Losing the Market, Owning the Task

Based on the research ofZou, "Platform Choice, Trust, and Privacy in the Consumer AI Assistant Market," arXiv preprint, 2026

Jennifer Zou's survey of roughly 2,000 US AI-assistant users, fielded in June 2026, describes a market that looks, at first glance, exactly like the textbook case for winner-take-most dynamics. ChatGPT is the primary assistant for 58 percent of users. Gemini takes 25 percent. The scale gap shows up outside the survey too: OpenAI CEO Sam Altman said ChatGPT had reached 800 million weekly active users by October 2025, and Google announced in August 2026 that the Gemini app had crossed 1 billion monthly users, calling it the fastest-growing product in the company's history. By the usual logic of platform markets, brand recognition, marketing budgets, and default placement on billions of phones and browsers, that is where the story should end: two incumbents split almost all the usage, and everyone else fights over scraps.

A Market That Looks Tipped, and Isn't Quite

Zou's data complicates that ending. The market is concentrated, but it is also internally differentiated, and the differentiation is organized by task rather than by user. Claude holds just 7 percent of primary usage overall, yet it captures roughly a third of coding tasks specifically, a defensible niche sitting inside a market its owner is losing on every aggregate measure. Task allocation, the paper finds, is driven far more by which platform a person routes a given job to than by which platform they consider their main one. A user can name ChatGPT as their primary assistant and still send their coding work somewhere else entirely.

That pattern is not confined to Zou's 2,000 respondents. A separate, independently sourced dataset points the same direction from a different population. Menlo Ventures' mid-2025 survey of technical decision-makers at enterprises and startups found Anthropic capturing 42 percent of the code-generation market by production usage, more than double OpenAI's 21 percent, even as OpenAI held the larger share of general enterprise LLM workloads earlier in the market's development. Two surveys, drawn from different populations (general consumers in one case, technical buyers in the other) and measuring different things (self-reported task routing versus production API usage), land on the same shape: a task-specific niche in coding that the aggregate market-share numbers do not show.

What Builds Trust When Usage Doesn't

The second pattern in Zou's data helps explain why that niche can exist without a matching share of the broad market. Users build trust through direct experience with a platform, not through its reputation or its marketing reach. Claude is ranked most trustworthy in every head-to-head comparison among people who have used it alongside a competitor, and the paper finds by far the largest gap between how Claude's own users rate it and how non-users rate it, wider than the corresponding gap for ChatGPT or Gemini. Read plainly, that means trust in this market is not evenly distributed by exposure; it concentrates specifically among people who have actually used the product, and forming an opinion from the outside, from ads, headlines, or a friend's description, produces a meaningfully different (and lower) assessment than forming one by using it. The same dynamic applies in principle to every platform in the market: distant impressions and hands-on experience do not track together, and a provider cannot buy the second kind with the tools that build the first.

Privacy Concern Is Universal; Privacy Action Is Not

Zou's third finding shifts from competitive dynamics to what users will pay for. Privacy concern among AI-assistant users is close to universal, but the paper finds that acting on it is gated by awareness rather than by the underlying level of concern; people worry broadly, then act only on the specific risks they understand. In a choice experiment, the single feature users valued most was keeping humans, not models, out of their conversations, worth about $11.20 per month on average, with valuations rising further for more sensitive tasks.

That finding lines up with how the industry has actually been handling data in the same period the survey covers. In August 2025, Anthropic announced that Claude Free, Pro, and Max users would need to choose, by October 8 of that year, whether their conversations and coding sessions could be used to train future models; the in-app prompt defaulted the training toggle to "on," and choosing to share data extended retention from 30 days to five years. Around the same time, OpenAI was fighting a court order, sought by The New York Times as part of its copyright lawsuit, that would have forced indefinite retention of ChatGPT and API conversations rather than OpenAI's standard 30-day deletion window; OpenAI's chief operating officer called the demand a conflict with "the privacy commitments we have made to our users," and the order was lifted in September 2025, restoring the 30-day policy. Neither episode is about keeping humans out of conversations specifically, the exact feature Zou's respondents valued most, but both show the same underlying gap the paper describes: the terms governing what happens to a user's data are being set through settings toggles and courtroom filings that most users never read closely, which is a plausible mechanism for why concern runs so far ahead of action.

None of the three major consumer AI providers currently sells "no human review of your conversation" as a distinct, separately priced feature the way Zou's willingness-to-pay number suggests they could. Data-handling choices exist today mostly as default settings to accept or decline, not as a line item a user consciously buys. That gap between a quantified $11.20 monthly valuation and the absence of a product built around it is, per the paper's framing, exactly what an awareness-gated market looks like: the demand is measurable, but it has not yet been packaged into something a user can knowingly choose to pay for.

What the Data Suggests for a Non-Leader

Put together, Zou's three findings describe a market that behaves less like the classic tipped platform story and more like a landscape with one dominant generalist, one fast-growing second generalist, and durable pockets where neither generalist's scale is the deciding factor. ChatGPT's 58 percent and Gemini's 25 percent are real, and so is the growth behind them; a billion-plus monthly users is not a rounding error. But a 7 percent overall competitor still converts into a third of a specific, high-value task, and it does so through a trust mechanism, direct use rather than brand reach, that scale alone does not buy.

Trust concentrates specifically among people who have actually used the product; forming an opinion from the outside produces a meaningfully different assessment than forming one by using it.

The practical read for a platform that is not going to out-market the leader is that market share and task share are not the same contest, and neither is trust the same contest as reach. A defensible niche, earned through demonstrated performance on one workflow, can coexist indefinitely with a small aggregate footprint. And privacy, treated here mostly as a compliance surface, toggles, retention windows, court filings, is sitting on top of a quantified consumer valuation that none of the current products have turned into a feature people knowingly buy. In a market this concentrated, the fight that is still open is not for the whole thing.