2026年已经过半，AI旅行规划这件事，产品供给端越来越热闹了。DeepSeek可以随口给你一份行程，飞猪的「问一问」标榜自己是多智能体驱动的AI旅行助手，马蜂窝的「AI小蚂」已经生成了130多万份攻略。更小众一些的赛道里，圆周旅迹、指北旅行、Gooh旅记这些垂类工具也在各自的方向上发力。

Half of 2026 has already passed, and the AI travel planning landscape is getting crowded. DeepSeek can casually spit out an itinerary, Fliggy's "Ask" markets itself as a multi-agent AI travel assistant, and Mafengwo's "AI Xiaoma" has already generated over 1.3 million guides. In the more niche tracks, vertical tools like Pi Travel, Zhibei Travel, and Gooh are pushing forward in their own directions.

听上去，用AI做旅行规划应该已经是一件很成熟的事了。但我们用了两周时间，拿同一个场景——「成都三天两夜，两个人，喜欢美食和历史，不想太赶」——把市面上几款有代表性的产品测了一遍。结果是，没有一款产品能让人放心地把行程交给它之后直接出发。每一款产品都在某一步做得不错，又在另一步让你不得不回到手动操作。

It sounds like AI travel planning should be a mature proposition by now. But we spent two weeks testing several representative products on the market using the same scenario: "three days and two nights in Chengdu, two people, into food and history, don't want to rush." The result: not a single product made us comfortable enough to hand over our itinerary and just leave. Each one does well at some step, but forces you back to manual work at another.

对用户来说，这个链条上的断点，比「AI还不够聪明」更致命。

For users, the breaking points in this chain are more deadly than "AI isn't smart enough yet."

---

PART 01 信息是有的，空间感是缺失的

PART 01 The Information Is There, the Spatial Awareness Is Not

先说通用大模型这一路——以DeepSeek为代表。输入「成都三天两夜，两个人，喜欢美食和历史，不想太赶」，DeepSeek的回复很快：第一天宽窄巷子→人民公园→锦里，第二天大熊猫基地→文殊院→春熙路，第三天杜甫草堂→武侯祠。每个景点都有简短介绍，末尾还附了美食推荐和大概预算。读起来很体面。

Let's start with the general-purpose LLM route — represented by DeepSeek. Input "three days and two nights in Chengdu, two people, into food and history, don't want to rush," and DeepSeek responds quickly: Day 1 Kuanzhai Alleys → People's Park → Jinli, Day 2 Panda Base → Wenshu Monastery → Chunxi Road, Day 3 Du Fu Cottage → Wuhou Shrine. Each attraction has a brief introduction, with food recommendations and a rough budget at the end. Reads quite presentable.

但稍微对成都有一点了解就会发现，这份行程在空间上是不成立的。大熊猫基地在城北，文殊院在市中心，春熙路在市中心偏东——三个地方排在同一天，相当于要走一个三角，交通上并不合理。这其实不是DeepSeek的问题，而是所有通用大模型做旅行规划时的共同缺陷：它们的信息排列逻辑是「语义相关性」，不是「地理空间关系」。一个景点的介绍和另一个景点的介绍在语义上相关，就会排在一起——但实际地理位置可能在城市的两端。

But anyone with even a slight familiarity with Chengdu will notice that this itinerary doesn't hold up spatially. The Panda Base is in the north of the city, Wenshu Monastery is in the center, and Chunxi Road is east of center — putting these three on the same day means zigzagging in a triangle, which doesn't make logistical sense. This isn't really DeepSeek's problem; it's a shared defect among all general-purpose LLMs doing travel planning: their information arrangement logic is "semantic relevance," not "geographic spatial relationship." If one attraction's description is semantically related to another's, they get grouped together — but their actual geographic positions might be on opposite ends of the city.

北二外发布的《AI旅游行程助手类应用能力评测报告》也印证了这一点：在8款主流产品的横向评测中，DeepSeek排名垫底，主要短板是「个性化推荐能力不足」和「多模态能力缺失」。通用大模型能给你的是一份「看起来像行程」的文字，而不是一份「可以按照走」的路线。对大部分人来说，拿到了这份文字之后，下一步的整理工作才刚刚开始。

BISU's "AI Travel Itinerary Assistant Application Capability Evaluation Report" confirms this: in a horizontal review of 8 mainstream products, DeepSeek ranked last, with the main weaknesses being "insufficient personalized recommendation capability" and "lack of multimodal ability." What a general-purpose LLM can give you is text that "looks like an itinerary," not a route you can "actually follow." For most people, receiving this text is only the beginning — the real organizational work starts next.

