The buyer on your page has changed, and the change is now measurable. In December 2025, Anthropic ran the largest qualitative study of AI users to date — roughly 81,000 people across 159 countries and 70 languages, interviewed over the course of a single week by a conversational AI agent and classified at scale by another. The result is the clearest population-level read available on the mindset your visitor carries when they arrive at a purchase decision.
The study and why its method matters
The interviewer was not a survey form. Anthropic deployed a frontier AI assistant in a conversational, open-ended format, conducting adaptive interviews that followed each participant's own framing — responding to what they said, probing where they hedged, pursuing threads that static questionnaires would have lost. AI classifiers then structured the responses across multiple dimensions, turning population-scale open-ended talk into comparable, categorized findings. That pipeline — AI-conducted interviews, AI-structured results — is itself a methodological proof of concept. Structured inquiry by an AI agent, at a scale no traditional qualitative program could reach in the time available, produced signal that holds up to analysis. This is not the output of a panel study conducted over months; it is a snapshot of the active user population, captured in a week.
The headline finding: a buyer held in tension
The study's central result is not a sentiment score. It is a structure: four paired tensions, each held simultaneously by the same people. Productivity against work pressure. Learning against skill loss. Emotional support against dependency. Economic opportunity against job displacement. Anthropic did not find optimists and pessimists in separate segments. They found optimism and anxiety occupying the same person, in the same response, about the same tool. That coexistence is the condition your page meets.
The productivity data illustrates it precisely. Self-reported productivity averaged 5.1 on the study's scale — 'substantially more productive' — but the distribution underneath that average tells the real story. About half of participants reported genuine, unreserved time savings. Roughly one in five described a treadmill: the time AI freed up was absorbed immediately by rising expectations, higher output standards, or a workload that expanded to fill the gap. And the people reporting the largest personal speedups were, as a group, among the most anxious about whether their role survives AI adoption. The gain and the threat arrived together.
'The people reporting the largest speedups were also among the most anxious about job displacement.' That pairing is the psychological signature of the AI-shaped buyer: genuine enthusiasm, held on the same breath as genuine risk.
How paired tensions produce a faster, more skeptical reader
A buyer shaped by this experience does not approach a product page with open credulity. Daily use of AI has trained them to evaluate output, not just receive it — to notice when a claim is unsupported, when a mechanism is missing, when a brand is performing competence rather than demonstrating it. The same critical posture they apply to AI-generated content follows them into the purchase environment. They evaluate faster because they have spent months pattern-matching against plausible-sounding text that did not deliver. They extend less default trust because they have learned that confident presentation is not the same as accurate content.
This is not generalized consumer skepticism. It is a specific learned behavior: the habit of checking whether a claim is grounded before treating it as true. On a landing page, it expresses as a reader who moves from hero claim to evidence before they move from evidence to interest — and who withdraws the moment a page asks for trust it has not earned. The AI-fluent buyer's stance moves on proof, not on adjectives.
The study's learning-versus-skill-loss tension reinforces this. Users who worried about cognitive atrophy from over-reliance were, in many cases, already noticing it in themselves. That self-awareness about what uncritical adoption costs them extends to their commercial choices: a buyer alert to dependency in their own work is alert to dependency in a product promise too. 'Just trust us' is the exact register they have trained themselves to discount.
What one buyer's voice looked like in the data
The classifier categories give the aggregate picture, but the qualitative texture matters for understanding what conviction actually sounds like in this population. A recurring pattern in the study was the 'support versus dependency' tension: participants described real relief — AI catching errors, explaining unfamiliar concepts, holding the scaffolding of a task — alongside a named fear that they were hollowing out a capability they would need later. The emotional valence was not resigned; it was active, watching, conditional. 'It helps me do this faster, but I am not sure I am learning anything.' That is not the voice of a buyer who will be won by efficiency claims alone. It is the voice of a buyer who wants to know whether the product's efficiency comes at a cost, and whether that cost is acknowledged.
How eLLMo's personas model this buyer exactly
eLLMo's simulation panel is calibrated to the buyer the study describes — not to a friendlier abstraction of them. Each persona in a simulation run carries the specific psychological disposition that produces the reading behavior the study measured: the high-Conscientiousness buyer checking every claim against visible evidence; the high-Neuroticism buyer reaching for risk signals before she reaches for the CTA; the Openness-forward buyer who responds to mechanism and is unmoved by outcome promises that skip the 'how.' These are not demographic segments. They are the OCEAN-structured expression of the same hope-and-anxiety pairs the Anthropic study found at population scale.
A simulation reads the page the way this visitor does — weighing each claim against the support available on the page, checking whether risk is addressed before the ask, whether the product's mechanism is legible enough to hold up to scrutiny. The output is first-person friction, ranked by how widely it surfaces across the panel: 'the hero promise is credible but the proof arrives four screens below where a skeptical buyer has already made their call,' or 'the guarantee that would settle my displacement anxiety is absent from the page.' Those are the same hesitations the study found at population scale, surfaced one page at a time.
The methodological parallel is exact. Where the Anthropic study interviewed a population of humans through an AI interviewer and classified the results into structured findings, eLLMo interviews a single page through a calibrated panel of AI buyers and ranks the results into prioritized friction. Both turn open-ended AI inquiry into structured, decision-grade output. The study read a population to understand AI use. eLLMo reads a page to understand whether it converts — and to tell you, in that buyer's own register, what is standing between interest and action. See how the ocean methodology conditions each persona to the segment it represents.
The sample skew is an advantage
The 81,000 participants were drawn from existing Anthropic users — meaning the sample skews toward people already in a formed, active relationship with a frontier AI assistant. That skew is not a limitation on the study's usefulness; it is a specification of its signal. This is the buyer population that is arriving at DTC pages now, in numbers, and it is the buyer population that will define the baseline over the next two to three years as AI adoption moves from early-adopter density to mass-market norm. The study does not describe where the average consumer sits today. It describes where the market is heading — and brands that design for that reader now are working ahead of adoption, not scrambling to catch up.
For eLLMo simulations, this is the useful direction. The personas are calibrated to the buyer the market is producing, not the buyer the market had. A simulation run against this standard surfaces the friction that will matter at scale, before the broader population has arrived to make it visible in your own analytics. The productivity compression research adds a further dimension: as AI compresses the time buyers spend on category research, the window in which a page has to earn trust shrinks. The AI-shaped buyer is not only more skeptical — they make the call faster.
The commercial implication
The study does not recommend a page treatment. What it establishes is the operating context: the buyer arriving at your page now carries an AI-formed evaluation habit, a paired set of real enthusiasm and real risk-awareness, and a trained eye for the gap between confident presentation and substantiated claim. Pages built for a less-informed, more credulous reader are not meeting that buyer where they are. The brands that close the gap — that bring evidence to the place where skepticism concentrates, that address risk before the ask, that make their mechanism legible to a reader who will check — are the brands designing for the buyer that the data describes. That buyer is already here.