what is commonly referred to as persuasive technology. Persuasive technology uses insights from psychology, behavioural sciences, and human-computer interactions to design systems that make users change their behaviour (Fogg, 2002). Using big data, it is possible to tailor advertising to the individual and to adapt strategies in real-time to changes in behaviour. Persuasion profiles can be constructed, allowing product recommendations and advertising to be tailored to these profiles in such a way that individuals are most likely to click on an advertisement or buy a product. This practice is also known as behavioural targeting and is widely used in online services (Martijn & Tokmetzis, 2016).

Persuasive technology is continually being improved through a ’cyber-kinetic loop’ between the physical and the digital world. Processes in the physical world are measured and the resulting datasets are analysed to optimise business processes. The resulting actions taken by firms can then be adjusted in real-time to new insights. The eﬀects of these actions can in turn be measured and analysed to indicate how successful a strategy has been in influencing each individual’s behaviour and to adjust strategies accordingly (Kool et al., 2017). ’Nudging’ by using persuasive technology is thus becoming increasingly refined; personalised behavioural strategies are used to persuade users, utilising the specific kind of argumentation past behaviour has shown them to be sensitive to (Kaptein et al., 2015). The data collected by platform operators may then be used selectively, in order to change the behaviour of one or more sides of the platform (Martens, 2016).

It is possible that firms’ analysis of aggregate user data can help them to understand a market better and thus to adapt their products to consumer preferences (European Commission, 2016). Individual data are used for consumer profiling, which has the potential to reduce search costs and provide benefits to all parties involved. When many (complex) options are available, consumers may experience choice overload (Ezrachi & Stucke, 2015b). The paradox of choice holds that increasing the number of options beyond a certain threshold can reduce consumer utility (Schwartz, 2004). This would then imply that oﬀering less, better-targeted options could increase consumer utility and therefore increase demand for searching for and buying products through a platform that successfully does this.

However, it is also possible that by means of persuasive technology, marketing and advertising strategies are adjusted rather than products. If this is the case, the products on oﬀer do not match the relevant user’s preferences better, but he or she will be more likely to buy the product. Thus, profiling users may not only generate benefits, but could also engender substantial costs. These costs may stem from behavioural biases that can be induced by such profiling, nudging users into decisions that they might regret (Martens, 2016). To be more precise, data-driven behavioural targeting could

Knowledge is (market) power? : analysis of the effects of 'free' goods in online platforms on competition