A Closed Market Has More Than One Meaning
The subject of this series is the consumer AI assistant market: the small set of conversational platforms that US adults use for information, work, coding, personal advice, creative tasks, companionship, and sometimes highly sensitive disclosure. This opening chapter asks what it would mean to call that market “closed.” The practical value of the distinction is straightforward: it lets us tell the difference between a dominant default, a market with meaningful alternatives, and a production system whose influence has not yet been connected to consumer choice.
That distinction matters because an assistant is not merely a tool a user opens once and forgets. Repeated conversations can accumulate history, preferences, uploaded files, and expectations about how a provider responds. The more context a user entrusts to one platform, the more consequential it becomes to decide whether another platform is genuinely available as a substitute. The question is not only who has the largest audience. It is who receives the user’s important tasks, who is trusted with disclosure, and whether another provider can win a sufficient share of valuable occasions to discipline the default.
The main evidence comes from a June 2026 survey of 1,999 US adults who were aware of and had used at least one AI assistant during the previous month. The estimates are weighted to the AI-using population, not to all US adults. That frame is important: the survey tells us how current users distribute their activity and evaluate providers; it does not measure adoption among people who do not use assistants. A separate Bank for International Settlements study maps 1,246 AI-producing firms across 32 economies. It helps locate the upstream production system, but it does not tell us which firms power a given consumer assistant or how production geography causes downstream choice.
The market has several units
A market can look closed at one unit of analysis and open at another. At the broadest level, we can count platforms and ask which one is a user’s primary assistant. That is a useful measure of default position. It is not the same as asking whether a user has ever tried another system, whether they route coding to a different provider, or whether they would switch if a privacy condition changed.
The survey separates three stages that are often collapsed into the phrase “market share.” Awareness measures whether people know a platform exists. Past-month usage measures whether they have used it recently. Primary choice is narrower: it is the platform through which a respondent routes the majority of their tasks. Each stage describes a different competitive asset. Awareness makes trial possible. Usage shows some actual adoption. Primary choice indicates where reliance is concentrated. None of these measures, by itself, establishes exclusive use.
The distinction between the stages also protects against a misleading comparison. A platform can be used by many people but still lose the role of default if those users distribute most of their tasks elsewhere. Conversely, a platform can have a large primary share without being the only system its users know or use. The denominator changes the meaning of the result: a percentage of all aware users, a percentage of recent users, and a percentage of primary users are not interchangeable quantities. The survey’s weighting further aims at the composition of the AI-using population rather than the general adult population, because the people who have already adopted assistants are not a random slice of everyone else. That does not remove uncertainty, but it keeps the estimate pointed at the population the question actually concerns.
Primary reliance also has a practical asymmetry. The default platform benefits whenever a new task is routine enough that the user does not reconsider the choice. A challenger must often make a reason to switch visible at the moment of use: better fit for a job, a trusted experience, or a data-handling condition that matters for that particular conversation. This is one reason accumulated context can matter without proving that users are locked in. Context raises the stakes of changing providers, but task-specific value can still create a path for another platform to receive work. The market’s openness is therefore better tested by observed allocation across occasions than by asking whether a second icon appears in someone’s history.
The same logic applies to the word “user.” A person may use several assistants without treating them as interchangeable. They might ask one system for everyday information, another for code, and a third for a sensitive personal deliberation. In that case, the competitive unit is not only the person or the platform. It is the task occasion—the repeated job for which a provider is chosen. A platform can be small in overall primary share yet important in a task that users consider difficult, valuable, or costly to get wrong.
This is why the series will treat closure as multidimensional. Default concentration asks who receives most of the activity. Task competition asks where alternatives win. User and payment patterns ask whether a platform serves a distinctive audience or reaches people through employers and schools rather than household subscriptions. Trust asks whether reputation is inherited or changed by experience. Privacy asks which data-handling promises users value enough to alter behavior. These are not interchangeable definitions of competition; they are different tests of how much room remains beneath the default.
Two concentration stories, not one
The consumer interface already presents one concentration story. The survey’s headline pattern is a hierarchy led by ChatGPT and Gemini, with smaller platforms occupying much narrower overall positions. That hierarchy matters because a primary assistant is not just another app in a crowded folder. It is the place where the largest share of a user’s questions and context is routed. If one provider is the default for most occasions, challengers must overcome more than awareness; they must become useful enough, trusted enough, or distinct enough to receive repeated work.
