During the May Day holiday, AI travel assistants became a hot topic. This article explores the performance and limitations of AI in travel planning through actual experience with multiple AI travel planning tools, pointing out that AI is not fully qualified for the role of travel planners.
If the scene that AI is first integrated into is ranked, the tourism scene will definitely be among the best. Whether it is scenic spots to improve service capabilities, tourists formulate travel strategies, or OTA platforms to fight AI search, it seems that AI can be used to find better solutions.
The general trend of AI towards application and the popularity of Agent have further accelerated the AI exploration process of the travel industry on the C-side.
Before May Day, Fliggy and Mafengwo both launched new AI capabilities. Fliggy’s “Ask a Question” claims to be a multi-agent-driven AI product that can think about problems and perform tasks like professional tourism service practitioners. Mafengwo has officially launched the AI travel assistant “AI Xiaoji”, which supports real-time Q&A, itinerary route planning, online travel guides, and personalized recommendations. At the same time, Mafengwo has also launched the “AI Road Book” function that focuses on strategy planning.
In addition to the vertical platform, general AI search and general agent products will also become one of the tools for some users to formulate May Day travel guides. In addition to DeepSeek and Doubao, Quark, Secret Tower, Button Space, Manus, etc. will also be chosen by some users to become their new tools for formulating travel strategies. We don’t find a quality partner who is willing to spend time and effort preparing for a strategy every trip. AI is believed to be able to play this role well.
This May Day holiday, we also chose products such as Secret Tower, Quark, Manus, Button Space, Fliggy Ask, AI Little Ants, AI Road Book, etc., and each generated a two-day Wuhan travel guide under the same requirements. We want to take this opportunity to truly feel whether AI has brought a tangible and effective experience upgrade to the tourism scene. Moreover, the tourism scene also has a certain degree of representativeness, which can represent the maturity of AI towards practical application.
1. The strategy is mediocre, and there are not many surprises in the experience
The subheadings of this section represent the final feelings of our experience. Although the strategy content given by each product has its own details and emphasis, the overall itinerary differentiation is not obvious. Even if I write to be more personalized, most of them are planning and recommending around several popular attractions such as the Yellow Crane Tower, Tanhualin, Lihuangpi Road, East Lake Scenic Area, Jianghan Road Pedestrian Street, Hankou River Beach, Wuhan University Campus, etc., which is a bit mediocre.
To achieve these three challenges, product managers will only continue to appreciate
Good product managers are very scarce, and product managers who understand users, business, and data are still in demand when they go out of the Internet. On the contrary, if you only do simple communication, inefficient execution, and shallow thinking, I am afraid that you will not be able to go through the torrent of the next 3-5 years.
View details >
In the research mode of the Secret Pagoda, although there is also a summary of the highlights of each attraction and two niche route recommendations, the calculation of the time required between attractions lacks basis. In addition, some of the information cited by the Secret Tower is slightly old and single, with a café recommendation page for 2022, and the food recommendations for the first day of the itinerary are basically from the same article on Ctrip.
The introduction of the highlights of the attractions given by Quark Deep Search is relatively simple, and the two-day itinerary basically covers the popular attractions in Wuchang and Hankou, and there is no personalized attraction recommendation. At the same time, quarks also have the problem of inaccurate calculation of the time required between attractions. But Quark can give some tips at each attraction, such as the Yellow Crane Tower photo spot recommendation, Gude Temple needs to buy incense coupons, etc.
Manus’s strategy is more detailed, although the attraction recommendations are similar to Quark, but the itinerary will be divided according to the time period of morning, noon and afternoon, and it will also give itinerary suggestions from one attraction to the next. For example, Manus recommends walking or cycling from Lihuangpi Road to Hankou River Beach, and lists three things to do at Hankou River Beach. At the same time, Manus will give some seasonal play items or food recommendations, but rarely give recommendations for specific stores.
It is also a universal Agent product, and the buckle space is similar to Manus in terms of body feel. From the perspective of the content of the strategy, the button space gives the ticket price and recommended play time of the attractions, but does not realize that it also takes time to switch between attractions, and the introduction of attractions, dining recommendations and accommodation recommendations are relatively detailed. In terms of overall effect, there is still a certain gap between the buckle space and Manus in terms of product maturity.
Then there are AI Xiao Ant and AI Road Book. In the Mafengwo App, you can directly chat and interact with AI Xiaowang to get route recommendations, or click on the AI road book to generate more detailed travel guides. There will be differences between AI Xiaowang and AI Road Book in terms of itinerary details. The AI Road Book Club includes expansion parts such as city overview, transportation guides, cost analysis, and tips; AI Xiaoji is relatively simple, with only core content such as general routes, recommended stores and gameplay.
