“Alexa, I need to book a doctor’s appointment.”

The assistant answers immediately with a single recommendation, voiced as a friendly suggestion rather than as one option among many. The user hears just one choice. No list of alternatives, no ranking, no visible logic. The interaction feels conversational and helpful, almost personal. Yet behind this simple exchange lies a profound shift in how digital markets function. Increasingly, users interact with the digital world not through screens but through language – asking, responding, clarifying, and relying on conversational systems to mediate access to information and services.

The definition of the „relevant market“ is a central element when assessing mergers and most antitrust cases. The European Commission’s 2024 Market Definition Notice (henceforth “the Notice”) provides guidance on the way the European Commission applies this concept. It is the first revision of the Market Definition Notice since its adoption in 1997 and is generally viewed as a welcome step towards modernising competition analysis for digital markets. It embraces concepts such as multi-sided platforms (para 94), data as a competitive parameter, and zero-price markets (para 97), and recognises that competition increasingly occurs through ecosystems (para 99) rather than isolated products.

Despite these advances, the Notice ultimately treats large language models (LLMs) – which are the basis for conversational systems like voice assistants or chatbots – as if they were variants of established digital services. While this is not explicitly stated in the Notice, it implicitly treats them as such by placing LLMs within broader service categories (e.g., search, operating systems, digital assistants, cloud services) without acknowledging that language-based intermediation is fundamentally different from these services. Because LLMs engage users through conversation, they can shape choices, habits, and behaviour in ways that create new forms of dependence and, potentially, market power. Existing market definition tools are struggling to capture these structures because they were designed for markets where power stems from prices, switching costs, or control over technical infrastructure – not from continuous, personalised linguistic interaction. This interaction generates semantic, pragmatic, and personal data that continuously adapts the service to the user, producing feedback loops and lock-in dynamics that the current Market Definition framework is not designed to detect or assess.

This post argues that the revised Notice, though modernised in many respects, remains ill-equipped to define markets in which undertakings compete not through price or interface design but through the phrasing, tone, and behavioural influence embedded in natural-language systems. By treating linguistic interaction as peripheral, the framework risks causing authorities to overlook key parameters through which firms may exercise market power. If conversational AI becomes a dominant gateway to digital services, as present trends suggest, then competition law will need to rethink how market boundaries are identified in environments shaped by linguistic interaction rather than traditional product features.

Advancement of the new Market Definition Notice

The Notice undoubtedly reflects a significant effort to adapt competition law to digital realities. It incorporates insights about multi-sided markets (paras 94-95), acknowledges the importance of data, and provides guidance for zero-price services (paras 97-98) where traditional price-based tests cannot be applied. It recognises that quality (paras 23, 27, 30), innovation (paras 15, 23) and privacy (para 15) may represent key dimensions of competition, and that ecosystem dynamics (para 99) often determine user choice. These additions correct many of the limitations of the previous framework, which was written for traditional offline markets where prices were clear, switching costs low, and substitution patterns relatively straightforward.

Yet, even these improvements continue to rely on assumptions about user behaviour that do not translate well to conversational environments. The Notice implicitly assumes that digital services are evaluated visually: users see lists, compare alternatives, and make choices based on attributes such as functionality, ranking, or price. This assumption collapses in language-centric markets. A voice assistant often does not present a set of options; it synthesises a single answer. A chatbot does not display competing providers of the goods or services it suggests; it frames information in narrative and dialogue-based form. Most large language models do not offer a menu or search results; they provide a linguistic construction that feels tailored to the user, confident, and authoritative. These systems do not merely deliver content. They perform the answer, mirroring a unique exchange with an emotional being.

Treating such interaction as functionally equivalent to screen-based visual interfaces obscures key aspects of linguistic intermediation. Voice and chat-based systems shape behaviour through the way they formulate responses – through wording, intonation, pacing, and implied authority. The conversational format encourages trust and reduces cognitive effort, as the user is relieved of the need to compare alternatives. It also integrates past interactions into current ones, forging a sense of continuity and familiarity that deepens user reliance.

Linguistic gaps in the Market Definition Notice

This gap becomes clearer when examining how the Notice approaches its core analytical tools. The framework relies heavily on demand-side substitution (paras 25 et seqq.), which asks whether users would switch to another service if the focal one worsened. But many users do not experience a switch from Google Assistant to Google Search or from Siri to a browser as a natural substitution, even if both services ultimately provide information. Interaction through language feels different. Asking a question aloud is not the same as opening an app and typing. The interface is embodied through a voice, often personalised, which strengthens the perception of assistance or advice. Substitution becomes difficult to describe in economic terms when the very nature of the interaction is distinct.

Similarly, the Notice suggests the use of SSNDQ (a small but significant non-transitory decrease in quality; see esp. para 98) where monetary prices are unavailable. Yet quality is especially hard to quantify in conversational systems. What constitutes a decline in quality: more hallucinations, less natural human-like phrasing, slower response times, inconsistent tone, or reduced contextual understanding? These features affect user behaviour in different ways, and no established metric exists for assessing them. Linguistic and pragmatic qualities do not map neatly onto technical performance indicators.

