Using AI to Pick Restaurants When You Travel as a Group

Travel is where AI restaurant recommendations became normal. You land somewhere you do not know, you have no instinct for which neighbourhoods are good, and typing "best dinner spots near the old town, not too touristy" into an assistant is genuinely faster than reading twelve listicles.

It works well right up until you are travelling with other people. Then the same question quietly gets harder, because the answer you get back reflects the preferences of exactly one person: whoever did the typing.

The context problem

A general assistant only knows what is in the conversation. If you mention that you like seafood and want somewhere walkable, it will optimise for that. It does not know that your friend is vegetarian, that two people in the group are staying across town, that one of you is on a tighter budget for the trip, or that half the group has an early flight the next morning.

So you get a confident, well-written recommendation that is wrong for the group in ways the assistant had no way to see. Then the recommendation goes into the group chat, someone points out they cannot eat there, and you are back where you started — except now you have also spent twenty minutes on it.

What actually needs to be in the model

For a group meal on a trip, a useful recommendation has to account for more than taste. In practice that means:

• Dietary restrictions across everyone, not just the organiser — allergies, vegetarian, vegan, kosher, halal.
• Budget range, which varies more on trips than at home and which people are reluctant to raise out loud.
• Where each person actually is. On a trip that is often several hotels or rentals, not one shared starting point.
• Schedule. A group with a 6am flight and a group with a free morning want different dinners.
• What the group has already eaten. Three days into a trip, the fourth great pasta place is not a great suggestion.

None of this is exotic. It is just information that lives in five different heads and never makes it into the prompt.

A better shape for the question

The fix is not a better prompt. It is collecting the group's constraints once, from the group, and letting the recommendation run against all of them. Instead of one person asking an assistant "where should I eat?", everyone answers "here is what I need," and the tool answers "where should we eat?"

That also changes what happens after the suggestion. A chat answer is a dead end — you still have to relay it, defend it, and get agreement. A shared plan can be voted on, adjusted, and confirmed in the same place the suggestion appeared.

How this works in Pixx

In Pixx you create a meal for the trip and invite the group with a link. Each person adds their own dietary restrictions and preferences, and says where they are coming from — which on a trip is the accommodation, not a home address. Pixx uses that to propose restaurants that work for the whole group and sit somewhere fair for everyone, not just near whoever planned it.

From there the group votes, chats about the options in a thread attached to the meal, and gets live updates as things firm up. If the group would rather not deliberate at all, "Just Pick For Me" makes the call from the same filtered set.

The point is not that AI is the wrong tool for finding restaurants abroad. It is very much the right tool. The point is that on a group trip, the useful question was never the one about you.