Compared with the above products, Fliggy’s Ask is more unique, more like an AI attempt of an OTA platform. From the perspective of the strategy itinerary, the effect of asking is similar to AI Xiao Ant, the route recommendation is not characteristic, and it is quite concise, and there will be an unreasonable arrangement to arrive in Wuhan in the afternoon in the time configuration. But from the perspective of travel experience, Ask is a good machine wine booking assistant, which can recommend different air tickets and hotels according to your budget and click directly to book.
Personally, none of these AI strategies can meet my travel needs, and they are still very tasteless products. From the perspective of tour route recommendations, except for the Secret Tower and Manus, which can give some niche route recommendations, the information given by other products is not much different from the information that can be obtained from a random Wuhan travel guide. From the perspective of catering and accommodation recommendations, due to insufficient judgment basis, it is difficult for me to believe the recommendations given by AI, and I still need to find and confirm them myself.
2. Three key points that have not been met
This tasteless feeling comes from the fact that the maturity of AI in tourism products is far from meeting the requirements for daily use by mass users. Most of the products are not highly finished.
First, many products are still in internal testing, limiting the number of times users can use them. Manus free users only have 1000 points, and it takes more than 500 points to complete a mission. Fliggy’s invitation code is only given to high-level members; Each invitation code of Mafengwo AI Road Book only represents one opportunity to use, and there will also be queues and errors.
Second, most products do not support fine-tuning based on the generated results, nor do they have a contextual memory function, and each time a new requirement is added, a guide will be regenerated. The lack of continuity is constantly restarting the production of the guide. This experience does not match the actual usage habits of users. AI products need to find a smoother travel guide creation process.
Third, this will involve how AI disassembles user needs. One of DeepSeek’s major breakthroughs is that AI has the ability to understand human needs and break them down into task execution steps. How to use this ability to more accurately disassemble user needs is a function that AI products need to be further improved, and it is also a good start to a smooth travel guide creation process.
At present, it seems that most of the AI assistants launched by travel products cannot well disassemble the needs of users. Traveling is often a whim, locking in a point and then exploring more unknowns. Rather than completing user tasks in one stop, it is more important to understand user needs in as much detail as possible. I would think that Mafengwo’s AI road book and Manus’ mechanism of asking for more details can make users’ needs more accurate and make strategy production more effective.
The granularity of demand disassembly will also affect the personalization of the strategy to a certain extent. At present, many users let the AI-generated travel guides are also the same. AI needs to learn to browse information like a human, and instead of humans, it can find its preferred personalized route from a large amount of information. On the one hand, AI needs to better understand human needs and preferences, and on the other hand, it requires more accurate data as support.
Fourth, AI still has hallucinations and needs to learn to cross-verify. AI search products often call public information to produce strategies, but this will face the problem of insufficient information accuracy and timeliness. Building your own knowledge base can be a solution. In the practice of IMA, some people put the collected information on travel destinations into a personal knowledge base, and let AI make strategies based on the knowledge base.
Fliggy has connected to the quotation engine to ask a question, calling air tickets, hotel prices and other data in real time, and Mafengwo has provided AI Xiao Ma with massive real travel data accumulated over more than ten years, which are efforts to provide accurate information and data for AI, and also reduce the “nonsense” of AI to a certain extent. But we can still see some stores that have closed or disappeared from the guide. On the basis of obtaining sufficient data, AI also needs to learn to cross-verify information.
3. Tourism + AI, can be done better
Through experiencing these products, we also realized that although everyone is talking about AI having stronger agency execution capabilities, what can be done by AI on our behalf is still the collection and processing of basic information, completing some preparations, and then letting people make judgments and decisions. The travel guide produced by AI is now only a reference and is the preparation for humans to complete the final strategy.
Even if Fliggy’s question gives us the exact flight and hotel recommendations, we still won’t ask directly to help us make a reservation. Google had previously planned to launch an agent that could help people book flights, but in the end it was just a trip planning tool, because Google also had no way to make the agent perform the actions that users expect very reliably. Perplexity’s Buy With Pro and Amazon’s Buy for me also only provide product information.
Regardless of execution capabilities, data connection can also make AI do better. If Fliggy’s air ticket and hotel data are combined with the personal strategies of Mafengwo and Xiaohongshu, and are called and checked by an agent product like Manus, you should be able to get an experience that is difficult to achieve in their respective products. This needs to solve the problem of information islands left over from the mobile Internet era, which is an ecological change.
General Agent tries to obtain accurate and timely information by configuring virtual machines for AI or connecting vertical services and data through MCP protocols, but the current results are not good. Button space can replace users to search for Xiaohongshu content, but it lacks the ability to filter effective information; There are MCPs for Amap maps, but they cannot be called in a complex way. Even within the ecosystem of the same Internet giant, it is not easy to tear down the department wall.
Obviously, during this May Day holiday, or even the future November holiday, it will be difficult for AI to become a qualified travel planner.
