404: cocktail not found
field notes from using an AI travel agent
This post is an interlude from the app-building saga borne out of a road trip. My traveling companions didn’t trust our itinerary to AI, but that didn’t stop me from using it for travel planning anyway.
June 22–27, 2025: Toronto → Montréal → Adirondacks
We were on our third day in Montreal and none of the bartenders we’d met had heard of a Caribou let alone knew how to make one. This was in spite of ChatGPT’s insistence that bars would break out this otherwise wintry cocktail during Quebec’s Fête nationale. After the second bewildered bartender I would have thought it a hallucination — Caribou is also the brand name of a Canadian Whisky — except that a real human friend of ours also encouraged us to seek one out.
Turns out both the human and the AI overestimated its availability.
the journey to the journey
Because it costs next to nothing, I sent both ChatGPT and Gemini down Deep Research rabbit holes to find me the optimal itinerary for a road trip. I’d had some success using ChatGPT to scout out dining options for an earlier road trip around the Southern United States, so I was eager to see how AI would fare in a more comprehensive planning context.
I dutifully advised both models that they were “expert travel advisors” and that their purpose was to “create a flexible itinerary for an upcoming road trip that will emphasize unique, cultural experiences while avoiding overly tourist-y attractions unless the attraction has immense cultural significance.” ChatGPT asked me a few clarifying questions while Gemini went straight to work.
The differences in output1 in spite of their starting with the exact same prompt are striking: compare ChatGPT’s “Montreal Exploration” to Gemini’s brooding “Montreal - A Tale of Two Solitudes”.
ChatGPT’s report seems more direct and condensed, whereas Gemini concocted something that veers into literary territory. They ended up thematically in tune but differed quite a lot in the details, with ChatGPT (anecdotally) making the most mistakes that seemed to bubble up from outdated intel. Neither were particularly opinionated, nor did they provide what I’d consider to be an itinerary, although that’s probably my fault for asking for something flexible.
In true human fashion, I got just what I’d asked for but not what I wanted.
Faced with an overwhelming number of options and flexibility, I completely backtracked, shoved both research dumps into a new prompt, and asked ChatGPT to “construct an opinionated, day-to-day itinerary for the trip, including meal recommendations for breakfast, lunch, and dinner as well as activities and timing for every day of the trip.”
And finally we had a plan for adventure.
how’s my driving?
You need to keep in mind that this was all shadow AI. Neither of the people with whom I was traveling had any idea I’d done this, although they probably suspected I’d do something along these lines. Regardless, I always knew the itinerary would be more of a playbook than an agenda, giving us opportunities to take hold of something specific when decision fatigue was high or to let serendipity guide us when we were feeling whimsy.
Where AI consistently excelled was helping me quickly establish a mental model of a place — for example I had no idea I should be seeking out smoked meat in Montreal — and of course it lacked some context about my traveling style and the preferences of my traveling group.
When I asked ChatGPT to retro its plan versus what actually happened, it simply could not get over the fact that I had a bagel every morning. It also discovered that my friends and I exude boozy energy (what it hilariously referred to as “group stamina”) when we travel and visit every brewery we see. It even kept score as if incredulous: 6 different bagel shops and at least 12 brewpubs.
Its plan resulted in some authentically delightful stops, though, especially when it recommended places we should stop and eat along the way that we would never have discovered on our own. I don’t think that I took credit for our stop for lunch at Valcour Brewing in Plattsburgh, New York because I didn’t know that either Valcour Brewing or Plattsburgh existed without ChatGPT’s help.
Especially where the chance of serendipity is low, like when driving a rural route in upstate New York, such specific recommendations are welcome beacons that so far have panned out wonderfully.
off the rails
ChatGPT also predictably fell down where it had outdated information in its training data.
The most notable example was everything surrounding Montreal’s Saint-Jean-Baptiste Day celebration. ChatGPT’s predictions of localized, neighborhood celebrations never materialized, and it repeatedly advised us to try this mythical “Caribou punch” that no bartenders in Montreal seem to know is a thing people in Montreal drink.
Another completely avoidable although less impactful failure mode was the weather: ChatGPT relied on seasonal norms rather than getting a forecast from the internet showing that Montreal would be atypically tropical and sweltering.
More esoterically, the plan when taken as a whole reveals that, in spite of how familiar it seems, ChatGPT doesn’t really know me at all. Just like a travel planner coming in cold, it couldn’t know how much I love to discover a city by wandering aimlessly through its neighborhoods or how often I’ll ask a server where they like to eat. And so a lot of our deviations from its plan were simple happenstance — our server at a brewery told us to go to Verdun because that’s where he lived, and so we spent an afternoon exploring Verdun.
In any event, ChatGPT acknowledged these missteps in hindsight and of course crafted a wonderful outline of “lessons for next time” that it’ll lose when its context window evaporates as if it were any of the supporting characters in Groundhog Day or Russian Doll.
righting the ship
Maybe the mythical Caribou2 is the price of letting robots plan.
Fortunately my own ability to parse AI’s overconfidence becomes more reliable the more I use it. Like a frequent travel buddy, I’m starting to learn its rhythms, when to follow it blindly to that rural brewpub and when to call it on its bullshit. It’s a surprising emulation of human interaction in that way.
Will it be as immersive when it’s eventually an infallible, superhuman travel advisor? Not sure, but I bet we’ll find out sooner than we think.
I did pick up a bottled version from SAQ before leaving the city. It’s delicious.