The future of anticompetitive self-preferencing: analysis of hypernudging by voice assistants under article 102 TFEU

ABSTRACT With the nascent rise of the voice intelligence industry, consumer engagement is evolving. The expected shift from navigating digital environments by a “click” of a mouse or a “touch” of a screen to “voice commands” has set digital platforms for a race to become leaders in voice-based services. The Commission's inquiry into the consumer IoT sector revealed that the development of the market for general-purpose voice assistants is spearheaded by a handful of big technology companies, highlighting the concerns over the contestability and growing concentration in these markets. This contribution posits that voice assistants are uniquely positioned to engage in dynamically personalized steering – hypernudging – of consumers toward market outcomes. It examines hypernudging by voice assistants through the lens of abuse of dominance prohibition enshrined in article 102 TFEU, showcasing that advanced user influencing, such as hypernudging, could become a vehicle for engaging in a more subtle anticompetitive self-preferencing.


Introduction
Voice assistants (hereinafter: "VAs") are becoming a ubiquitous feature of modern life. Integrated into smart home devices, wearables, vehicles, computers, and smartphones, they offer support for mundane everyday tasks while continuously and silently analyzing their owners' characteristics, habits and emotions. The consumer Internet of Things (IoT) sector has recently come under closer regulatory scrutiny in Europe.
The European Commission's (hereinafter: "the Commission") inquiry into the sector highlighted several concerns related to the development and competitiveness of consumer IoT and the market for general purpose VAs, specifically: restrictions on multi-homing, concerns about default settings and pre-installations on VAs, data accumulation and lack of interoperability. 1 It also showcased that the development of the voice intelligence industry is spearheaded by big technology companies such as Amazon (Alexa), Google (Home Assistant), and Apple (Siri). 2 This contribution posits that VAs by leading providers are uniquely positioned to engage in dynamically personalized steeringhypernudgingof users towards specific market and non-market outcomes and thus seamlessly influence and shape their preferences. 3 Importantly, hypernudging should not be viewed as a single behaviourally informed intervention or design element deployed to steer the user. Instead, it represents multiple interventions and elements delivered within the context of complex systems that may not be indicative of harmful effects on their own. 4 The scope of this article is limited to examining hypernudging by VAs in an economic activity context, namely, VAs providing information (and recommendations) about consumers' purchasing decisions and helping them execute pre-determined commercial tasks, such as renewing household items orders. Positioning recommendations by VAs within the hypernudging framework provides a new lens for studying their potential influence on consumers' personal spaces and aggregate effects on the market. When hypernudging is used to protect and/or expand firms' market power to the detriment of consumers, it is a cause for closer regulatory scrutiny. In the EU, competition law rules are applied to curb the negative manifestations of market power to safeguard inter alia consumer welfare and the system of undistorted competition. 5 Article 102 TFEU provision deals with sanctioning dominant undertakings that abuse their market power in a specific relevant market. The Treaty does not contain an exhaustive list of abuses or an explicit definition of abuse. Instead, the concept of abuse develops through the case law of the European Courts. 6 This contribution examines hypernudging by VAs vis-à-vis the selfpreferencing form of abuse, which has recently been clarified in the Google Shopping judgement of the General Court (hereinafter: "the Court"). 7 It deviates from existing competition law literature which examines personalized business practices through exploitative abuse lens by showcasing that hypernudging could also lead to exclusionary effects on the market. 8 Exclusionary effects reference a dominant firm engaging in a conduct which artificially raises barriers to entry and expansion, limiting consumer choice and quality, and ultimately raising prices for end consumers.
Against the backdrop of the Digital Markets Act (hereinafter: "the DMA"), which contains prohibitions against self-preferencing behaviour by VAs designated as gatekeepers, this article provides a justification for not overlooking European competition law as a relevant instrument in sanctioning anticompetitive next-generation consumer influencing practices such as hypernudging. Ultimately, it is set to answer the main research question: under which circumstances can hypernudging by VAs be considered a vehicle for platforms to engage in self-preferencing behaviour, and could such self-preferencing fall under the scope of article 102 TFEU?
It is important to note from the onset, this article does not posit that hypernudging processes should be labelled as a specific form of abuse, or an inherently problematic form of self-preferencing. Instead, it is assessed as potential means for anticompetitive self-preferencing to take place. After all, European competition law does not offer a one-size-fits-all solution to various forms of hypernudging and may trigger considerations under different theories of harm or fall within the scope of legitimate business strategies.
The article proceeds as follows. The first section will set out the stateof-the-art of the market for VAs while contextualizing its development concerning the features of dynamic digital markets. After establishing that leading VAs' providers possess market power, it will describe European competition law developments in digital markets to highlight the momentum created for sector-specific regulation. The second section will demonstrate why VAs are so well-positioned to hypernudge consumers and, in turn, shape their preferences and behaviour and the circumstances in which such conduct may threaten the functioning of competitive markets. Finally, the third section will closely examine hypernudging by VAs vis-à-vis European competition law's theory of harm of anticompetitive self-preferencing. It will do so by deconstructing its elements and development through case law, concluding with asserting European competition law's relevance in addressing potential market threats of hypernudging.

The rise of market power in the voice intelligence industry
The voice intelligence industry is at its nascent stages. 9 The development of the market for general-purpose VAs is led by a handful of big technology companies that shape consumers' experiences and interactions online. 10 The recent Commission's inquiry into the consumer IoT sector highlighted several concerns over the concentration of market power and, in turn, potential threats to a competitive process. This section will provide a comprehensive overview of the market for general-purpose VAs. In light of the presence of substantial market power, it will examine European competition law and the DMA as appropriate legal regimes to address its potential negative manifestations, ultimately justifying this article's focus on article 102 TFEU perspective.

