The Default Layer Is Already Concentrated
For the consumer AI assistant market, the first question is not simply whether ChatGPT is large. It is how concentration changes as we move from being known, to being used, to becoming the platform that receives most of a user’s tasks. Reading that funnel and then comparing trial with intentional multihoming gives us a way to judge default power without mistaking it for exclusive control.
The distinction is strategically important. A platform that everyone recognizes has an advantage in attracting the next query, but recognition alone does not tell us where work is actually done. A platform that many people have tried may still be peripheral. Primary choice is a stronger signal of reliance, but even primary users may keep another assistant for coding, work, or a sensitive conversation. The evidence shows a strong default layer; it does not by itself show that alternatives have disappeared.
Reach is not reliance
The survey’s weighted estimates create a three-stage funnel. ChatGPT is known by 98.9% of the AI-using population, has been used by 85.7% in the past month, and is the primary assistant for 58.2%. Gemini follows at 95.8% awareness, 77.4% recent usage, and 25.4% primary choice. The gap becomes much wider as the measure moves toward reliance. Claude is known by 79.4% and used by 34.1%, but is primary for 7.0%. Copilot is known by 78.7%, used by 35.0%, and primary for 4.1%.
Those figures are percentages of the survey’s target population: US adults who already use at least one assistant, with the estimates weighted to the AI-user population. They are not adoption rates for all adults, global market shares, revenue shares, or shares of total time spent with each system. The denominator matters because a platform can be familiar to nearly everyone in the frame while still receiving only a portion of their tasks. Conversely, a platform with a smaller awareness base can be highly important to the people who do use it.
Awareness, usage, and primary choice also represent different stages in a competitive process. Awareness makes a platform available for consideration. Recent use demonstrates that the platform has crossed the threshold into someone’s behavior at least once during the month. Primary choice shows where the respondent routes the majority of their tasks. A large difference between usage and primary choice can mean that a provider is easy to sample but less likely to become a default. It can also mean that a platform is used for a narrow but valuable job while another system handles ordinary requests.
The ordering in the primary funnel is not a statistical accident in this sample. The paper reports significant separation between each adjacent primary rank down through fifth place: ChatGPT exceeds Gemini, Gemini exceeds Claude, Claude exceeds Copilot, and Copilot exceeds Grok. The bottom three—Meta AI, DeepSeek, and Perplexity—are statistically indistinguishable from one another, with each having an upper 95% bound at or below roughly 1% of primary choice. The exact lower end matters less than the shape: two platforms dominate the top, a small middle tier follows, and several named alternatives remain marginal as general-purpose defaults.
The concentration is more than a reach effect. ChatGPT’s past-month usage lead over Gemini is 8.3 percentage points, from 85.7% to 77.4%. Its primary-choice lead is 32.8 points, from 58.2% to 25.4%. Conditional on having used ChatGPT, respondents are much more likely to make it their main tool than the usage comparison alone would predict. That pattern is consistent with a strong default or reliance position. It could reflect product fit, habit, accumulated context, distribution, network effects, or several of these at once. The survey does not identify which mechanism causes the lead, so the correct conclusion is about the position itself, not an unproven explanation for it.
Uncertainty changes how precisely the hierarchy should be described, not whether the broad hierarchy exists. The reported 95% confidence-interval half-width is about 2.3 percentage points for ChatGPT’s primary share and 2.1 points for Gemini’s; the intervals are much smaller than the gap between them. The paper also tests adjacent ranks rather than merely displaying a sorted table, finding statistically significant separation down through fifth place. At the bottom, the evidence does not justify a confident ordering among Meta AI, DeepSeek, and Perplexity. This is an important discipline in reading concentration: large, well-separated positions can be stated plainly, while small differences among marginal platforms should not be turned into a league table.
Nor should the hierarchy be read as a forecast. A primary share records where users route the majority of their tasks at one point in time. It does not reveal how often a user reconsidered the choice, how much time or money each task represented, or whether a platform would retain the same position after a price change, a privacy incident, or a rival improvement. The funnel establishes an unequal starting point. It does not establish that the starting point is permanent.
What “primary” captures—and what it leaves out
The survey defines a respondent’s primary platform as the one through which they route the majority of their tasks. This is a useful measure because it captures a practical default rather than a vague preference. If someone uses two systems but sends most of their information requests, work, and personal questions to one of them, that platform has a meaningful advantage in repeated interaction and context accumulation.
But “majority” is not “all.” A user with an 80–20 split is counted in the same primary category as a user with a 51–49 split, even though the competitive implications may differ. The measure also does not say whether the minority platform receives the most sensitive task, the most technically demanding task, or the task with the greatest willingness to pay. A primary-share table compresses a distribution of occasions into one label. It is therefore strong evidence about default position and weak evidence about the absence of alternatives.
This distinction changes how switching costs should be interpreted. Conversation history, saved preferences, uploaded files, and learned expectations can make a primary platform sticky. A user may not want to reconstruct that context elsewhere for every routine request. Yet the same user can still switch when another provider is noticeably better for a specific job. Switching is not one decision made once for the whole relationship; it can be a series of decisions made at different task occasions. The market may be hard to enter at the default level while remaining accessible through a specialized use case.
The practical comparison is not between a perfectly loyal user and a perfectly mobile one. It is between a default that receives work automatically and an alternative that must earn reconsideration. The dominant platform benefits from inertia even if no policy prevents switching. The challenger needs a reason that is salient at the moment a task arrives: a better fit, a different workflow, a trusted experience, or a guarantee that changes the user’s calculation. That is why primary choice is a serious concentration signal without being a complete lock-in measure.
