Vertical niche apps are fighting for survival by burning through every last resource they have.

From ordering milk tea via Qwen to booking hotels through Doubao, AI is evolving from a mere chatbot to a capable task performer.

Demands that used to be scattered across a range of separate apps are now being consolidated into a single dialog box. This has put vertical and utility-focused apps in an identity crisis. They are losing direct touchpoints with users, putting their advertising and subscription business models under mounting pressure.

It seems that AI is devouring applications, and the content that vertical apps have painstakingly accumulated over years can only end up as "data blood packs" feeding AI models.

Recently, Fliggy launched "Fliggy Helper". It not only answers travel-related questions and plans itineraries, but also handles flight and hotel reservations, order modifications and refunds, online seat selection, hotel room upgrades, and airport transfer arrangements, taking care of the entire process.

Two days later, Ant Group announced that the "Good Doctor Doctor Version" has been officially upgraded to "Ant Afu Doctor Version". This product is no longer just an online consultation tool, but integrates functions including patient management and follow-up visits, medical knowledge retrieval, and scientific research assistance into an AI workstation for medical practitioners.

One handles travel arrangements, the other supports medical diagnosis and treatment. Launched two days apart, these two vertical apps made the same choice: convert years of accumulated professional materials into data that AI can understand and utilize, break down business capabilities into tools that large models can call, and proactively transform themselves into vertical Agents.

Why go through such a thorough transformation? The answer is far from romantic: AI is siphoning off the traffic of vertical apps.

In the heyday of the mobile internet, one demand corresponded to one dedicated app: translation for translation needs, question searching for question searching needs, image editing for image editing needs, and medical consultation for medical consultation needs. The home screen of mobile phones was like a commercial street, with each storefront sign corresponding to a specific business, and users would visit them one by one with their respective demands.

However, after the emergence of AI-native apps such as Doubao and Qwen, they have stationed themselves right at the entrance of this commercial street, becoming the first place users think of when they encounter problems. These AI-native apps may not be masters of every field, but they excel at having a basic understanding of almost everything, and are able to "steadily attend to" users' needs.

For a large number of daily demands, what people need is not the most professional answer, but a fast, passable answer that works, preferably without paying for another separate subscription. Users have limited patience, and their willingness to pay separately for each type of demand is even more limited.

According to data from QuestMobile, in June 2026, the total user base of AI-native apps reached 499 million, with an average of 92.7 uses and 183 minutes of usage per person per month.

Converted to daily terms, users open these AI apps on average three times a day, and interact with AI for six minutes each day.

At the same time, among the mobile internet industry categories counted by QuestMobile, 28% of the industries saw a simultaneous decline in both per capita usage frequency and duration. The underlying logic of vertical apps is being gradually and systematically reconstructed by these AI-native apps.

What is truly being squeezed is the necessity for users to open other apps. The traffic, advertising and subscription models that apps rely on to operate through repeated user visits are having their foundations eroded little by little by AI.

Education apps are the most intuitive example. Photo-based question searching used to be the signature function of such products, with a clear commercial path: when encountering a problem you cannot solve, take a photo to get the answer, then subscribe to the membership to view the detailed explanation.

Now, AI-native apps can already recognize questions, provide answers, and accept follow-up questions from users around the problem-solving process. Qwen can even automatically generate variant questions with the same test points after grading assignments and identifying wrong questions. The signature features that used to be encapsulated in education apps are becoming ordinary side dishes for AI-native apps.

Chegg, the US edtech company, has already felt this substitution effect. With declining traffic and accelerating loss of new subscribers, Chegg has carried out layoffs, closed offices, and once considered multiple options including the sale of the company.

It seems that vertical apps are left with only one outcome: being replaced by AI, then quietly exiting the historical stage. But the story does not end here.

What AI first weakens is the status of vertical platforms as independent entry points, not all the value they have accumulated. The fact that users no longer visit a dedicated platform does not mean that AI no longer needs the products, experience and service capabilities accumulated by that platform.

The professional content, real-time data and business systems that vertical apps have accumulated over years are being separated from the product interface, and respectively turned into training materials for models, knowledge bases that can be called at any time, and execution interfaces for completing tasks.