---

PART 02 预订打通了，规划却成了配角

PART 02 Booking Is Connected, But Planning Became a Sidekick

飞猪的「问一问」是另一个路数。作为2025年4月上线的多智能体AI产品，它可以直接调用飞猪的机票和酒店库存。我们在实际使用中用同样的需求测了一次，问一问会先追问预算、酒店偏好、是否介意转机——这种追问机制确实比通用大模型的一次性回答更像在做规划。行程出来之后，直接附带了机票和酒店推荐，点一下就能订。

Fliggy's "Ask" takes a different approach. As a multi-agent AI product launched in April 2025, it can directly access Fliggy's flight and hotel inventory. We tested it with the same requirements in actual use — Ask first follows up with questions about budget, hotel preferences, and whether you mind layovers. This follow-up mechanism does feel more like actual planning than a one-shot response from a general LLM. After the itinerary comes out, it includes flight and hotel recommendations right there — one tap and you can book.

但问题出在行程本身。推荐的景点以大众打卡地为主，线路安排平庸。36氪旗下「窄播」在2025年五一的评测中记录了一个细节：问一问给武汉行程的时间安排中出现了「下午到达武汉」的不合理配置——这意味着行程的第一天实际上被砍掉了半天。

But the problem lies in the itinerary itself. Recommended attractions are mostly mainstream check-in spots, and the route arrangement is mediocre. "Narrowcast," under 36Kr, recorded a detail in its 2025 May Day review: Ask's Wuhan itinerary included the unreasonable configuration of "arriving in Wuhan in the afternoon" — meaning the first day of the itinerary was essentially cut in half.

这不是一个偶然的bug，它反映出的是OTA平台做AI的逻辑矛盾：交易是核心，规划是获客手段。问一问强在打通了预订链路，但行程规划这件事本身，它并没有比DeepSeek的通用回答强出多少。或者说，问一问更像一个智能预订助手。如果你想的是「帮我订张机票顺便看看怎么玩」，它是合适的。但如果你想要的是「先帮我把行程排清楚」，它的能力是有上限的。

This isn't an accidental bug — it reflects the logical contradiction of OTA platforms doing AI: transactions are the core, planning is a customer acquisition tool. Ask excels at connecting the booking pipeline, but for the actual task of itinerary planning, it isn't significantly better than DeepSeek's general responses. Or rather, Ask is more like an intelligent booking assistant. If you're thinking "help me book a flight and顺便 see how to get around," it's suitable. But if you want "help me sort out the itinerary first," its capability has an upper limit.

---

PART 03 攻略库庞大，但「可执行」是另一回事

PART 03 The Guide Library Is Massive, But "Executable" Is Another Matter

马蜂窝的「AI小蚂」和「AI路书」又是另外一套逻辑。马蜂窝有十几年的攻略和游记数据积累作为底座，2025年4月上线的AI小蚂支持实时问答、行程规划、在线向导和个性化推荐。据官方数据，到2025年底已生成131.5万余份深度旅行攻略，覆盖55个国家、416个城市，累计用户节省约471万小时规划时间。

Mafengwo's "AI Xiaoma" and "AI Roadbook" operate on yet another logic. Mafengwo has over a decade of accumulated guide and travelogue data as its foundation. AI Xiaoma, launched in April 2025, supports real-time Q&A, itinerary planning, online guides, and personalized recommendations. According to official data, by the end of 2025 it had generated over 1.315 million in-depth travel guides, covering 55 countries and 416 cities, saving users approximately 4.71 million hours of planning time.

