3 Chatbots, 2 Cities, 1 Lesson: Better Prompts Fix AI Trip Planning

For ML practitioners, this reads less like a travel story and more like an unscientific but useful real-world benchmark of three frontier assistants on a messy, multi-constraint planning task. The undercover methodology — fresh accounts, two distinct traveller personas, and free and paid tiers — controls for the model's memory of the user and isolates the variable that actually moved results: the prompt. The observation that paid versions go 'deeper, faster, and don't run out of steam' is a practical signal about reasoning-tier inference and context retention, while the 'blandness' the author flags is the alignment-temperature trade-off showing up in the wild.

The stubborn belief that AI 'sucks at planning holidays' is really a failure of prompting rather than of the underlying models — that is the central finding of a hands-on experiment published by The Age and Brisbane Times on September 23, 2026. The journalist built two fictional five-day itineraries — São Paulo for a first-time visitor and New York for a veteran traveller who had exhausted the main sights — then queried three chatbots (ChatGPT, Gemini and Claude) in both free and paid tiers. To remove the models' ability to lean on her own history, she used fresh accounts and 'temporary chats' so the systems could not infer that she had lived in both cities for years. The result was a striking reversal: given enough upfront context — dates, budget, interests, pace and prior experience — the bots produced recommendations the author called 'impressive,' even if not enough to retire guidebooks, friend tips and late-night research.

“

For ML practitioners, this reads less like a travel story and more like an unscientific but useful real-world benchmark of three frontier assistants on a messy, multi-constraint planning task.

Wharton professor Ethan Mollick supplied the mental model that the test validates: 'They're not oracles... Think of it like working with a smart person.' A smart collaborator needs context before it can be useful, and so does a large language model. This reframes the popular 'AI is bad at X' complaint as partly a user-input problem. Prompt engineering, long a niche practitioner skill, is becoming an everyday consumer competency, and the quality of a model's output on open-ended planning tasks now scales as much with the brief a user writes as with the model's raw capability. That is a meaningful shift for an industry that has spent most of its marketing energy on benchmark scores and model size.

For AI builders, the test doubles as an unscientific but informative real-world benchmark of three frontier assistants on a messy, multi-constraint task. Two findings stand out. First, the paid-versus-free gap was observable and directional: paid versions 'tended to go deeper and work faster, and were less likely to run out of steam.' That is a practical signal that reasoning-tier inference, longer effective context and stronger instruction-following materially improve subjective output quality on long-horizon tasks — and a warning that free tiers may quietly underdeliver for heavy use. Second, outputs converge on bland middle ground. The author's dig — 'who doesn't like good food?' — is a symptom of the alignment-temperature trade-off: models default to the lowest-risk, most universally palatable answer when the prompt lacks the constraints that would force specificity.

The article also surfaces a genuinely important and under-covered risk: generative engine optimisation (GEO). Just as search engine optimisation reshaped the web, GEO sees businesses 'scramble to game the models so they appear' in AI-generated answers. In travel, that means restaurants, hotels and tour operators optimising their digital footprint to win placement inside chatbot itineraries — a recommendation that can look editorial but is effectively manipulated placement, usually invisible to the consumer. It is a trust and disclosure problem with no settled regulatory answer, and its scale will grow in direct proportion to how many people delegate planning to assistants.