The market for general purpose voice assistants
VAs are "voice-activated pieces of software that can perform various tasks, acting both as a platform for voice applications and a user interface." 11 They represent a sub-set of virtual assistants that use voice as input. 12 This contribution focuses on general-purpose VAs that enable access to a broad range of functionalities in response to users' voice commands, such as providing recommendations, controlling smart home devices, and executing daily tasks. 13 While primarily associated with smart home speakers, VAs are integrated into an increasing number of smart applications and devices, including smart home appliances, wearables, connected vehicles, and smartphone applications. 14 The market for general-purpose VAs has an oligopolistic competition structure, with a handful of big technology companies competing for the market. 15 The Commission's inquiry into the consumer IoT sector showed that in Europe, Amazon (Alexa), Google (Home Assistant), and Apple (Siri) are leading players in the development of the voice intelligence industry. 16 Fierce competition among leading providers is wellillustrated by the smart speakers' market developments. In 2018, Google and Amazon were engaging in price wars to plant their respective products at consumers' houses. 17 Companies recognize that the stakes for entrenching themselves as consumers' go-to IoT brands extend beyond VAs' market: winning platforms are likely to control a significant user interface (UI), with VAs becoming a likely gateway of consumer e-commerce and, ambitiously, world wide web experiences. 18 The movement towards voice-based services could be understood in the context of UI shifts comparable to the web and smartphones. 19 Each of these shifts has changed the way people interact with and access digital content: the web gave us a "click" enabled by the computer 11  mouse, smartphones introduced "touch" and "swipe," while voice further simplifies users' interactions by allowing them to "speak." 20 The development of commerce is, too, mirrored in these UI shifts. Just as e-commerce and mobile commerce became ubiquitous with the adoption of web and mobile applications, respectively, voice commerce is expected to follow this trend, despite the slow uptake. 21 According to the Commission's findings, the projected use of VAs worldwide will double between 2020-2024, from 4.2 billion to 8.4 billion, with 11% of EU citizens surveyed in 2020 already using VAs. 22 Voice and text-assisted AI are increasingly utilized in customer services, product information, marketing, and sales support. 23 The covid-19 pandemic further accelerated virtuallyassisted, staff-free shopping experiences. 24 However, the adoption of VAs in consumers' customer journeys is yet to mature, with consumers currently focusing on purchasing small and quick items that do not require visualization. 25 The success of the leading VAs is reinforced by the dynamics of the respective platform ecosystems they operate inthe utility of the service to the users is shaped and determined by those ecosystems. 26 Platform ecosystems consist of two key elementsa platform and its complementary applications. 27 Here, a software-based product for voice assistants serves as a foundation on which outside parties, such as smart home device producers or software developers, can build complementary goods and services around. 28 29 By the same token, it is compatible with an increasing number of third-party hardware with a "works with Amazon Alexa" label, including brands such as Sonos, Marshall, Bose, and Libraton Zipp 2. 30 In addition, "Alexa Skills Kit" is a software development framework that allows developers to create skillsvoice activated applicationsfor Amazon's VA. 31 The way these platform ecosystems connect and integrate with thirdparty consumer IoT products and services depend on their design. For instance, when setting up their smart home environments that can be controlled with the help of VAs, consumers have to choose how they will bring the different devices together; a logical starting point is choosing a VA. A distinction can be made between voice assistants operating as part of a more open ecosystem, such as the described Amazon's Alexa or Google Assistant, and more closed ecosystems, such as Apple's HomeKit controlled by Siri. 32 Whichever ecosystem they choose, consumers are likely to be locked in to build their environments based on compatibility with that ecosystem. Even in cases of multi-homing, specific areas of a consumer's life, such as a smart home or commuting, can be compartmentalized in a way that requires building those environments considering the compatibility of devices.
The Commission's inquiry into the consumer IoT sector highlighted several concerns stemming from the market power dynamics within the general-purpose VAs market. One such concern relates to the lack of standardization in the industry, which exacerbates the dependencies upon the leading VAs providers and further reinforces consumer economics literature, ecosystems can be viewed as "multi-actor groups of collaborating complementors (i.e., "theory of the firm" alternatives to vertical integration or supply-chain arrangements) and multiproduct bundles offered to customers (i.e., horizontally or diagonally connected goods and services that are "packaged" together), focused on customer ease-and lock-in. lock-in effects while stifling potential competition. 33 Currently, thirdparty consumer IoT services providers seem to cater to their service offerings and future business strategies, focusing on leading providers' standards. 34 Furthermore, the identified competition features in the general-purpose VAs market corroborate that these leading companies are expanding and shielding their market power by strategically using their application interfaces, algorithms, and contractual restrictions to ensure interconnectivity and interoperability for final consumers. 35 The market is characterized by high barriers to entry and expansion, with most data being collected and held by leading companies, pointing to the need to oversee the developments driven by firms with substantial market power. 36 2.2. Addressing market power in digital markets: from article 102 TFEU to sector-specific regulation The observed market dynamics in the consumer IoT sector led European policymakers to bring forward regulatory initiatives aimed to address some areas of concern, with focus being placed on removing barriers to entry and expansion. 37 This article posits that in addition to the emergent regulation, competition law provides a logical legal avenue in curbing the negative manifestations of market power in the context of general-purpose VAs. Hypernudging by VAs could be assessed as a way to engage in exclusionary self-preferencing behaviour sanctioned under Article 102 TFEU. By following the relevant developments in the competition law enforcement in digital sector, this section sets out the background necessary for further competition law analysis. 2.2.1. Abuse of dominance Abuse of dominance prohibition is drafted in a broad and abstract manner, leaving its interpretation and defining of specific concepts up to enforcement and judicial bodies. When building an abuse of dominance case, the first step necessitates establishing that an undertaking in question does hold a dominant position. To do so, it is necessary to define the relevant market, to delineate "the boundaries of competition between firms." 38 Defining the relevant market entails considering its product and geographic dimensions, which are determined by assessing demand substitution, supply substitution, and potential competition using economic tests. In practice, both the Commission and the Courts tend to define the relevant market in narrow terms. Once the relevant market is established, the competitive constraints and undertaking's position in that market are examined. The assessment necessitates considering the undertaking's market shares and other economic factors, such as performance indicators, price levels, profits, and barriers to entry and expansion. 39 Determining that an undertaking holds a dominant position is insufficient to trigger Article 102 TFEU intervention. Firms that gain strong market positions due to rigorous competition and innovation should not be penalized for their success. It is only when undertakings abuse their dominant position by engaging in behaviour that deviates from competition on merits that they ought to be sanctioned. 40 The Treaty does not contain an exhaustive list of abuses. 41 However, the literature and case law identify three broad categories: exclusionary abuses that exclude competitors from the relevant market, exploitative abuses that are harmful to consumers directly, and discriminatory abuses that apply dissimilar conditions to equivalent transactions between various 38 Commission Notice on the definition of relevant market for purposes of Community competition law ( customers. The subject of this article is self-preferencing, which the Court construed as an independent type of exclusionary abuse. Finding article 102 TFEU infringement necessitates establishing a logically consistent theory of harm, which must articulate how the dominant undertaking's behaviour harms competition and consumers. It is done relative to a counterfactual scenario, not deviating from the various parties' available empirical evidence or incentives. 42 The standard for establishing anticompetitive effects that leads to an infringement of Article 102 TFEU is, nevertheless, disputed. A distinction could be made between "capability and likelihood of anticompetitive effects taking place." 43 The former relates to a situation where a credible mechanism through which anticompetitive effects would manifest is identified. 44 The latter refers to conduct more likely than not to lead to an anticompetitive outcomea higher threshold for enforcement authorities to meet. 45 There is generally no requirement to show that conduct has actual effects on competition. 46 Instead, there must be a probability of anticompetitive effects taking place, albeit those effects cannot be purely hypothetical. 47 The Court also stressed that the Commission is required to analyze all the relevant circumstances in the case. 48 Absent objective justification for dominant undertaking's behaviour, establishing a credible theory of harm would lead to article 102 TFEU infringement.