Users often “pay” for digital services with their data. While this framing is helpful, it treats data largely as a uniform input. In language-centric ecosystems, however, language data is uniquely valuable: it contains semantic, pragmatic, emotional, and contextual layers that go far beyond what clickstream data reveals. It discloses intentions, preferences, habits, and social meaning. Users produce this data through spontaneous conversation, not by filling structured fields. It is not publicly available and cannot be purchased in raw form. It can also not be replicated synthetically at comparable quality. The Notice’s general treatment of data does not account for this specificity, nor for the significant competitive advantage that high-quality conversational data confers.

Switching costs in conversational ecosystems also operate differently from those described in the Notice (see e.g. para 57). Users learn how to speak to a specific assistant – what phrasing works, how the system behaves, and how it responds to certain requests. Over time, they adapt to the system’s linguistic routines, building familiarity and expectations. Switching to another assistant requires cognitive effort: one must learn new commands, adjust to a new conversational style, and re-establish context. These behavioural switching costs are real, yet the Notice primarily focuses on technical or contractual switching barriers (esp. para 57-58).

Even the treatment of network effects** **(paras 94-96) requires further nuance. Classical data network effects are already difficult to incorporate into market definition, as more users generate more data, improving services and further attracting users. In language-centric systems, these effects are magnified. Each additional user does not just generate data; they generate conversational data that directly improves the system’s linguistic competence, often in a personalised manner. Performance gains arise rapidly from increasing interaction volume, creating feedback loops that are significantly stronger and more entrenched than those in conventional digital markets. These “linguistic network effects” are not yet addressed in the Notice, even though they may determine market power in AI-driven ecosystems.

This dynamic also illustrates why models such as ChatGPT and Gemini may not, in practice, be strong substitutes, even if the Notice would place them in the same product market based on their similar functionalities. When users interact with one system, it adapts to their language, preferences and routines, creating behavioural sunk costs that make switching to a different model feel less personalised or less effective. As switching becomes harder and substitutability weakens, linguistic network effects can entrench market power in ways the current framework does not yet recognise.

From real world examples to theory

The competitive significance of these dynamics becomes visible in practical examples: Reports have indicated that Amazon Alexa tends to favour Amazon-affiliated products in response to consumer queries, framing these suggestions as conversational advice rather than as search results. Because Alexa usually provides only one recommendation, the user does not see competing options, and the system’s linguistic framing exerts substantial influence over consumer behaviour. Google Assistant similarly integrates voice-based responses with Google’s own services, often phrased in ways that present the recommendation as neutral fact-finding rather than preferential routing. And although the CJEU’s 2024 Meta Platforms ruling did not involve conversational AI, it underscored the competitive relevance of cross-platform behavioural data – an insight that applies with particular force in linguistic ecosystems, where behavioural cues are encoded in interaction itself. Similarly, the OLG Cologne dealt with issues concerning the DMA’s ban on “combining” personal data in a recent national decision. Even though this case actually concerned large language AI models, linguistic or behavioural aspects of language data have not been addressed in the judgment.

These examples highlight a recurring theme: conversational interfaces do not merely mediate access to digital services; they shape the way markets function. They determine which options the user does or does not see. They influence the framing of choices. They accumulate data that competitors cannot match. Such interfaces can create or reinforce market power, not through exclusionary contracts or price strategies, but through the linguistic architecture of interaction.

The way forward

The key question, therefore, is how market definition should adapt to this new environment. Importantly, doing so does not necessarily require revising the Market Definition Notice. The tools for a more behaviourally informed approach already exist within EU competition law; they simply need to be applied with an understanding of linguistic intermediation.

A more functional approach would focus on how users actually use conversational systems. The relevant market may not coincide with traditional categories such as “search engines” or “virtual assistants”, but could instead reflect tasks like product discovery, service navigation, or information retrieval. A behavioural perspective would incorporate linguistic switching costs and user dependence on personalised dialogue, treating them as relevant indicators of substitutability. And a more nuanced understanding of data would recognise that conversational data is not simply an input but a behavioural resource with significant competitive implications.

Competition authorities could also place greater emphasis on assessing intermediation power: Which undertakings control the conversational channel(s) through which users access digital markets? Who determines the default pathways for voice-based product discovery or AI-mediated booking processes? These questions are increasingly urgent because the linguistic interface is becoming the first – and in some cases only – gateway through which users encounter services, information, and commercial choices. As control over this gateway expands, so does the potential for undertakings to steer, shape or limit users’ decision-making in ways that traditional competition law tools struggle to capture.

Conclusions

The 2024 Market Definition Notice is a significant achievement, but linguistic intermediation exposes conceptual blind spots that the updated framework does not yet resolve. As conversational systems become more integrated into daily life, they will increasingly serve as gateways to essential markets. Market power may emerge not from traditional conduct but from control over language itself: the phrasing of responses, the accumulation of conversational data, the behavioural lock-in created by continuous dialogue.

If market definition fails to reflect these dynamics, downstream assessments under Article 102 TFEU or merger control may overlook critical sources of competitive advantage. EU competition law has always adapted to technological change; it is now confronted with a shift from screen-based to language-based market mediation. As this transition accelerates, competition law must learn to interpret the linguistic signals reshaping digital markets.

The market is speaking. Whether the law is prepared to listen remains an open – and increasingly urgent – question.

Barbara Justen is a PhD student at the University of Vienna, focusing on competition law and new technologies.