Everyone uses AI to plan travel these days. But have you noticed—same AI itinerary, some read like a seasoned travel expert, others read like something a robot generated. I took the same request, tested 5 mainstream AI agents, and the gap was ridiculous.
This post breaks down the test and analyzes why AI travel planning quality varies so much.
The Test: 5 AIs, Same Request
I set one unified test request and had 5 AIs each plan:
Plan a 2-day Hangzhou trip from Shanghai. Me and my girlfriend. Total budget 2000 RMB. We like photography and avoiding crowded spots. Plan the itinerary, hotel, and transport.
Then I scored across four dimensions (out of 10):
- Itinerary reasonableness: Is the route smooth? Timing reasonable? Too rushed?
- Detail richness: Specific times and locations? Transport options? Restaurant recommendations?
- Price accuracy: Are quoted prices reliable? Budget control reasonable?
- Personalization: Did it account for "likes photography" and "avoids crowds"?
The Results: How Big Is the Gap?
Total scores first:
| AI Agent | Itinerary | Details | Price Accuracy | Personalization | Total |
|---|---|---|---|---|---|
| GPT-4o | 9 | 9 | 7 | 9 | 8.5 |
| Claude 3.5 | 8 | 9 | 7 | 8 | 8.0 |
| ERNIE Bot | 7 | 7 | 6 | 6 | 6.5 |
| Gemini | 7 | 6 | 6 | 6 | 6.3 |
| Kimi | 6 | 7 | 5 | 5 | 5.8 |
See? First place 8.5, last place 5.8. A gap of nearly 3 points. This isn't "which one is slightly better." This is "professional vs. amateur."
First Place: GPT-4o, Why Does It Sound Like an Expert?
First place, GPT-4o. Why does it plan like an expert?
1. Itinerary is especially reasonable
The itinerary it planned: morning high-speed rail to Hangzhou day one, afternoon photography by West Lake (crowd-free route, not squeezed onto Broken Bridge), evening snacks at Hefang Street. Day two, Lingyin Temple (early, fewer people), afternoon tea and photography at Longjing Village, evening back to Shanghai.
Route is smooth. No backtracking. It deliberately scheduled Lingyin Temple in the morning and Longjing in the afternoon to avoid crowded time windows. That's the kind of detail someone who actually understands travel considers.
2. Personalization works
User said "likes photography, avoids crowds." It actually listened. It recommended photogenic but relatively uncrowded spots—like Longjing Village and Yanggong Causeway—instead of crowded Broken Bridge and Leifeng Pagoda. That's real understanding of user needs.
3. Details are on point
Which high-speed train, which area to stay, how to allocate budget—all spelled out. It even recommended photogenic restaurants, with per-person costs. The detail level rivals a professional travel planner.
Last Place: Kimi, Why Does It Sound Clueless?
Last place. Why the gap?
1. Itinerary is unreasonable
What did it plan? Morning arrival in Hangzhou day one, afternoon Leifeng Pagoda, evening West Lake night view. Day two morning Lingyin Temple, afternoon Songcheng, evening back to Shanghai.
What's wrong? Leifeng Pagoda and West Lake night view crammed into the same afternoon and evening. Too rushed. And it completely ignored "avoids crowds"—Leifeng Pagoda, Broken Bridge, Songcheng, all the most crowded spots. User said no crowds, and it ignored it.
2. Made-up prices
It quoted hotels at "around 300 a night." Actually, West Lake area hotels in Hangzhou, 2 nights for two people, 300 doesn't get you much. Total budget 2000 for two, and it blows the budget just on hotels. No money left for food and attraction tickets.
3. Personalization is nearly zero
User said "likes photography." It recommended the most generic spots. No thought toward "photogenic." It's like a templated itinerary with the generic "Hangzhou 2-day trip" template applied. It didn't read the specific user request.
Why the Gap? Three Reasons
Same LLMs. Why does travel planning quality vary so much? Three reasons:
First, whether it connects to real-time data (MCP)
This is the biggest gap source. If GPT-4o is connected to a hotel MCP and flight MCP, it queries real-time prices and availability. Prices are accurate. If it's not connected, it makes things up from training data memory. Prices are wrong.
Many AI travel products now connect MCP. Like using RollingGo's hotel MCP to connect real hotel data. Thousands of developers have already integrated. Supports 40+ major LLM clients. Behind it: 2M+ global hotel properties and 110K+ direct-contracted hotels. Direct inventory with real-time price confirmation. What you see is what you can book. Completely free, no call-volume limits. Provides an all-in-one booking skill for ClawHub and Coze. Developer partners set country-based markup rates and earn commission.
MCP-connected: prices accurate, availability accurate. Not connected: all made up. Want your AI assistant to stop sounding clueless? Follow the quick-start guide and connect it.
Setup in Codex:
[mcp_servers.RollingGo-Hotel]
url = "https://mcp.rollinggo.ai/mcp"
http_headers = { "Authorization" = "Bearer YOUR_API_KEY" }
Restart Codex and you're calling hotel search, room detail, and price confirmation directly. Whether an AI travel assistant sounds like an expert or a robot comes down to whether it connects to real MCP data.
Second, prompt engineering quality
Same LLM. The quality of your prompts makes a big difference. Some AI travel plans have very detailed prompts—consider user preferences, control budget, arrange reasonable routes. Others are written loosely, letting the AI freewheel. Results naturally differ.
Third, LLM's own understanding ability
This is the fundamental gap. GPT-4o's understanding is stronger than smaller models. It truly understands what "likes photography, avoids crowds" means. It allocates budget reasonably. Smaller models can't understand this level of nuance. They apply templates.
Advice for Users: How to Pick an AI for Travel Planning
After this round of testing, a few tips:
1. Prioritize AI that connects to MCP
If this AI can check real-time hotels and flights, its prices are accurate. If it's pure chat, prices are made up. Don't trust it.
2. Pick a strong-understanding LLM
For tasks like travel planning that need complex requirement understanding, don't use tiny models. Bigger models understand more accurately and plan more reasonably.
3. Ask multiple rounds. Don't trust the first result.
Don't use the first generated itinerary directly. Ask a few more rounds: "Can this hotel be cheaper?" "Is this attraction crowded?" Adjust a few rounds and the result gets much better.
Same AI travel planning. The gap is ridiculous. Some sound like professional travel experts. Others sound clueless. The reasons: whether it connects to real-time data, prompt engineering quality, and the LLM's own strength.
Next time you use AI to plan travel, don't just use the first result. Pick a reliable one, adjust a few rounds, and the experience gets much better.
Which AI do you think plans travel best?
Top comments (0)