我们测了AI小蚂和AI路书两个功能。AI小蚂比较简洁，只有大概路线、推荐店铺和玩法，像是一个长版的景点清单。AI路书详细得多，包含城市概况、交通出行指南、费用分析。两个功能生成的内容有差异，但核心景点推荐高度趋同。在评测中，AI小蚂给出的攻略「与随便一篇武汉旅行攻略能得到的信息相差不大」，个性化程度有限。并且，产品整体的使用链路较为冗长，生成旅行攻略的过程需要"排队"，需要用户付出等待大模型输出的时间，最终生成的却是一份通用感很强的计划。

We tested both AI Xiaoma and AI Roadbook. AI Xiaoma is relatively concise — just a rough route, recommended shops, and activities, like an extended attraction list. AI Roadbook is much more detailed, including city overviews, transportation guides, and cost analysis. The content generated by the two functions differs, but core attraction recommendations are highly similar. In testing, AI Xiaoma's guides were "not much different from what you'd get from any random Wuhan travel guide," with limited personalization. Moreover, the overall product workflow is rather lengthy — generating a travel guide requires "queuing," demanding users wait for the large model to output, and the end result is a plan with a strong sense of genericness.

这不是说马蜂窝的数据不丰富，而是说从一个庞大的攻略数据库中提取信息、然后组合成一份「可执行的行程」，不是单纯靠「信息量」能解决的。攻略本身是文本，行程是空间和时间的组合——中间需要一层转换。AI小蚂现在做的更像是把攻略的内容结构化了一下，但从「内容」到「行程」这一步，还不够。

This isn't to say Mafengwo's data isn't rich — it's to say that extracting information from a massive guide database and then combining it into an "executable itinerary" isn't something that "information volume" alone can solve. A guide itself is text; an itinerary is a combination of space and time — there needs to be a transformation layer in between. What AI Xiaoma is doing now is more like structuring the guide content, but the step from "content" to "itinerary" isn't there yet.

---

PART 04 垂类工具冒了出来，但还没人完成闭环

PART 04 Vertical Tools Have Emerged, But No One Has Closed the Loop

除了大模型和大平台，一批专注旅行规划的工具在过去一年里也冒了出来。指北旅行走的是「从零规划」路线：先问你喜欢自然还是人文、热闹还是安静，再根据回答推荐目的地组合，自动生成按天划分的行程。如果从零开始，这个流程比通用大模型要合理一些。但它不支持导入你已有的攻略——如果你已经在小红书上收藏了几十条笔记，需要全部手动输入。实际使用中"个性化"的程度很有限，并且在接入预订的功能使用上明显不如飞猪、马蜂窝等资源丰富。

Beyond large models and big platforms, a batch of travel-planning-focused tools has emerged over the past year. Zhibei Travel takes a "plan from zero" approach: first it asks whether you prefer nature or culture, bustling or quiet, then recommends destination combinations based on your answers, and automatically generates day-by-day itineraries. If you're starting from scratch, this workflow is more reasonable than a general LLM. But it doesn't support importing guides you already have — if you've already bookmarked dozens of notes on Xiaohongshu, you need to manually input them all. In actual use, the degree of "personalization" is quite limited, and its booking integration is clearly inferior to resource-rich platforms like Fliggy and Mafengwo.

轻舟旅记与Gooh旅记则是另一个方向，这一类产品更偏向旅行记录而非行程规划，核心功能是旅行账单、行李清单、手帐制作和组队分享等。

Qingzhou Travel Notes and Gooh take another direction — these products lean more toward travel journaling than itinerary planning, with core features like travel budgets, packing lists, scrapbook creation, and group sharing.

圆周旅迹是目前这个赛道里定位最特别的一个。它把"AI行程规划"做到极致，在聊天模式的对话中用户能够最大程度地自定义自己的出行需求，情侣/小孩/老人/家庭/朋友出行等众多场景，高原/热带/海边/徒步等众多模式都能够给出一份合理的旅行规划。

Pi Travel is the most uniquely positioned player in this track so far. It pushes "AI itinerary planning" to the extreme — in chat-mode conversations, users can customize their travel needs to the fullest extent. Scenarios like couples, kids, elderly, family, or friends traveling; modes like plateau, tropical, beachside, or hiking — it can produce a reasonable travel plan for each.

