After wading through a half-dozen of its recommended itineraries, I asked ChatGPT why it hadn’t put Lyon on any of them. What I got back was a 500-word, four-part essay on why I was wrong, with its final conclusion helpfully bolded so I wouldn’t miss it: “Lyon is a 10/10 city — just not a 10/10 fit for this specific trip.” All it was missing was a bold and centered Q.E.D.
We’d booked a bike tour with Backroads through Burgundy for the last week of April and I thought it’d be interesting to see what AI would suggest we do on the week before. So I fired up ChatGPT and Claude and them to fed them the same prompt:
- We’re flying out of Nashville (BNA);
- At least two nights in every stop so we’re not bunny-hopping;
- We have to be in Auxerre the night before the tour starts;
- We’re here for the local food, wine, cheese, and beer;
- We like walking through towns to see what there is to see (I struggled mightily not to use the possibly-past-trend word flâneuring);
- And we’d like a lunch or dinner in at least one Michelin-starred restaurant.
I already had what I thought was a pretty good itinerary — fly BNA to CDG, catch the TGV from the station inside the airport to Lyon (which, as my post, The Long Way to Lyon, showed, wasn’t quite as straightforward as I thought), spend a couple of days there, train two hours up to Dijon for a couple more days, then over to Auxerre to hang out until we started the tour.
Reasonable, I thought. The models disagreed. When I pushed ChatGPT to justify itself, that’s when it got a little shirty. We started in Lyon anyway and had a great time.
The Last Time I Tried This
It’s a long way from just three years ago; the last time I tried using AI to plan a trip. Back in 2023, shortly after ChatGPT 3 launched, I asked it to:
Develop a taproom circuit of microbrewery taprooms in Asheville, NC starting at the Aloft Hotel in Asheville. Optimize it to minimize walking distance and maximize Untappd ratings.
It spat out a nice little table of eight taprooms that looked solid — at first glance. But then looking closer, none of the ratings actually matched Untappd, and the “circuit” was less a loop and more a random wander around town. The problem: back then, ChatGPT 3 had no web access and a bad habit of confidently making things up to fill in the gaps.
After that experiment, I shelved AI for travel planning. I kept using it more and more for other things — technical analysis, research, and image generation — but not for trips. Then we locked in the Burgundy trip, which got me thinking about giving AI another shot at planning a trip. I’d felt the models getting dramatically better everywhere else I was using them. And honestly, I didn’t have many ideas on what to do with the week beyond “eat and drink in Lyon and Dijon”, so I gave AI another go.
Model Triangulation
Another thing that’s changed since Asheville is the sheer number of large language models (LLMs). Back then, it was ChatGPT. Today, it’s easy to paste the same prompt into seven state-of-the-art models — ChatGPT, Claude, Gemini, Mistral, Kimi, Deepseek, and Qwen. Three American, one French, three Chinese.
Seven was about four too many for me. So after skimming the first round of answers, I cut it down to three: ChatGPT and Claude — my usual suspects — and Mistral, mostly because I thought a French model might have some unique local insight buried somewhere in its training data.
I do what I call model triangulation. I give each model the same prompt and watch how their answers line up. If two or three of them agree on something, I figure it’s pretty safe to believe it. I started doing this as an error check — a way to swerve the hallucinations, and it has mostly worked.
AI is the new mid-level travel agent
The three models gave me pretty similar results, all sitting right in the “fat part of the curve” of tourist recommendations. In Lyon: don’t miss Vieux Lyon (the old town) and get into a traditional bouchon for beef bourguignon or a pike quenelle. In Dijon: hit Fallot’s for a mustard tasting. None of it wrong; none of it bad; just none of it unique.
AI didn’t give me any blinding insights or uncover any hidden finds. What the models did do was collect all the usual suspects — those recommendations that show up in every guidebook, tourism site, travel blog, and review site — filtered them through all the wants and constraints I’d given it, and then hand back a clean, easy-to-read itinerary. Which has real value. It gave me in minutes what would’ve taken me a couple of days of Googling and stitching the results all together. Or what a competent mid-level travel agent would’ve given me — if those still exist.
The Part AI Didn’t (Couldn’t?) Do
What I didn’t get was the next-level research that leads you to the real highlight of the trip. To find great restaurants, Irene leaned on her old standby, the Michelin Guide. I went hunting for local opinions — Chrome-translated pieces from Le Figaro and Lyon Capitale — spelunking through specialty wine blogs and forums for wine bar recommendations and, as always, hitting Untappd to see where people were checking in interesting beers.
For me, AI quickly did the first-level work — the initial research that lays out the contours of the trip, and assembling an itinerary around them. Yes, that’s the easy part of the trip planning, but it’s still work that has to get done. And every hour AI saved me there was another hour I could spend digging further into the stuff it couldn’t reach.
Where AI Actually Won
The place where AI was able to shed its mid-level travel agent tag was weather forecasting. Not a big deal for two weeks in one spot — pull up AccuWeather, give it a look, and you’re done. But this was a touring vacation where we were pulling up stakes every two days. That’s a lot of forecast tabs, a lot of copy-paste to get it all into one place that I could look at and figure out what to pack. Sounded like a good job for AI.
So I gave ChatGPT and Claude the same prompt:
Use your web search tools to give me a day-by-day forecast for the two-week Burgundy itinerary in this public TripIt link.
Claude immediately pulled up short — “TripIt blocks automated access to their pages” — while ChatGPT ploughed ahead. Huh. Well, no matter — I printed the itinerary as a PDF, uploaded it to Claude, and off it went. About four minutes later, both models came back with daily forecasts for every stop, plus a very useful summary. Something that probably would’ve taken me thirty minutes to assemble was done in four.
But the best part came the day before we left. While packing, I went back to both chats and asked them to refresh their forecasts and flag every change of more than three degrees Fahrenheit — the kind of swing that would make me rethink, say, the ratio of short-sleeve to long-sleeve shirts I pack. Three minutes later, I had updated forecasts, and in Claude’s case, the big-change days color-highlighted. That’s something I would’ve never found time to do myself the day before leaving.
Treat It Like a Friend
So I will lean harder in AI travel planning. The models keep getting smarter and they’re getting wired into more real-time data and more travel services. But my Burgundy travel planning reminded me that they’re not psychic — at least, not yet. The more thought and detail you put into the prompt, the better the response.
I’m working on my prompting skills again for a road trip my son Andrew and I are planning through the Canadian Maritimes. Like my Gulf Coast Mardi Gras road trip, this trip has been knocking around the back of my mind for a while. But, like with Burgundy, I don’t have much of a plan beyond “drink local beer and eat local seafood”. Andrew has already uploaded his Untappd history to Claude and had it draft an initial itinerary. Now he’s unleashed a few AI agents to hunt down more ideas.
But even with agents scouring the Internet, AI isn’t all-seeing and all-knowing — not yet — so I can’t just take its itinerary and run. I’ll treat its recommendations the way I would a good friend’s or a sharp travel agent’s: give them strong consideration, check they’re still correct, and then make up my own mind.
Because after all, it’s my trip.
What about your trips? Any stories or tips for using AI to help with your travel planning? Let me know — in the comments below, by email at comments@travelcommons.com, or ping me at @mpeacock on X/Twitter or @travelcommons on Instagram.
Leave a Reply