Competition in the digital economy
It is a truism that the rise of the digital economy introduced challenges in applying competition law tools in abuse of dominance cases, which have been amply criticized for not being able to fully capture the power dynamics that play out in digital markets. 49 The fluidity of the boundaries of market power and, in turn, dominance, are highlighted in the context of big technology companies that form intricate multi-product and multi-actor ecosystems. These firms possess substantive market power in their respective core platform service markets, characterized by high market shares, network effects, data gathering and analysis capabilities, economies of scale and scope. 50 This market power is further reinforced by the interrelationships between market actors, with big technology companies creating organizational dependencies among their network of partners. 51 Although the characteristics of big technology companies are not fundamentally new, their power seems to have greater pervasiveness, scope, precision, and invasiveness in modern societies and individual lives, with effects spilling beyond the market into the social, political, and personal domains. 52 Over the past decade, defining dominance in digital markets has been a subject of intense debate. 53   sided nature of digital platforms and their structural roles within interconnected platform ecosystems, characterized by vertical integration, cross-sectorization, and private modes of governance. 54 To adequately address emerging challenges, more understanding is required regarding how the different business areas of these complex systems interact. 55 In digital abuse of dominance cases, the Commission and the Courts have continuously resorted to defining markets narrowly. 56 Illustrative is Google Android decision, which dealt with multi-product tying abuse and was largely confirmed by the Court. 57 In Google Android, the Commission concluded that non-licensable operating systems (hereinafter: "OS") do not belong to the same market as licensable ones. Consequently, Google's dominant position in licensable OSs was considered not to be meaningfully affected by the competitive constraint exerted by Apple or BlackBerry. 58 In reaching the decision, the Commission relied on non-conventional market indicators including small but significant non-transitory decrease in quality (hereinafter: "SSNDQ") test to examine the reaction of users and app developers to a hypothetical deterioration in quality of Android, as well as assessed user loyalty and switching costs, highlighting the limitations of traditional market definition toolkit in the digital sphere. 59 By the same token, Android app stores were considered to form a separate relevant market from Apple's App Store, showcasing that OSs and app stores were assessed as part of the same system. 60 Further demonstrating the salience of the topic, in March 2020, the Commission launched the evaluation of the Market Definition Notice and gathered evidence from stakeholders, with findings indicating the need for updating the Notice to reflect the realities of digital markets. 61 Similar hurdles are expected to emerge in defining dominance in the general-purpose VAs' market which, as said before, is currently dominated by three players: Amazon Alexa, Google Home Assistant and Apple Siri. It is noteworthy that establishing each company's respective market share is challenging given that VAs are integrated and preinstalled in a number of free services and devices, which lack reliable statistics. 62 Since this contribution is focused on the analysis of anticompetitive self-preferencing, it will operate under an assumption that dominance would be established either because the market tipped to favour a single firm or general-purpose VAs evolved in a way that allows for sufficient differentiation to constitute separate relevant markets. 63 The "system perspective" identified in Google Android decision may be informative in this regard since VAs' development is generally closely dependent on the platform ecosystem it belongs to. 64 Before examining what hypernudging by VAs entails and whether it could fall within the scope of article 102 TFEU, specifically regarding self-preferencing, it is necessary to acknowledge that digitalization of markets has also led to a surge of abuse of dominance investigations that tested the boundaries of existing theories of harm. companies have been at the centre of enforcers' radar. While building abuse of dominance cases proved to be a lengthy endeavour, the growing knowledge and enforcement experiences showed that the digital market dynamics and structure is prone to systemic concerns, instead of one-off competition law infringements. Therefore, the momentum has been created for sector-specific regulation, with the DMA adopted in July 2022. 65

The digital markets act
The DMA is a regulation that aims to foster fairness and contestability in the digital sector. 66 It identifies black-listed and grey-listed practices for companies designated as gatekeeperslarge online platforms with entrenched and durable positions that significantly impact the internal market and provide core platform services on which other business users and end users depend. 67 The final DMA text includes VAs within the definition of "core platform services." 68 Consequently, problematic self-preferencing practices that manifest through VAs technology may be, to a large extent, addressed by this regulation. Article 6(5) outlines an obligation to: "not treat more favourably in ranking and related indexing and crawling, services and products offered by the gatekeeper itself than similar services or products of a third party. The gatekeeper shall apply transparent, fair and nondiscriminatory conditions to such ranking." The outlined obligation in effect amounts to a per se prohibition on selfpreferencing. 69 While this contribution is not aimed to address the lively debate on the interaction between competition law and the DMA, 70 it is necessary to justify the choice for assessing dynamically personalized . Make a reference to Vestager's speech on the need to address systemic concerns and the DMA. 66 Ibid, Article 1(1). 67 Ibid, Article 3. 68 The text includes "virtual assistants", a broader term that also covers VAs. Ibid, Article 2(2)(h) and 2(12). 69  consumer steeringhypernudgingby VAs as a potential article 102 TFEU infringement. In light of the emerging voice intelligence industry, including VAs under the scope of the DMA is a forward-looking choice. However, the application and impact of promising provisions are expected to be heavily litigated. 71 As will be discussed in section 4, hypernudging by VAs is an elevation of existing forms of self-preferencing behaviour already familiar to regulators and enforcement authorities. In the context of intricately connected multi-product and multi-actor ecosystems, one could envision the next-generation of self-preferencing to manifest in more covert ways. Hypernudging has the potential to elevate the practice of self-preferencing, where instead of steering consumers' behaviour by ranking recommendations for a specific product uniformly and overtly across the different consumer groups, it moves towards presenting multiple offers, at different times, perhaps through various channels within the respective platform ecosystem the VA belongs to. 72 In such a scenario, individual recommendations and actions may not be indicative of harmful behaviour. 73 Even though the drafting of article 6(5) is wide enough to capture such next-generation self-preferencing behaviour, in practice, the challenge would emerge in pinpointing the specific features that lead to it. With the lack of observability of complex proprietary systems that are expected to facilitate such hypernudging, competition law may prove to be a logical instrument to deal with anticompetitive effects ex-post. Thus, even though the DMA may capture a great deal of harmful self-preferencing practices in digital markets, with the evolution of digital technologies and new ways developed to reach anticompetitive outcomes, the boundaries of article 102 TFEU may be tested further. As discussed in section 4.2, the Court's approach indicates a degree of malleability to it.

