Planning a family trip with a chatbot

A general assistant is genuinely good at the shape of a trip: what a city is for, how many days it needs, what to do with a wet afternoon. What it cannot do, structurally rather than because it is not clever enough, is tell you whether the place it just recommended is open on the Tuesday you are going.

Written by Kate Sinha · Updated 12 August 2026

We use these tools ourselves, every day, and this page is not an argument that they are bad. It is an argument about what kind of question they can answer.

What ChatGPT is genuinely good at

Shaping the trip. Ask it whether five days is enough for a city, what a neighbourhood is like, how to break a long drive with a toddler, or what to do when it rains, and you will get a useful answer in seconds for nothing. It is also good at the thing people underrate: telling you what to cut.

If your trip is straightforward, you are flexible, and being wrong costs you an hour rather than a day, that is often all you need. We say the same thing on the questions page and we mean it.

The failure it cannot engineer away

A language model answers from what it absorbed during training, plus whatever it can fetch while you wait. Neither of those is the same as somebody opening the restaurant's own page and reading the hours for your dates. The gap shows up in a specific, boring, expensive way:

• A place that closed since the model was trained still sounds open.
• Annual closure weeks, which are published once a year in the local language and almost never in English, are invisible.
• A weekday closure gets averaged away into "open daily".
• Travel times come from a map rather than from your front door, and never include the walk from the car park.

Search-connected assistants close some of this, and are much better than they were. But retrieval finds a page; it does not check that page against your dates, and it will happily summarise a listing that was itself last updated two years ago.

What testing has actually found

The published evidence is thinner and older than the confident headlines suggest, so here it is with its dates on.

Forbes, March 2026. Reporting on a travel insurance company's test of three assistants, Forbes noted a suggested restaurant that did not exist and Michelin ratings that were misrepresented. Worth reading with its size in mind: it was three tools on one trip, not a study. It is also worth noting that the publisher has since edited the specifics out of its own article, which is its own small lesson about how fast this ground moves.

The Points Guy, June 2023. A writer handed a Lisbon city break to ChatGPT and was sent to two squares described as street food destinations, both of which he found empty, and to a market on a day it was closed. This is the most concrete published account we could find, and it is three years old, which means it describes a much earlier generation of the technology. We include it because it is real, and flag its age because pretending otherwise would be the same trick we are complaining about.

The agreeableness problem, stated carefully. That these tools tend to agree with you is well documented in general: one 2025 evaluation found sycophantic behaviour in 58 percent of cases across maths and medical questions, and further work has measured the same tendency on open-ended advice. What nobody has done is test it for trip planning. We looked, and the published travel research measures something else entirely, bias across traveller profiles rather than agreement with the user.

So the honest version is this: a documented tendency to tell people what they want to hear, in a domain where it has not been measured, applied to a question people arrive at having already half decided. Our own experience is that you get told your plan looks great. Take that as an observation from someone who sells the alternative, not as a finding.

The honest comparison

The useful distinction is not clever versus stupid. It is generated versus checked.

What we would actually suggest

Use the chatbot for the shape and the shortlist. It is fast, free and good at it. Then check anything whose failure would cost you a morning: opening days on your exact dates, the walk you assumed was ten minutes, and the one booking that has to exist before you land.

That checking is the whole job, it takes hours, and it is the part nobody enjoys. If you would rather not do it, that is what we sell. If you would rather do it yourself, the four checks page describes exactly how, and you are welcome to it.

Our destination research is free to read and shows what a checked place looks like written down.

If you would rather not do the checking

Guides are $99, $179 or $299 depending on how many places you are staying. See the three tiers.