并且，圆周旅迹"一键抄作业"的功能让我们感到很惊喜，把小红书攻略、全网各种文字或视频链接丢进去，AI可以在三十秒内自动提取出地点与天数，给出一份按天排列、标注交通距离与天气等信息的行程计划；圆周旅迹还会根据地理位置自动排线，不会出现同一天跑三个不相邻区域的问题（当然，这一点在实际使用过程中需要提前输入当日住宿地点，不然AI也有可能排点混乱）。

And Pi Travel's "Copy Homework" feature genuinely surprised us: throw in a Xiaohongshu guide or any text or video link from across the web, and within 30 seconds the AI automatically extracts locations and days, producing a day-by-day itinerary with transit distances, weather, and other information. Pi Travel also auto-sorts by geographic position, avoiding the problem of visiting three non-adjacent areas in one day (though in practice, you need to input your accommodation location for that day in advance, or the AI might still get the sorting confused).

还有一些常规的功能也做的不错，例如说支持多人协作和附近的实时推荐，行程可以拖拽排序、在地图上看路线走向。

Some routine features are also well-executed, such as multi-person collaboration and nearby real-time recommendations. Itineraries can be drag-and-drop sorted, and you can see route directions on the map.

据QuestMobile数据，2026年2月圆周旅迹月活249万，App Store评分4.5分。当然，最好的一点是圆周旅迹完全免费，无广告无内购无预订。它解决的是行程规划这个链条上最具体的一个断点：你看了很多攻略，但不知道怎么把它们变成一份能出发的行程。这个断点看起来很小，但其实是最耗时间的那一步——很多人之所以「做了攻略也等于没做」，就是卡在这一步。但它的短板对于某些人群来说也很明显：不管预订。你排好了行程之后，订票和住宿还是要跳到携程去。

According to QuestMobile data, Pi Travel had 2.49 million monthly active users in February 2026, with a 4.5 App Store rating. And of course, the best part is that Pi Travel is completely free — no ads, no in-app purchases, no booking. It solves the most specific breaking point in the itinerary planning chain: you've read a lot of guides, but you don't know how to turn them into a ready-to-depart itinerary. This breaking point seems small, but it's actually the most time-consuming step — the reason many people "did the guide work but it was as good as not doing it" is getting stuck here. But its短板 is also obvious for some users: no booking. After you've sorted your itinerary, you still need to jump to Ctrip for tickets and accommodation.

---

PART 05 问题不只是「AI还不够聪明」

PART 05 The Problem Isn't Just "AI Isn't Smart Enough"

测试结束，我们的感受是：AI旅行规划的产品选择变多了，但核心体验没有质的飞跃。但这不是因为「AI还不够聪明」。真正的问题在于三件事。

After testing, our feeling is: there are more AI travel planning product options, but the core experience hasn't seen a qualitative leap. And it's not because "AI isn't smart enough." The real problem lies in three things.

第一，每一款产品都只覆盖了旅行规划链条上的一小段。DeepSeek给你创意和信息，飞猪给你预订入口，马蜂窝给你攻略内容，圆周旅迹给你行程结构——但没有一个产品做到了从灵感到预订的闭环。用户还是得在不同的App之间跳来跳去，自己拼接。

First, each product only covers a small segment of the travel planning chain. DeepSeek gives you ideas and information, Fliggy gives you a booking portal, Mafengwo gives you guide content, Pi Travel gives you itinerary structure — but no single product has achieved a closed loop from inspiration to booking. Users still have to jump between different apps and piece things together themselves.

第二，规划和交易的割裂。能做规划的产品没有预订能力，有预订能力的产品规划做得一般。理论上，这两件事应该连在一起——先排好行程，才知道要订哪天的机票、住哪个区域的酒店。但在产品层面，它们是分开的。

Second, the split between planning and transactions. Products that can plan don't have booking capability, and products with booking capability do average planning. In theory, these two things should be connected — you need to sort the itinerary first before you know which day's flight to book and which area's hotel to stay in. But at the product level, they're separate.

第三，用户已经在自救。一个越来越常见的使用模式是：用DeepSeek做灵感收集和目的地研究，把核心信息丢进圆周旅迹或手动整理成结构化行程，最后回到携程或飞猪订票。三个步骤，三个App，没有一个能替代另一个。

Third, users are already saving themselves. An increasingly common usage pattern is: use DeepSeek for inspiration gathering and destination research, dump core information into Pi Travel or manually organize into a structured itinerary, then go back to Ctrip or Fliggy to book. Three steps, three apps — none can replace the other.