Hypernudging by voice assistants as a threat to competitive markets
This section will examine why VAs are uniquely positioned to engage in dynamically personalized user steeringhypernudgingprocesses, 71  which can be used to shape consumers' preferences and in turn market behaviour. They are designed to address two essential needs that consumers have when shopping onlineconvenience and trust. 74 When hypernudging by VAs allows for seamless consumer steering towards outcomes that do not fully align with their interests, the concerns over distortion of competition occur. 75

The mechanics of hypernudging by voice assistants
The premise of this article is that VAs of the leading providers are uniquely positioned to engage in a highly dynamically personalized user steeringhypernudgingtowards specific market outcomes, such as purchasing decisions, and to seamlessly shape their preferences. Theoretically, hypernudging is built on the insights of linkages between behavioural economics and information systems (IS) literature, which demonstrates that people's behaviour is influenced by their environmental and cognitive constraints. 76 Thus, their behaviour may be shaped by external actorsthe choice architectsthat can re-assemble their choice environments based on their specific context and circumstances, such as personal characteristics. 77 Hypernudging processes could be visualized as a staircase: "it is no longer about a single step placed by the choice architect to steer the user, but multiple steps that might come in different shapes, at different times, all with the goal to gently push her towards a specific outcome." 78 In other words, it is a system of dynamically personalized nudges, not a single design feature or behavioural intervention. 79 As discussed earlier, general-purpose VAs enable users to access a broad range of functions. At their core, however, many relate to 74  providing information and helping execute pre-determined tasks. 80 User profiling is an integral part of the functioning of VAs to achieve successful service personalization. The goal is to move away from simply responding to a query or executing a task but instead determining and predicting user's needs to give them a dynamically personalized experience, based on their preferences, needs, behaviours and interests. 81 A user profile provides information representing user's specific characteristics and context. 82 The human voice is loaded with information about the user, opening opportunities for voice profiling. Rich research shows that voice holds the cues to detecting not only physical parameters of a person, such as their gender, weight, or height, but also physiological (age, heart rate), demographic (nativity, education, skin colour), medical (general state of the health, autoimmune/genetic/neurological disorders), behavioural (perception of dominance, dynamism, leadership, sexual orientation), and environmental parameters as well as personality and emotions. 83 Furthermore, the probability of two people, even identical twins, sharing precisely the same voice is highly improbable. 84 It is unsurprising that the terms of service of the (leading) VA providers allow companies to process and retain user interactions such as voice inputs to, as they state, "provide, personalize and improve [their] services." 85 When it comes to user profiling, however, the critical point relates to the processing part. Regardless of the quantity of information, the accuracy of the user profile ultimately depends "on the user profiling process in which the information gathered, organized and interpreted to create the summarization and the description of the user." 86 Big technology companies that dominate the VAs' market are well-positioned to profile users accurately. Due to the workings of their respective platform ecosystems, they have not only amassed vast amounts of user data across their many business domains but also are the leaders in artificial intelligence (AI) and machine learning (ML) 80  algorithms. 87 Furthermore, users have an incentive to share their information across the services by the same provider on the cloud, as this creates synergies among those services and allows for better functionality for the user. For instance, if a busy parent asks Amazon Alexa to order household items, they may want their VA to have access to their previous purchasing history. They may also be interested in their child's activity and time spent on Twitch (owned by Amazon) or want entertainment while cooking with Amazon's Prime Video in the background. While many studies have flagged the privacy concerns related to using VAs, 88 it also shows that, for many, these do not outweigh the benefits of having one. 89 Convenience remains one of the most prioritized consumer values, driving e-commerce and, more broadly, the digital economy. 90 User experience is a critical factor, revealing the potency of hypernudging opportunities by VAs. In addition to convenience, trust also plays an essential role in consumers' purchasing decisions online. 91 VAs are purposely designed to feel normalthey can express disappointment and excitement or adjust their voice and tone according to the customer's wishes. 92 Psychological studies confirm that subconsciously people react to devices with human-like qualities as if they were human; they are also often referred to in human pronouns. 93  forms an emotional connection with the device, albeit a one-sided one. 94 Studies have further shown that human-like characteristics in nonhuman objects can induce a high-level of trust and allow a person to sustain a stronger relationship with them. 95 To illustrate, a recent study showed that a robot asking people not to shut them off ignited a social response from the participants. 96 Children, in particular, tend to view a VA as a social partner and want to get to know them. 97 It is also not uncommon for consumers to pose queries about all kinds of intimate questions, including asking to look up illness symptoms or baby names, that indicate trust and allow the platform to glean into their future needs. 98 It is noteworthy that even though consumers effectively give up some of their decision-making powers to the algorithmic agent 99 , they may retain a sense of control over their digital assistant's decisions due to the narrative of a master-servant dynamic. This perceived power is crucial for increasing consumer confidence and the technology's adaptability. 100 People's status can affect their wariness to the VAs as "[a]nthropomorphism increases risk perception for those with low power, whereas it decreases risk perception for those with high power." 101 Similarly, consumers demonstrate increased and more enjoyable interactions with VAs when they feel superior to their devices. In effect, this perceived power mediates their willingness to purchase products with the help of a VA. 102 The state of the art of voice-enabled consumer profiling combined with the design of VAs' technology point to potent consumer influencing opportunities. Leading VAs' providers can engage in such influencing in a large-scale systemic manner. While it might be tempting to assign the potential adverse effects of hypernudging by VAs to fall under the remit of regulation as different regulatory fields play a positive role in safeguarding against harm, they are usually not as such concerned with competition concerns. 103 A further distinction can be made between individual and systemic harms, with relevant regulations to a large extent covering the former. 104 Competition law may complement regulation to address systemic effects on the market. The following section will proceed to showcase why hypernudging by VAs should, at the very least, come under the European competition authorities' radar.