Trying another assistant is not the same as routing work to it
The survey provides a second test. About four in five users say they have tried a second assistant. That sounds like a highly open market until we ask what the trial changed. Only about 43% intentionally route different task categories to more than one platform. The instrument defines intentional routing narrowly and usefully: a respondent must identify at least two distinct primary platforms across the task categories they completed. In other words, the person is not merely reporting a forgotten experiment; they are assigning different jobs to different providers.
The gap between roughly 80% trial and 43% intentional routing is the difference between availability and reliance. Many users are curious enough to test another system, encounter it through work or a bundled product, or use it once for comparison. Fewer make platform choice part of their routine. This is exactly the distinction needed for a market-closure judgment. Trial shows that the incumbent is not literally the only option people will touch. Functional multihoming shows that a substantial minority use alternatives in a way that can discipline a task-level decision.
The pattern is uneven across primary-platform groups. About 51% of Claude’s primary users intentionally multihome, compared with about 38% of Gemini’s. Claude’s users also report the highest platform-match quality, at 4.0 on a five-point scale, and the weakest habit-driven inertia, at 3.3. Gemini’s users report the lowest switching cost, at 2.9. These measures should not be read as a universal ranking of mobility. A low reported switching cost can coexist with a strong default if the platform is easy to use and already receives most routine tasks. Nor does high multihoming prove that Claude will take share from ChatGPT. It does suggest different relationships to the market: one platform can be a settled general-purpose choice, while another attracts users who are more willing to maintain a portfolio of tools.
The comparison has to keep its criteria visible. “Open” might mean that a user can access another assistant at little effort; “competitive” might mean that another assistant repeatedly wins a task; “contestable” might mean that the default must improve or lower its price because users can credibly divert valuable work. The survey gives evidence for the first and some evidence for the second. It does not estimate the third directly. Reported switching cost is a perception, while actual discipline would require observing what happens when a task, price, or policy changes. Keeping these criteria separate prevents the analysis from treating any act of experimentation as proof that incumbency is weak.
The different group patterns also warn against a single market-wide story. A platform can serve users who are comfortable keeping several tools while another serves users who have settled on one, even if both platforms are technically available to everyone. That difference may arise from task breadth, technical engagement, or user selection. It affects how a challenger should be evaluated: not only by how many people it can attract, but by whether the people it attracts use it repeatedly and whether their alternatives are credible enough to matter to the default.
Multihoming is valuable to a challenger for a specific reason. It lowers the requirement for total replacement. A platform does not have to become the primary assistant for every category to matter economically; it needs to become the preferred option for enough repeated occasions. A coder who keeps a generalist for everyday questions but routinely uses a second platform for code is not a fully converted user, yet the coding provider has won a defensible part of the relationship. That position may generate payment, trust, and more opportunities for use.
The reverse is also true. Multihoming does not automatically make the market competitive in a strong sense. If users try alternatives only rarely, if switching occurs only for low-value tasks, or if the default can copy the feature that attracted the diversion, the incumbent may retain most of the economic value. The survey measures the existence of intentional routing, not its frequency within each category, its revenue, its persistence over time, or the direction of future switching. It shows a live margin of competition, not a settled balance of power.
The measure also cautions against treating reported quality and inertia as causes. Users who feel that a platform matches their needs may be less likely to switch because it is good, because they have invested context, or because they have already selected into a particular workflow. Users who multihome may be more technically engaged, have more varied tasks, or simply be more willing to experiment. These explanations can coexist. The evidence identifies different user-platform relationships; it does not isolate the one that produced them.
Consider two users who are both counted as having the same primary platform. One sends nearly every task to that system and rarely thinks about alternatives. The other sends ordinary questions there but uses a second platform whenever a technical task demands a different fit. The primary label treats them alike; their competitive exposure is not alike. The second user gives a challenger a recurring foothold and gives the default a reason to defend that task, even though neither user has “switched” in the total sense.
That is why the most informative follow-up is not simply “How many platforms have you used?” It is whether the same task would be routed elsewhere when the primary option is unavailable, less suitable, more expensive, or less trusted. The survey’s task-level design moves in this direction by asking for primary and secondary choices across categories. Still, the reported multihoming result remains a snapshot of self-described allocation. It does not tell us how much diversion occurs within a category, whether the alternative receives the highest-value work, or whether the pattern survives over time.
The funnel and the routing branch should be read together. The first path shows why the consumer interface is concentrated: recognition and recent usage culminate in a highly unequal distribution of primary reliance. The dashed branch shows why that inequality is not enough to declare the market closed: primary users can still try another assistant, and some deliberately assign different jobs to different platforms.
The bounded conclusion
At the default layer, the evidence is clear. ChatGPT holds the largest primary position by a wide margin, Gemini is a strong second, and the primary ranking is cleanly separated through the middle of the table. This is not a market in which all general-purpose assistants begin from equal footing.
At the same time, “most users have a primary platform” does not mean “most users have only one platform.” The trial rate is high, intentional multihoming is substantial though clearly smaller, and the perceptions of quality, inertia, and switching cost differ across user groups. The market is therefore concentrated, but not behaviorally uniform.
The next question is where that remaining activity goes. If alternatives win only incidental tasks, their presence may be too weak to challenge the default. If they win recurring jobs with distinct requirements, they may have a more durable position than their overall share suggests. To answer that, we have to stop treating the user as the only unit of competition and examine the tasks, audiences, and payment channels that divide the market beneath its dominant default.