What the paper studied
The authors investigate how travelers interact with generative AI (GenAI) systems when planning trips, treating the exchange as a cooperative, multi-turn process rather than a one-shot query. Using reflexive thematic analysis, content analysis, and inductive taxonomy on real user-GenAI interaction data (supplemented with interviews), they categorize single-turn user questions and GenAI responses, then abstract higher-level interaction dimensions to derive multi-turn interaction patterns.
Key findings
- User questions in single-turn interactions fall into 10 categories spanning courtesy, exploration, and self-expression behaviors.
- GenAI responses fall into 11 classes, covering information exchange and emotional expression.
- Across multi-turn conversations, interaction patterns cluster into four types: convergence-guidance, convergence-non-guidance, divergence-guidance, and divergence-non-guidance.
- A core mismatch exists between users' diverse, context-shifting questions and GenAI's templated responses, leaving needs unmet and reducing decision-making efficiency and satisfaction.
- Existing research treats interaction as a static outcome; the dynamic, turn-by-turn process of cooperative decision-making has been largely unmapped.
- Each interaction pattern surfaces distinct challenges (vague answers, rigid scripts, missed personalization cues) that require pattern-specific optimization strategies.
Why it matters for hospitality
Travel planning is increasingly happening inside GenAI conversations — 61% of surveyed travelers are willing to use GenAI for travel planning. For hotels, OTAs, and destination marketers building or integrating GenAI assistants, understanding which conversational pattern a guest is in (exploratory vs. converging, guided vs. unguided) is the difference between a useful AI concierge and a frustrating one. The taxonomy gives operators a diagnostic vocabulary for evaluating their own AI-driven booking, concierge, and itinerary tools.
Practical takeaways
- Treat GenAI travel interactions as cooperative processes, not single Q&A exchanges — measure success across the multi-turn arc.
- Detect whether a user is in a convergence (narrowing toward a decision) or divergence (exploring options) mode and adapt response style accordingly.
- Detect whether the user expects guidance (recommendations) or non-guidance (information they will synthesize themselves) — mismatched stances are a top source of friction.
- Avoid templated answers in exploratory or personalized contexts; build branching response strategies tied to the four interaction patterns.
- Use the taxonomy as an evaluation rubric when auditing third-party AI booking and concierge tools.
- Capture and review conversation transcripts to identify which patterns dominate your guest interactions and where GenAI fails to adjust.