Hypernudging effects on competitive digital markets
The assessment of hypernudging (by VAs) under European competition law necessitates demonstrating that the conduct actually or potentially harms the competitive process. Following modernization, European competition law enforcement is guided by economic principles and is focused on safeguarding economic values, placing consumer welfare at its forefront. 105 In the EU it is equated to consumer surplusit is not enough that a firm's behaviour would increase producer surplus due to efficiencies, possibly at the expense of consumers. 106 With the consumer at the heart of European competition policy 107 , the theories of harm that trigger Article 102 TFEU enforcement generally relate to negative effects upon consumer welfare, including those on price, output, choice, quality, or innovation. 108  embraced a more pluralistic approach. 109 Competition rules are not to be applied in isolation from other Union policies but instead their interpretation requires balancing of different Treaty's objectives against each other. 110 With the changing nature of consumer engagement, trusted VAs are well-positioned to hypernudge consumers towards commercial decisions. One could argue that in a commercial context, critical functions of VAs are to limit consumers' search and information costs. They may facilitate better market transparency and discovery of new, and better, products and services. 111 VAs are expected to assist consumers best when recommending items or shopping for them in the circumstances where "(i) there is effective competition in [voice assistants] market, (ii) the [voice assistant supplier] is independentno integration or contracts with store operators, and (iii) the user may perfectly control the VA's shopping." 112 However, vertically integrated companies have incentives to introduce selection bias that favours their (partners') interests over the consumers' interests, and individuals may not be aware of such misalignment. 113 In addition, automated systems may contain imperfections that unintentionally steer consumers against their interest. 114 While the normative discussion on what constitutes consumer's "true interest" is outside the scope of this article, some observations on decision theory are helpful when considering hypernudging scenarios in digital markets. From neo-classical economic theory perspective, which is the basis of the current European competition policy, it is generally assumed that consumer's revealed preferences (actions) reflect their normative preferences (actual interests). 115 Accordingly, when consumers follow VA's recommendations or allow it to shop for them, the assumption is that the VA serves their preferences and therefore contributes to maximizing consumer welfare. However, behavioural insights show that, at least in some cases, revealed preferences cannot be treated as normative. The dichotomy between revealed and normative preferences is apparent in decisionmaking situations where: (1) consumer is exposed to a default, and therefore is making a passive choice; (2) decisions are complex, requiring consumer incur cognitive costs; (3) consumer lacks personal experiences; (4) marketing with branded commodities is involved; (5) consumer follows impulses and does not account for the long-term consequences. 116 Thus, consumer choices do not always equate to their preferences, instead they could be viewed as a combination of outcome of preferences and application of some heuristics, as well as decision-making errors. 117 Accordingly, people's decision-making is heavily influenced by environmental and cognitive constraintsthey can be (hyper)nudged towards market outcomes that are contrary to their self-interest. 118 In this regard, consider the recent empirical study which demonstrated how conversational robo advisors influence consumers' perception of trust, the evaluation of a financial services firm, and decision-making. The results indicated that consumers are significantly more likely to follow investment advice from a conversational robo advisor compared to non-conversational one, even if the investment advice was inconsistent with their risk profile or invoked larger annual management fees. 119 While financial products are highly complex even for sophisticated 115 John Beshears, James J. Choi, David Laibson and Brigitte C. Madrian, 'How are preferences revealed?' [2008] 8-9 Journal of Public Economics 1787. 116 Ibid, 1788-89. 117 Ibid. 118 To illustrate the power of context in steering consumers' behavior consider the example of addictive design. The term emerged in relation to the design of social media digital user interfaces, such as Instagram, which deliberately leverage human attentive and affective systems to make them stay and engage with the platform longer, often at the expense of their mental health, whilst being exposed to ads. consumers, potentially influencing their inclination to follow the advice, the findings should raise curiosity for future research as to what extent consumers do follow the recommendations of VAs for unfamiliar products and services without double-checking whether their attributes fit their interests or exploring alternatives. When it comes to competition law analysis, hypernudging, and consumer influencing more generally, are not directly addressed by European competition law. Even though personalizationone of the key features of hypernudginghas gained some traction in the literature 120 , its welfare effects are ambiguous and no article 102 TFEU investigation directly about personalization has been opened in the EU at the time of writing. Furthermore, most contributions focus on collection of big data, personalized pricing and behavioural manipulation vis-à-vis exploitative abuses, leading with the argument that consumers are harmed directly.
This article deviates from existing literature by showcasing that hypernudging could lead to potential exclusionary effects on the market. Even though consumers may experience economic harm by being exposed to biased dynamically personalized offerings that deviate from their best interests, competition in the downstream market may, by the same token, be harmed due to firms using behavioural insights on a large scale to shape market's demand side. Therefore, when it comes to consumer influencing, one can establish a link between exploitative and exclusionary effects; the former reinforces the latter.
It should be noted that the connection between exclusion and exploitation has been implicitly touched upon in 2019 Bundeskartellamt's decision against Facebook, which was appealed and ultimately referred for a preliminary ruling to the ECJ. 121 It was found that Facebook abused its dominance in the market for social networks by using its terms of service to collect consumers' personal data on third party websites to provide greater personalization of services. According to the Federal Supreme Court, Facebook's personalized user experience was equivalent to "imposed extension of services", as consumers were forced to accept on-and off-Facebook data processing as a whole package irrespective of whether they wanted such extension. 122 This exploitative behaviour was found to impede competition by limiting consumer choice and degrading service quality. 123 The conduct was subsequently considered to indirectly contribute to creating exclusionary effects, as by virtue of imposed unfair terms Facebook was able to amass huge quantities of consumer data, raising barriers to entry and expansion for competitors on the advertising side of the market. 124 In other words, exclusion was reinforced by exploitative conduct. Similarly, a nexus between exploitative and exclusionary effects could be identified in the context of hypernudging by VAs.