这意味着用户需要的不是「一个更聪明的AI」，而是一个能减少切换次数的工具链。即便AI本身的规划能力没有达到理想状态，只要能把现有产品的断点接上，体验就会有明显改善。从这个角度看，目前在规划这个环节上做得最深入的产品，反而是一些不起眼的垂类工具——它们不追求广度，但在自己覆盖的那一小段里，比大平台更细、更具体。

This means what users need isn't "a smarter AI," but a toolchain that reduces the number of switches. Even if the AI's own planning capability hasn't reached an ideal state, as long as it can connect the breaking points of existing products, the experience would improve significantly. From this perspective, the products currently doing the deepest work in the planning segment are actually some inconspicuous vertical tools — they don't pursue breadth, but in the small segment they cover, they're more detailed and more specific than big platforms.

我们在调研过程中注意到的一款叫圆周旅迹的产品就是这样——它不做预订、不做社区、不追求炫技感，只是很朴实地解决了「攻略碎片化到可执行行程」这一步。

A product called Pi Travel that we noticed during our research is exactly like this — it doesn't do booking, doesn't do community, doesn't chase flashy features. It simply and honestly solves the step from "fragmented guides to executable itinerary."

它能不能走得更远，取决于接下来的选择：是继续把规划这件事做深，还是有一天开始做预订和社区。两种路径对应着完全不同的产品逻辑。

Whether it can go further depends on its next choice: continue deepening planning, or one day start doing booking and community. The two paths correspond to completely different product logics.

---

常见问题

FAQ

Q：AI旅行规划工具靠谱吗？

A：取决于用途。通用大模型适合收集灵感和做初步研究，但路线顺序不可靠；OTA平台的AI擅长预订整合，但行程规划能力一般；垂类规划工具在空间排序上更精准，但没有预订能力。最实用的方式是组合使用：AI做灵感，专业工具排路线，OTA订票。

A: It depends on the use case. General-purpose LLMs are good for gathering inspiration and preliminary research, but route order is unreliable. OTA platforms' AI excels at booking integration, but itinerary planning capability is average. Vertical planning tools are more accurate in spatial sorting but lack booking capability. The most practical approach is to use them in combination: AI for inspiration, specialized tools for route sorting, OTA for booking.

Q：DeepSeek规划的行程能直接用吗？

A：不建议直接使用。通用大模型按语义相关性排列景点，不考虑地理位置，生成的行程经常在地图上走不通。需要在地图上检验一遍路线顺序再出发。

A: Not recommended for direct use. General-purpose LLMs arrange attractions by semantic relevance, not geographic position, so generated itineraries often don't work on an actual map. You need to verify the route order on a map before departure.

Q：圆周旅迹和飞猪/马蜂窝的区别是什么？

A：飞猪和马蜂窝的核心是预订和社区，规划是辅助功能；圆周旅迹专注行程规划，不做预订。飞猪的优势在于可以直接订票订酒店；马蜂窝的优势在于攻略内容丰富；圆周旅迹的优势在于基于地图的路线排序和攻略链接一键解析。

A: Fliggy and Mafengwo's core is booking and community, with planning as a supplementary feature. Pi Travel focuses solely on itinerary planning with no booking. Fliggy's advantage is direct flight and hotel booking; Mafengwo's is rich guide content; Pi Travel's is map-based route sorting and one-click guide link parsing.

Q：圆周旅迹的"一键抄作业"怎么用？

A：打开圆周旅迹，把小红书攻略链接或全网各种文字、视频链接粘贴进去，AI会在30秒内自动提取地点和天数，生成按天排列的行程计划，并自动按地理位置排线。支持多人实时协作，完全免费。

A: Open Pi Travel, paste in a Xiaohongshu guide link or any text/video link from the web, and the AI will automatically extract locations and days within 30 seconds, generating a day-by-day itinerary sorted by geographic position. Supports real-time multi-person collaboration, completely free.

Q：AI行程规划工具收费吗？

A：各家不同。DeepSeek基础功能免费；飞猪、马蜂窝的AI功能免费使用但引导预订消费；圆周旅迹完全免费，无广告无内购无预订。

A: It varies. DeepSeek's basic features are free; Fliggy and Mafengwo's AI features are free to use but guide toward booking purchases; Pi Travel is completely free — no ads, no in-app purchases, no booking.