Ask an AI to plan three days in a city and it will answer with confidence. It will also, now and then, send you to a restaurant that closed last year, a museum wing that is shut for renovation, or a tram that stopped running before you booked your flight. The model is not lying. It is doing what it was built to do: produce a fluent, plausible answer. Plausible and true are not the same thing, and on a trip the gap between them costs you an evening.
We built citysense around that gap. The short version: the AI in our guides chooses and writes, but it never supplies the facts.
Drafted with AI from our own notes on how citysense works, then reviewed by our editor.
Two layers, kept apart
Every citysense guide is made from two separate layers.
The first is the facts layer: a catalog for each city, built before anyone orders a guide. It holds the neighborhoods and what each one is like, the landmarks that matter, how the city eats, how to get around, the phrases worth knowing and the local codes, from greetings to meal hours to what to wear in a church.
The second is the words layer. When you order a guide, AI reads that catalog together with your answers and decides what to put in front of you and how to say it. The rule it works under is simple: choose from the catalog, never add to it.
How the facts layer gets built
AI helps here too, but in a very different role, and with checks on it.
A first AI session drafts the catalog and has to search for and cite a source for every dated or named claim. A second AI session then audits that draft from scratch. It shares nothing with the first one: not its sources, not its reasoning, not its conclusions. It searches every claim again independently and writes up what it finds. Corrections go back into the catalog, anything that needs judgment goes to a person, and only then is the catalog locked.
Facts that repeat across cities, such as entry rules for visitors, live in one shared reference that is verified before new cities are added, so one outdated sentence cannot quietly copy itself into five guides.
Why we recommend types, not addresses
A guide that names a specific bar is out of date the week the bar changes hands. So citysense recommends by type and anchors to landmarks: a family trattoria near the old market, a covered market hall, a neighborhood you can cross on foot. Landmarks stay put. Types of places stay true. Your guide is still useful the week you travel.
What the AI actually does
The personal part is real, and it is where AI is genuinely good. Two people going to Rome will get different guides. Someone who travels for food and goes to bed late is pointed to different neighborhoods than someone who wants the ancient core at the door and an early start. The pace, the three neighborhoods, the experiences, the order of the days: that is matching and framing, and a model does it well.
What the model does not do is decide what exists. If it is not in the catalog, it does not go in the guide.
Two guides, same rule
That is also why we can offer two guides. The City Guide is the read every visitor needs, the same for everyone, straight from the catalog. The Personalized Guide starts from the same facts and adds the three neighborhoods, the experiences and the days that fit how you travel, built from five minutes of questions. Different amounts of AI, one set of checked facts underneath.
The trade-off we chose
This makes us slower to add a city and less "creative" than a chatbot that will happily invent a hidden gem. We think that is the right trade. A travel guide is only worth reading if you can act on it.
Researched and fact-checked. Personalized by AI. That is the whole idea.
If you want to see how it reads, the first pages of every city are free at citysense.guide.
Top comments (1)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.