Coming back to the previous example, once a busy parent requests to order or recommend diapers, a VA may point the consumer towards the home (or partner) brand, reciting their best-selling points. It may do so in a way that frames the product to meet individual consumer's requirements; it may adjust recommendations according to the consumer's mood; or, in time, it may recognize a good moment to request a consumer to make it an automatically re-occurring purchase. In all these scenarios, the VA would be hypernudging an individual towards their profitdriven choices that may not accurately reflect consumer's interests and preferences. 125 What makes hypernudging by VAs more challenging to identify and assess than more traditional forms of steering, such as search results by a ranking algorithm or even a personalized recommendation delivered by a recommender system, is that the hypernudging mechanism allows presenting multiple recommendations at different times, perhaps through different channels within the respective platform ecosystem the VA belongs to. Consequently, the consumer steering becomes not only more covert but also more potent. The abovementioned scenario points to problematic market-level outcomes when VAs engage in a systemic diversion of consumer attention and consequently distort demand, especially in the context of biased recommendations that favour some goods and services over the others. 126 Better-quality and value offers may end-up hidden from consumers, resulting in loss of consumer surplus and profits for competitors. 127 However, the firms are incentivized not only by increasing their profits, but also the possibility to control the dissemination of innovations in their respective platform ecosystems, thereby limiting the risk of being disrupted. 128 The systemic diversion of attention has been brought up in abuse of dominance assessments as an issue that could lead to exclusionary effects and therefore reduction of consumer choice. In the abovementioned Google Android decision, which concerned tying Google Search app with the Play Store, one of the main points of contention was examining the extent to which granting a default status to an app will result in significant changes in its levels of usage. 129 The Commission noted that by foreclosing access to rival search engines the company was able improve their search service by gathering more search queries and user data. Thus, by securing a user's attention on one market, the company had an additional advantage over rivals in other markets within its platform ecosystem. 130 This was confirmed by the General Court on 14 September 2022. 131 Notably, VAs are different from other intermediaries in the "attention economy", such as social media platforms. 132 They generally do not aim to extend consumer's time engaging with them. For instance, Amazon has limited advertising options for Alexa, while Google prohibits advertising via its VA, despite earlier attempts to do so. 133 Instead, VAs promise to serve consumers in micro-moments and leave them with more time and cognitive resources to spend elsewhere. At the same time, VAs continuously listen to consumers' mundane interactions and may divert their attention to products and services closely tied to VAs' respective platform ecosystems. This opens doors for more subtle influencing than other obvious forms of advertising.
Convenience seeking consumers' incentives to critically evaluate presented offers, or look out for alternatives, may be further diminished by the personalization aspect of VA's recommendations. In the user-centric digital economy, the ultimate aim of mass segmentation is having consumers that each constitute their own "unique markets" 134 ; the concern is that consumers will be stuck in "targeting pockets" 135 where they are not exposed to diverse assortment of products and services. This exacerbates an information asymmetry not only between the consumer and the firm, but also the firm and its business customers. 136 Since consumers appreciate convenience and VAs typically present a single offer at a time (with a possibility to reject that offer to hear another one), non-discriminatory conditions for being recommended are particularly important for business customers whose sales depend on being discovered. 137 To illustrate the competitive concerns related to hypernudging by VAs, a parallel could be drawn with the recent Commission's investigation into Amazon's "Buy Box." In its preliminary view, the Commission held that the company artificially favoured its own retail business or of those sellers that use Fulfilment-by-Amazon (FBA) service, when selecting a winner of the "Buy Box." 138 Being crowned as a winner is important for marketplace sellers as prominent placing of their offer stimulates a vast majority of sales. In December 2022, the Commission accepted Amazon's proposed commitments, including the application of non-discriminatory conditions and criteria for featured offer and the display of a second offer on the offer display. 139 As will be further discussed in section 4.2, the Commission and the Courts start to recognize the power of biased prominent product placing in digital environments for potentially distorting consumer demand. The difference between the abovementioned "Buy Box" example and VAs is that in the former scenario, consumers are simultaneously exposed to the featured offer and several alternatives that match their search criteria on the Amazon Marketplace webpage or app interface. The visual images and reviews may be helpful in supplementing decision-making, often they must still actively take a few steps before the product is purchased. In contrast, with VAs, this power of biased prominent product placing is further exacerbated since consumers are exposed to one product at the time. Without visual cues and very little information provided, they effectively rely on the VAs recommendation even more.
The above discussion illustrates that hypernudging by VAs is no longer about merely influencing an individual and may lead to negative effects on the market level. Even though exploitation due to direct consumer harm has increasingly been considered a viable route that competition authorities may take in instances involving personalized services and behavioural manipulation, exclusionary effects should not be overlooked since with technological developments, the familiar abusive conduct is expected to morph and advance in form. When it comes to anticompetitive self-preferencing, hypernudging may prove to be a vehicle for more covert and potent "self-preferencing on steroids."

Anticompetitive self-preferencing analysis of hypernudging by voice assistants
Hypernudging by VAs present new challenges to the functioning of the digital markets. The concerns are well summarized by the Stigler Committee: even though consumers retain an illusion of control over their digital interfaces and decisions, it is digital platforms that have a detailed "minute-by-minute control over their interfaces and can present a façade of competition, choice, autonomy when in fact users are directed by behavioural techniques." 140 Nevertheless, while consumer steering may be viewed as problematic or unethical, it does not have to constitute a competition law infringement. This section will examine the legal position of self-preferencing under European competition law and apply its legal criteria to hypernudging by VAs. Given the technological developments, ongoing UI shifts and novel ways, such as hypernudging, to engage in problematic market behaviours, this section will conclude with remaining queries about the future of anticompetitive selfpreferencing.
4.1. Self-preferencing under article 102 TFEU Self-preferencing practices have become a contentious subject in European competition law enforcement, particularly in digital markets. 141 Considered to be one of the flavours of discriminatory behaviour 142 , it is broadly defined as "giving preferential treatment to one's own products or services, or one from the same ecosystem, when they are in competition with products or services provided by other entities." 143 The practice of self-preferencing can take two forms. The first corresponds to competitive distortions on a downstream market induced by a vertical integration of the upstream market dominant player. The second relates to preferential treatment that benefits an independent player, instead of a downstream subsidiary. 144 Self-preferencing is a specific technique of leveraging behaviour, which refers to the extension of an undertaking's market power to a neighbouring market. 145 As such, leveraging is a generic term that is not indicative of article 102 TFEU infringement. 146 It may take various forms, some of them having been found abusive in the past. 147 With digital markets characterized by conglomeration and platform integration into vertical and neighbouring markets, the concerns over anti-competitive self-preferencing have also increased. 148 The platform architecture and ecosystem governance play a role in enabling digital platforms to implement self-preferencing strategies. 149 The incentives for such conduct emerge as self-preferencing can be profitable as soon as it protects the upstream position or allows leverage into adjacent markets. 150 Article 102 TFEU does not prohibit such conduct if it falls within the scope of competition on merits; each case is subject to the effects test. 151 After all, in specific circumstances, self-preferencing is a legitimate business strategy, and giving own products or services preferential treatment could be viewed as a reward for the firm's management that generates efficiencies for both the platform and consumers. A typical selfpreferencing example includes supermarkets introducing home brand products in their assortment, which creates more choices and lower price offerings for consumers, increasing welfare. 152 However, such an example in a brick-and-mortar context does not account for the complexities of the digital markets where market power is not only present, unlike most supermarket scenarios, but is also reinforced by the dynamics of intricately connected platform ecosystems, which may lead to a distortive effect on downstream markets. 153 In markets characterized by high barriers to entry with a specific platform serving as an intermediation infrastructure, clear benchmarks for identifying anticompetitive self-preferencing are ever-more important. 154 In practice, few abuse of dominance cases have focused on the self-preferencing theory of harm. 155 However, in the watershed Google Shopping judgement, the General Court for the first time confirmed that self-preferencing constitutes an independent form of abuse, differentiating it from other forms of leveraging cases, such as refusal to deal.

