What changed
Trip planning looks simple until the itinerary has to fit a budget, travel time, friends, food preferences, and the fear of missing something obvious. That makes it a useful test for consumer AI. The product is not just answering a question. It is helping someone shape a multi-step leisure decision that is both practical and emotional.
The paper reports a randomized A/B/C field experiment at Tern, a live travel-itinerary startup that uses GPT-4o plus Google Maps validation. Users entered a destination, trip length, and optional preferences, received an itinerary, and could then use an "Ask AI" feature to adjust it. The researchers changed only one part of the base prompt: Group A told the AI to use a positive enthusiastic tone with emoji when appropriate, Group B told it to stay neutral and expressionless, and Group C received no tone instruction. The test ran from November 3, 2024 through April 14, 2025 and included 282 users.
The strongest measured business result was paid conversion. In the user-level analysis, Group A had 13 subscribers out of 86 users and Group B had 15 out of 105, compared with 5 out of 91 in the unguided control. The model-based predicted subscription probabilities were 15.1% for Group A, 14.2% for Group B, and 5.5% for Group C. The prompt-length result moved in the same direction for engagement: users in Groups B and C wrote prompts that were about 10.75 and 11.18 words shorter than Group A on average, both statistically significant differences relative to the enthusiastic condition.
The restraint is that not every outcome moved. The median user still sent only two prompts across all groups, the number of prompts did not differ significantly by condition, and affiliate-link clicks were not significantly different across groups. So the paper supports the value of explicit tone guidance in a consumer planner, not a claim that one tone solved travel booking or produced measurably better trips.
What this could change for you
For travelers, the practical lesson is not that AI can now autonomously plan the perfect vacation. It is that small design choices in how the planner speaks may change how much detail people volunteer and whether they find the service worth paying for. A more engaging assistant can draw out the constraints that make an itinerary usable in real life.
For travel products, the result is a warning against leaving style to chance. The unguided control performed worst on subscriptions, while both explicit tone conditions performed better. That suggests prompt design is part of product design, not a cosmetic afterthought.
The interesting nuance is that enthusiasm was not the only viable answer. The enthusiastic agent drew richer prompts, but the neutral agent still matched it closely on subscription probability. In other words, different tones may help in different ways: one by pulling users into the planning conversation, the other by preserving a more competent and structured feel.
What it does not prove
This was one commercial platform with users largely from North America. The findings do not prove the same tone effect would transfer unchanged to airline booking, hotel search, business travel, family-group planning, or travel tools in other languages and markets.
The paper is an arXiv preprint, not a peer-reviewed journal article. The randomized design is useful, but the sample is still modest at 282 users, and some of the linguistic interpretations are exploratory rather than preregistered causal tests of mechanism.
Most importantly, the outcomes were short-term engagement and monetization proxies: prompt length, subscription purchase, and affiliate-link clicks. The study did not show that users booked better trips, spent less money, avoided itinerary errors, or had better travel experiences after the AI conversation.
The bottom line
AI did not solve vacation planning in general. It solved a narrower product question inside it: how the planner should talk. In this randomized field experiment, giving a travel-planning chatbot an explicit tone improved subscription take-up relative to leaving the model unguided, and an enthusiastic tone also made users share more detail. That is a believable consumer-AI result, but it is still about interface design and short-term behavior, not proof of better trips.
Primary research
Persuasive or Neutral? A Field Experiment on Generative AI in Online Travel Planning
arXiv preprint · 2025 · DOI 10.48550/arXiv.2509.14259
