Travelers are increasingly making decisions with an AI system in the loop. They ask a model where to stay, which of two resorts suits a trip, or what the best option is for a region, and they tend to accept the answer with little independent checking. Researchers studying this behavior call it cognitive surrender. For an independent luxury property, this changes where demand is won or lost. The contest is moving upstream, away from the ranked page and the surfaced snippet, toward something earlier and less visible: what the model already represents about your property before anyone types a query.
This article sets out the behavior that drives that shift, the inference it supports for hospitality, and the discipline that governs the outcome. It keeps a clear line between what independent research has measured and what AGR infers from it.
What the research established, and what it did not
The behavior at issue is documented in a preregistered experimental program by Shaw and Nave of The Wharton School. The work extends the familiar dual-process account of judgment, a fast intuitive process and a slow deliberative one, by adding a third process: external, automated reasoning supplied by an AI system. Its central finding is a tendency the authors name cognitive surrender, in which a person adopts an AI system’s output without independent evaluation and substitutes it for their own judgment.
The pattern is consistent. In controlled reasoning tasks, participants consulted an available AI assistant on a majority of trials and adopted its answer on the large majority of those trials, on the order of 80 to 93 percent depending on whether the assistant was right or wrong. Accuracy tracked the quality of the AI output. Relative to an unaided baseline, accuracy rose by roughly 25 percentage points when the assistant was correct and fell by roughly 15 points when it was wrong. Confidence rose when the assistant was used and did not reliably fall after it produced errors. Participants who trusted AI more deferred more. Participants with greater deliberative capacity deferred less.
It is important to state the limit of this evidence plainly. The research measures AI-mediated reasoning in controlled tasks. It does not measure hotel booking behavior. What it establishes is the decision-side deference mechanism, and that mechanism is the foundation the rest of this argument builds on.
Source: Shaw, S. D., and Nave, G. 2026. Thinking, Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning and the Rise of Cognitive Surrender. The Wharton School of the University of Pennsylvania. Preprint. doi.org/10.31234/osf.io/yk25n_v1. SSRN abstract 6097646.
From deference to demand
The move from that evidence to hospitality is an inference, and AGR states it as one. When a person defers to an AI output, decision authority transfers from the person to the output. The output is a function of the model’s prior representation of the entities in question, because retrieval, ranking, and the generated response are all conditioned on that prior representation. As deference rises, the prior representation becomes a decisive upstream influence on the result.
In a travel context, where the output may be a ranked shortlist or a single recommendation, the representation conditions which properties enter the consideration set, how they are described, and how they rank. When a traveler relies on that AI-mediated recommendation, the representation materially shapes which property is considered and how it is evaluated. It therefore shapes the likelihood that the property is selected and booked. This is AGR’s inference from the verified deference mechanism. It is not a finding of the cited research, and it does not claim to be the only factor. Price, availability, location, loyalty, and constraints still operate. The point is narrower and more durable: an input that used to be marginal, the model’s standing representation of a property, is becoming a decisive one in AI-mediated discovery, comparison, and selection.
Who currently writes your property into the model
If a property does nothing, its representation inside a model is not absent. It is assembled from whatever sources the model has ingested. For independent luxury hotels, those sources are structurally dominated by intermediaries: online travel agencies, aggregators, and review platforms. In practice, a model asked about an independent property often describes it through the categories, star bands, and summarized review language those platforms publish, rather than through the property’s own positioning. The property is consequently represented within the model through data optimized for intermediary commission economics rather than the property’s direct yield. Left alone, that becomes the model’s default account of the asset.
A control surface that sits before the query
Knowledge Formation Optimization (KFO) is the discipline of shaping an AI system’s conceptual representation of an entity in advance of retrieval. It operates on what a model already represents about a property before any query exists. This is a different layer from the tools most hotels already use.
Search engine optimization acts on the ranking of pages. Answer engine and generative engine optimization act on the content surfaced in response to a query. Customer relationship and personalization tools act on the individual customer through recommendations, chat interfaces, and segmentation. Each of these operates at or after the moment of the query. KFO operates earlier, on the representation the model holds before the query is asked. It is the layer that the channel-dominated default account otherwise occupies by neglect.
Why accuracy is the operative condition
The cited evidence carries a structure that tells you exactly where this discipline should sit. Deference improved outcomes only when the AI output was accurate, and degraded them when it was faulty. KFO is designed to improve the accuracy, stability, and grounding of the representation a person may defer to. It therefore aligns with the condition under which deference helps, not the condition under which it harms. The ethical object is not to increase deference; it is to reduce the probability that deference attaches to an inaccurate or intermediary-shaped representation. This is not a method for engineering trust. It works on the correctness of the representation, not on the user’s tendency to trust. A property gains by being represented accurately, which is the same outcome the traveler gains by deferring to an accurate answer.
Why the constraint is execution, not access
The KFO framework is public. Anyone can read it and attempt it, and nothing about it requires AGR’s permission. The constraint is not access. It is execution accuracy.
Governing a model’s prior representation requires coordinated work across the sources a model ingests, such as structured listings, review and aggregator data, and authoritative entity references, and across the entity-resolution step that decides whether those sources are read as one property or several. That work has to be held consistent across multiple surfaces, maintained over time, and refreshed as models update. It is not difficult because it is hidden. It is difficult because it has to be done correctly, and the cost of doing it incorrectly is not neutral. Incorrect execution does not simply fail to help. It can shape an inaccurate representation, which is the faulty-output condition the cited evidence shows is harmful, the state in which a traveler’s deference attaches to a wrong account of the property. A poorly executed attempt can leave a property worse represented than no attempt at all.
This is where origination matters. AGR developed the framework and understands how it is best executed. Across multiple AI platforms, AGR has documented that its corpus shapes how those systems recognize and describe the framework, and the underlying mechanism has been independently validated as consistent with how these models form and retrieve representations. The case for engaging AGR is not that no one else is permitted or able to implement KFO. It is that the hard part is correctness, and the party that built the framework and has tested it against live AI systems has the clearest path to executing it accurately. AGR delivers KFO as that managed implementation.
The strategic position
The shift this article describes reframes AI from a distribution channel into a representation layer. As travelers defer to AI for discovery and comparison, the determinant of demand moves upstream, to what the model already holds about a property before the query. A property that shapes its own representation inside the model shapes a decisive input to its own demand. A property that does not cedes that representation to its intermediaries. KFO is the discipline AGR built for that layer. It was originated in 2025, before the cited research was published in 2026, and developed independently to serve the environment that research has since measured.