A turning point: Google shopping
In November 2021, the General Court delivered its long-awaited Google Shopping judgement, described as an "edifice of article 102 TFEU enforcement in digital space". 156  in an algorithmic context, building upon the "equal treatment" principle of European Union law. 157 Google was fined 2.42 billion euros for favouring its comparison-shopping service compared to competing comparison-shopping services on its general search results pages. 158 In effect, the company systemically directed (nudged) consumers towards its service in a secondary market. 159 Google's behaviour consisted of two elements: the company was found to have consistently displayed its own comparison shopping services among the most prominent results on general search results pages and simultaneously actively demoted competing comparison shopping services on those results pages. 160 In this context, the Court started its analysis by rejecting "leveraging" as a relevant theory of harm here by stating that it is not a specific type of abuse but a more generic term encompassing several different practices. 161 The Court did not lie down a universal criterion for identifying a behaviour as anticompetitive self-preferencing. Instead, it proceeded by assessing the context in which the alleged abuse took place, focusing Google's conduct in relation to its business model. The distinction was made between Google's general search results pages infrastructure, which is in principle open as "the rationale and value of a general search engine lie in its capacity to be open to results from external sources" 162 and "other infrastructures referred to in the case-law, consisting of tangible or intangible assets (press distribution systems or intellectual property rights, respectively) whose value depends on the proprietor's ability to retain exclusive use of them." 163 Google's conduct was considered to be abusive because it compromised the open nature of the product in question. 164 For a general search engine to limit "the scope of its results to its own entails an element of risk and is not necessarily rational" 165 unless the company enjoys dominance that is not challengeable in the short term. By favouring its comparison-shopping services on search results pages, Google seems to have acted contrary to the "economic model underpinning the initial success of its search engine." 166 In other words, Google's behaviour ran against its business model, implying that self-preferencing could only be explained by Google's goal to foreclose competition. 167 In light of the effects-based approach increasingly adopted in abuse of dominance cases, the Commission relied on extensive economic analyzes and behavioural evidence showcasing the impact of Google's conduct on competing comparison-shopping service providers' traffic. 168 However, the judgement has been criticized for accepting a rather low standard for establishing anticompetitive effects. 169 According to the Court, it was sufficient to demonstrate potential restriction of competitionor that the conduct is merely capable of leading to the foreclosure 170without applying the as-efficient-competitor (AEC) test to non-price related practice. 171 The Court deviated from the approach taken by the England and Wales High Court in Streetmap.eu v Google Inc., which had been compared to the judgement in question. 172 The case concerned online map provider Streetmap.eu, which accused Google of abusing its dominant position in the search engines market by placing its Google Maps thumbnail image at the top of the search engine results page and therefore favouring own online map services over competitors. Mr. Justice Roth held that to establish an alleged infringement, Streetmap.eu needed to demonstrate the actual effect of the conduct on the market for online mapping services instead of merely a potential effect. 173 He also concluded that in such a leveraging case, "where the likely effect is on the non-dominant market, […] the effect must be appreciable." 174 Accordingly, it was held that Streetmaps.eu failed to demonstrate that Google's conduct would have an appreciable effect on competition, and even in a contrary case, it was objectively justified. 175 Google Shopping adopting a more relaxed threshold for establishing anticompetitive foreclosure signals that competition law enforcers should not be forced to wait for the materialization of actual effects on digital markets, where devising and adopting effective remedies is not only challenging but often too late. 176 The judgment also referenced Google's "super-dominant position" in the general search market, which has been interpreted to be relevant in assessing the effects of the undertaking's conduct. 177 The stronger the market position, the greater the likelihood of foreclosure effect. 178 Since the Court did not explain the use of the term "super-dominance", there is some room left for questioning how anticompetitive effects would be assessed and established when an undertaking engaging in self-preferencing behaviour does not hold such a strong market position. Considering this observation, it would be interesting to consider a case concerning general purpose VAsa market in which the assessment of market definition and dominance would be the initial hurdles in building a successful abuse of dominance case.
Finally, it is noteworthy that the Court has explicitly relied on "the general principle of equal treatment, as a general principle of EU law, [which] requires that comparable situations must not be treated differently, and different situations must not be treated in the same way, unless such treatment is objectively justified." 179 The principle of equal treatment was initially rooted in the market integration rationale and covered non-discrimination of imports, foreign companies and workers. 180 Its scope has since broadened to include the protection of natural persons, which links it closely to safeguarding fundamental rights. 181 The explicit mention of "equal treatment" in Google Shopping can be perceived as unprecedented in the context of European competition law. 182 By using the general principle of equal treatment as an aid in interpreting EU primary law the Court further legitimized its decision to expand the range of conduct that is sanctioned under article 102 TFEU. 183 Furthermore, a reference could be made to the Court placing emphasis on Google's market position in the general search market. 184 A specific mention was made to the common carriers' obligations of equal treatment laid down in EU regulations on net neutrality (Regulation (EU) 2015/2120) and roaming (Regulation (EU) No 531/2012). 185 In combination with the "super-dominant" position of Google 186 , the Court seems to point out that self-preferencing abuse is flexible enough to be transferred to other forms of discrimination. 187

Voice-based services vis-à-vis the future of self-preferencing
Google Shopping considered self-preferencing as an independent type of abuse and was criticized for providing vague standards for assessing when such behaviour deviates from competition on merits. 188 Since the facts of the case concern an investigation opened more than a decade ago, technological developments are expected to create new ways to foreclose competitors by self-favouring in digital markets. 189 Hypernudging by VAs provides a powerful depiction of an advanced way to engage in self-preferencing practices that could be considered by competition authorities. However, the functioning of the generalpurpose VAs' market and their impact, as well as the understanding about hypernudging are yet limited. When it comes to consumer influencing, while digital nudging is becoming a "hot topic" in the digital policy circles, the discussions are predominantly concentrated in consumer law and data protection areas, and the subject is barely touched upon in the competition law debates. 190 This is logical considering that, at least on a surface, the harms would firstly materialize at an individual level, leading to an infringement of rights. 191 Nevertheless, as discussed in section 3.2, exploitation of consumer characteristics and circumstances on a largescale, systemic manner may create opportunities for firms to distort demand side of the downstream market(s).
