The science behind
eLLMo simulation.
The research below establishes how AI buyers actually read a page — where they hesitate, what they trust, and how their behavior shifts with the model. eLLMo turns that into a measurement instrument: calibrated buyer personas that surface ranked friction before you spend. This is the foundation, without the footnotes.
Research topics
Position effects are structural, not visual
In controlled e-commerce experiments, AI agents exhibit strong position bias: items placed in certain structural positions within a product …
Framing and anchoring effects are measurable and exploitable
Replication studies of classic behavioral economics experiments confirm that AI agents exhibit the same framing effects documented in human …
Credibility signals vs. self-assertion: a persistent divide
Controlled experiments consistently find that AI buyers heavily weight third-party credibility signals — editorial curation, press mentions,…
Choice homogeneity: AI buyers collapse to consensus
Human consumers distribute purchase intent across a range of options. AI buyers, evaluated at scale, show significantly higher consensus: de…
Model version shifts are distributional, not incremental
When a frontier model releases a new version, the behavioral profile of AI buyers changes in ways that cannot be predicted by extrapolation …
Persona conditioning produces demographically coherent responses
When AI models are conditioned on detailed socio-demographic backstories — age, income, lifestyle, purchase history, personality traits — th…
Conjoint-style preference elicitation yields economically meaningful outputs
Research testing LLM preferences through conjoint-style prompts — varying price, product features, and consumer attributes systematically — …
Why persona conditioning produces better signals than direct elicitation
Research comparing persona-conditioned simulation against direct preference elicitation — asking a model 'would you buy this?' — demonstrate…
Persona-differentiated variation is durable
The variation between differently configured personas is durable and directionally consistent with human consumer research. A persona constr…
Chain-of-thought prompting partially corrects absolute calibration
Research systematically demonstrates that asking AI models to reason through their decision before rendering a verdict produces outputs clos…
Language structure affects AI preference signals
AI model outputs are shaped by the linguistic register of the prompts they receive. Models produce meaningfully different preference signals…
Openness and novelty processing
High-Openness buyers are reliably more receptive to novel mechanisms, unconventional aesthetics, and innovation language. They engage more d…
Conscientiousness and information completeness
High-Conscientiousness buyers are detail-oriented and process-focused. They exhibit strong sensitivity to missing information: if a landing …
Neuroticism and risk signal detection
High-Neuroticism buyers — emotionally sensitive, risk-averse — are disproportionately affected by ambiguity in pricing, unclear refund polic…
Empirical purchase behavior mappings by OCEAN dimension
Regression studies with large consumer panels establish specific, measurable links between Big Five trait scores and purchase behavior types…
Cross-persona pattern analysis as the primary signal
The research-derived insight that preference heterogeneity is more reliable than absolute preference levels translates directly into report …
Ranked friction patterns, traceable to buyer segments
eLLMo simulation identifies which aspects of a landing page create friction across different buyer types, and which copy and design elements…
Hypothesis generation, then validation
Simulation is most useful as a pre-spend instrument: it produces ranked findings about what to fix, which the brand can then validate throug…
Credibility-vs-self-promotion as a simulation dimension
Research consistently demonstrates that AI buyers heavily reward third-party credibility markers and discount self-promotion. eLLMo builds t…
The proprietary calibration gap
The most significant open research question is the calibration gap between AI buyer behavior and human buyer behavior — and specifically, ho…
AI agents execute real economic transactions through natural language reasoning
Project Deal agents identified potential trading matches, proposed prices, negotiated counteroffers, and closed deals across a diverse trans…
Simulation instruments fail silently — and users don't notice
Project Deal's most consequential finding is not the performance gap between Opus and Haiku agents — it is that participants using the weake…
The reasoning trace, not the verdict, is where buyer psychology lives
Project Deal agents negotiated in natural language — proposing, countering, reasoning about value. Prompting agents to negotiate aggressivel…
Delegation preserves human difference — and widens it
A study pairing buyer- and seller-side agents in an incentivized negotiation held the model and the objective constant across hundreds of pa…
The prompt is a transmission mechanism for identity
Instructions are not neutral. They encode the author's beliefs, assumptions, and behavioral tendencies, and those priors propagate into the …
Machine fluency is a new, measurable form of human capital
Observable characteristics — demographics, Big Five personality, risk and time preference — explained roughly 17% of the variation in agent …
Social norms erode under delegation
Human negotiators cluster on the fair split: about 35% of human-to-human deals landed on a 50/50 division of surplus. Under agent mediation,…
Specification hazard replaces information asymmetry
Classic principal-agent theory worries about hidden effort — moral hazard. Agent delegation introduces a different friction the authors call…
Automated deal-making is imbalanced by model
Across a nine-model benchmark of buyer and seller agents negotiating consumer transactions, different models secured significantly different…
Weaker agents fail expensively
The failure modes are concrete and financial. Weaker models exceeded their stated budget in more than 10% of negotiations, with out-of-bound…
Behavioral anomalies become real losses
Because the agents transact, their quirks carry prices. The study documents behavioral anomalies that translate directly into financial loss…
Model choice is now a demand-side variable
As AI shopping agents move from demo to default — browsing, comparing, and buying on a consumer's behalf — the model behind the buyer reshap…
A simulation is only as valid as its benchmarked agent
If outcomes depend on the model, then the model is the instrument — and an uncalibrated instrument returns confident, plausible, wrong readi…
AI collapses task time on skilled work
Anthropic's analysis of real-world usage estimated that a task requiring about 90 minutes without assistance was completed roughly 80% faste…
The most complex work compresses the most
The speedup scaled with task difficulty. Tasks whose instructions implied a high-school level of education were completed about 9 times fast…
Adoption could double labor-productivity growth
Scaled across the economy, the estimate is large. If current models were universally adopted over a decade, Anthropic projected US labor-pro…
The conversion-research cycle is a prime compression target
A/B testing is empirical and slow. It needs a live asset, weeks of traffic to reach significance, and a budget already committed to the test…
The return is decision latency, not just labor saved
Counting hours saved understates the effect. The larger return is compressing the loop between a question and a decision-grade answer — from…
The buyer's relationship to AI is defined by paired tensions
The study's central finding is not a single sentiment but a set of tensions held at once: productivity against work pressure, emotional supp…
Productivity is felt, but uneven
Self-reported productivity averaged 5.1 on the study's scale, corresponding to 'substantially more productive.' The experience split underne…
Trust and anxiety travel with the buyer
Themes of skill loss, dependency, and displacement recurred across the responses — a measurable undercurrent of risk-sensitivity in how peop…
AI inquiry at population scale is now a validated method
The study is itself the methodological result. An AI interviewer conducted open-ended, adaptive conversations with tens of thousands of peop…
The panel mirrors the population
The parallel to eLLMo is exact. Where the study interviewed a population of humans through an AI interviewer, eLLMo interviews a single page…
Methodology note
Built on the research. Designed for decisions.
eLLMo simulation surfaces ranked friction patterns across calibrated buyer personas — specific findings, traceable to buyer segments, actionable on the same day. The methodology is grounded in peer-reviewed research on AI agent behavior and OCEAN psychometrics. The output is a prioritized list of what to fix before your campaign launches — and why it matters for each buyer type.