The first path is to sell data to AI and become training materials for models.

Large models do not lack ordinary text. What is truly scarce is the data accumulated by vertical apps in closed ecosystems that is difficult to obtain from external sources.

Such data usually does not appear in full on public web pages. After long-term accumulation, professional annotation and real business verification, they are high-quality corpora far more valuable than ordinary website content and plain text.

The financial sector has already proved that professional data can indeed widen the gap in model capabilities. BloombergGPT is trained with a large amount of financial corpora, and its performance in multiple financial natural language processing tasks is significantly better than general-purpose models of similar parameter scale.

However, data licensing may not be a sustainably growing business, and is likely to be just a windfall profit. After the model learns the data, it is difficult to sell the same batch of data at the same price year after year. Once the data is fully exploited, it has no further reuse value.

More subtly, AI thrives on high-quality content, but siphons off the resources that originally support content production. The content that the platform has painstakingly built up may end up training a product that replaces itself. Once the apprentice is fully trained, the master will be left with no means of survival.

The second path is to become a real-time knowledge base for AI and survive in the form of plugins.

Some users have already shifted their habit of querying enterprise information from Tianyancha and Qichacha to AI: without buying hundreds of yuan worth of membership, or manually browsing equity structures, legal risks and operational anomalies, they only need to ask AI to directly give the conclusion.

However, public industrial and commercial, legal and operational information is scattered across different web pages, with some outdated updates, some confusing names, and some duplicated and conflicting content. With the rise of GEO poisoning, some content is easier to be crawled and cited, but not necessarily closer to the facts.

The real moat of platforms like Tianyancha is their capability to clean, de-duplicate, annotate, correlate and update this information in real time. What they sell is not a single piece of industrial and commercial information, but the capability to organize scattered information into reliable conclusions.

Kimi has already integrated professional databases such as Tianyancha and Tonghuashun iFind into its membership services. Paid members of Kimi can make queries in the dialog box and generate professional infographics. There is no need to open the original products one by one, nor to buy separate database memberships anymore.

Tianyancha has thus become one of the capabilities and selling points of Kimi, and Tianyancha also gains revenue from this partnership. The specific settlement terms between the two parties have not been made public, but Tianyancha's business model has become more attractive: it used to sell query permissions to individual users, and now it provides plugins and data to AI platforms.

The third path is to become an Agent and directly deliver business results.

Turning a natural language sentence into a real execution in the real world is no easy task. Between being able to chat and being able to perform, there is a full supply chain to be bridged.

Take travel as an example. A large model can quickly write a travel guide for Toronto, but it does not know whether there are remaining seats on a certain flight, whether a transit requires a visa, or whether the hotel offers free room upgrades, nor does it have the authority to modify orders or process refunds. To get these things done, AI must connect to real-time inventory, transaction systems, and service provider networks.

Fliggy Helper conducts post-training based on the platform's real supply, proprietary data, transaction systems and industry experience, then encapsulates search, reservation, modification & refund, and fulfillment capabilities into tools that the large model can call. It verifies inventory, price and order status in real time during execution, to improve the authenticity and executability of the results.

After all, a wrong travel guide will at most waste half a day; a wrong order modification may ruin the entire vacation.

AI's capabilities have enabled the Fliggy app to derive a brand new usage scenario, and also given Fliggy the opportunity to overtake its competitors in the OTT industry. In the future, this set of capabilities can also be integrated into large models such as Qwen.

Users may open Fliggy one less time, but AI will become more dependent on Fliggy to get things done properly.

In general, vertical apps will not all collectively be reduced to "data blood packs" for AI, but will be re-divided in the AI ecosystem: some sell training data, some provide real-time knowledge, and some assist with transactions and fulfillment.

In the past, they competed to be opened by users; in the future, they will compete to be called by AI. Losing their own app as an entry point, they will gain access to an even larger traffic pool.

This article is from the WeChat official account "True Lab", written by Pan Da, edited by Zhang Duo, and published with authorization from 36Kr.