But the same consumer evidence points to a second story. Users do not make one undifferentiated choice for every task. Some deliberately route different categories to different platforms. Some smaller platforms have a strong task signature or a distinctive payment channel. Trust rankings can differ from reach rankings, and privacy preferences distinguish human review from model training or sponsored answers. These patterns do not erase the concentration of the default. They change what that concentration means. A dominant generalist may control the front door while other providers retain valuable rooms inside the house.
The upstream AI economy introduces another kind of concentration. The BIS study divides production into compute, cloud and related infrastructure, data tools, AI models, and AI applications. Its sampled firms are concentrated especially in the United States and China, while other economies specialize in particular layers. An economy can have firms in several layers yet hold most of its valuation in one of them. Investment can also be geographically home-biased and concentrated within the same supply-chain layer.
That upstream map matters because consumer assistants depend on capabilities that are not visible in a chat window. Yet it must be kept in its proper place. The BIS study describes where qualifying AI-producing firms are located and how they invest; it does not measure consumer assistant shares, task routing, trust, or privacy behavior. The relationship between production capacity and consumer control is a question for further evidence, not a conclusion that can be imported from firm counts.
There is a second reason to resist a single upstream number. Compute hardware, cloud capacity, data tools, models, and applications have different capital needs, technical bottlenecks, and routes to customers. Concentration in one layer does not automatically mean that the same firms control every other layer. The BIS sample itself is selective: it includes firms above a $500 million valuation threshold, combines public market capitalization with private estimated valuations, and classifies businesses into primary roles. Those choices make the map useful for identifying large production hubs and specializations, not for declaring that every relevant supplier—or every consumer-facing alternative—has been counted.
The map is therefore deliberately asymmetric. The consumer-side branches are measured directly by the survey. The upstream branch is measured by a different study and joins the question only as context. Treating the two as one market would create a category error: the location of AI-producing firms is not the same variable as the platform receiving a user’s next conversation.
What a defensible answer would require
A defensible answer cannot begin with the comforting observation that alternatives exist, nor with the dramatic observation that two platforms dominate aggregate choice. It has to ask what kind of control is being claimed.
If “closed” means that the general-purpose default is already concentrated, the evidence can speak directly to that claim. We can compare awareness, recent use, and primary reliance, then examine whether the primary platform receives most tasks because users never try alternatives or because alternatives are used only at the margins. The difference matters. A market can have a powerful default while still allowing meaningful diversion when a task exposes a weakness or when a user values a different guarantee.
If “closed” means that smaller platforms cannot build durable positions, the evidence must be more demanding. We need to know whether users intentionally multihome, whether task niches persist, whether a niche attracts payment or institutional distribution, and whether users return after trying a challenger. A single trial is weak evidence of competitive pressure. Repeated routing for a consequential task is stronger, although even that does not show that the niche will survive improvements by the default platform.
If “closed” means that users’ evaluations are already fixed by brand size, trust and privacy complicate the picture. A challenger may be less trusted by people who have never used it yet rank more favorably among its own users. Privacy concern may be widespread without becoming action, because users do not know what a provider does or cannot justify the price of a protective plan. Those mechanisms can create room for differentiation, but only if experience, information, credible guarantees, and affordability line up.
Finally, if “closed” means that the entire AI system has become structurally controlled by a small set of producers, the upstream evidence is relevant but incomplete. Geography, finance, and supply-chain specialization may influence who can build and scale alternatives. They do not, on their own, establish control over consumer distribution. The missing bridge is direct observation of how production firms, models, infrastructure, applications, and consumer assistants connect.
The time horizon matters as well. A niche that exists today may be absorbed by a generalist, weakened by a change in pricing, or strengthened as users learn which tool fits a recurring job. A privacy preference reported in a hypothetical choice may fail to change behavior when a real subscription requires money or when a user’s context is difficult to move. Conversely, repeated task routing, persistent trust advantages, and credible controls could make a small overall platform harder to dismiss than its aggregate share suggests. The available evidence can establish the starting positions and the mechanisms worth watching; it cannot settle their durability from one cross-sectional snapshot.
The thesis of this series is consequently narrow but consequential: the consumer AI assistant market is concentrated at the default layer, yet durable closure across tasks, users, trust, privacy, and production capacity is not established by these sources. The first test is the simplest one. Before discussing niches or privacy, we need to measure how large the default lead really is—and whether primary reliance should be mistaken for exclusive dependence.