In assessing hypernudging by VAs as a potential vehicle for self-preferencing that could lead to foreclosure of competitors, one can draw some lessons from the Google Shopping judgement. At its core, the problematic consumer steering in Google Shopping concerned framing of consumers' options. As demonstrated by the Commission's behavioural studies, prominent placing of Google's comparison-shopping services on the search engine results pages was effective because consumers were inclined to "paddle the path of least resistance" and click on the first results. 192 This framing was applied to consumers uniformly, meaning that every consumer making a query on Google Search would receive Google's comparison shopping services recommendation at a specific prominently placed area on the search engine results page. The recommendation placing, therefore, would not account for different consumers' preferences and inclinations.
Hypernudging, on the other hand, is well suited to address consumer heterogeneity. 193 By engaging in hypernudging, market actors can harness voice UI affordances to exploit consumer's vulnerabilities in a dynamically personalized manner. In addition, hypernudging mechanism allows presenting consumer with multiple (behavioural) interventions, at different times, through different intra-platform or interplatform channels, which assessed individually may not be indicative of problematic behaviour. 194 For instance, when a VA is asked to recommend a specific product, a consumer may have purposedly been exposed to VA's suggested brand or model in the respective platform ecosystem prior, be it through an ad, ranking of items on a marketplace or video-content. Having vast amounts of consumer data leading VA providers are well-positioned to identify where in the purchasing funnel the consumer is and when, as well as how, they should be gently pushed to move further towards a purchase. 195 This multidimensionality perspective of hypernudging is particularly challenging in terms of observability and inferring causality, necessitating novel detection methods and techniques to be placed on the future research agenda. 196 Google Shopping is one, among other, cases which has shown that behavioural evidence is becoming utilized in competition law enforcement. In the world where behavioural insights can and are used to exploit consumer vulnerabilities to strategically influence their behaviour in a large-scale manner, taking stock of the relevant empirical behavioural analyzes is valuable in supplementing enforcement, and does not as such necessitate replacing of pre-existing neo-classical economics theories and tools that guided competition law so far. 197 The growing concern over using consumers' cognitive biases and emotional trigger points for anticompetitive purposes has also been corroborated by the reports on digital competition. 198 As a result, in an increasingly consumer-centric digital markets, competition authorities can no longer ignore the impact of business practices, such as hypernudging, that primarily target (individual) consumer experience as a means to foreclose competitors.
In addition, there are some key differences between the nature of voice-based services and general search market analyzed in Google Shopping, which seem to point to wider range of opportunities to engage in differential treatment between consumers and business customers in the case of the former. In the context of VAs, it is important to note that the business models are yet to fully crystallize as the companies are still finding their way in monetizing voice services. The reporting in November 2022 signalled that leading VAs providers have lost revenues, are scaling back on different voice services and are reshaping their strategies. 199 For example, Google is sunsetting Conversational Actions, which allowed third-party developers to build a voice-only service for Google Assistant. 200 Instead, the focus is shifted to App Actions on Android, which allows giving voice commands to Android Apps, such as booking a rideshare or a table at a restaurant. The move seems to have realigned the incentives for developers to support Google's ecosystem as a whole.
It appears that the industry's development is moving towards VAs becoming a mode of engagement with digital products and services, voice being visualized a layer on top of them, rather than the assistance service being its own destination. 201 As the reach of VAs extends beyond a specific business line within the respective multi-product and multi-actor ecosystem that the VA is operating in, ensuring the adherence to the general principle of equal treatment is paramount. 202 Therefore, when considering the exclusionary potential of specific VAs' practices, the platform ecosystem perspective becomes particularly important. This observation further feeds into the discussion on the relevant market definition in digital markets, as the current tools do not adequately grasp the issues related to multi-sided markets, zero-price services and platform ecosystems. 203

Conclusion
The ongoing shift towards voice-based engagement with digital products and services is currently led by a handful of big technology companies that have also dominated the previous UI shifts. This article showcased why VAs by leading providers are well-positioned to engage in hypernudging -dynamically personalized steeringof consumers towards specific market outcomes, such as purchasing decisions, and to seamlessly shape their preferences in favour of platforms' economic imperatives. In such circumstances, hypernudging by VAs may be considered as a vehicle for engaging in anticompetitive self-preferencing that falls under article 102 TFEU.
The combination of hypernudging and voice-based services paints a picture of complexity, extent of which competition authorities have not dealt with before. This article highlighted the overlooked connections between direct consumer influencing and concomitantly direct consumer harm, and exclusionary effects, specifically when firms engage in self-preferencing behaviour by systemically diverting consumer attention towards favoured products or services. Hypernudging by VAs provides a powerful depiction of more advanced and novel ways for firms to engage in self-preferencing behaviour and points to the potential evolution of self-preferencing theory of harm, which was only recently confirmed in the Google Shopping judgement.
The voice-based services are still at early stages and the policymakers are in a favourable position to shape this industry. Recent regulatory initiatives, including the DMA and the proposal for the European Data Act, are important contributions in fostering the contestability of general-purpose VAs market. While the impact of (upcoming) regulations is uncertain, it is a step in a positive direction since policymakers are actively dealing with identified concerns in this market. Since from the business perspective, VAs are developing to become a mode of engagement with digital products and services, instead of providing core platform service in its own right, there is a risk that the (proposed) legislation will focus on the former, more limited, perspective of the VAs market. Therefore, in the context of hypernudging by VAs, to grasp the full potential for exclusionary behaviour, it is imperative to account for the respective platform ecosystem the VA is operating in, as the VAs have a reach for strengthening business lines across that platform ecosystem as a whole.