Author: cc Sun Congying

Editor: Hou Yu

As AI agents gradually acquire the capability to directly invoke external services, the long-debated issue of traffic allocation in the internet industry has resurfaced.

Traditional search engines can only display a list of webpage links. AI products like WeChat Agent and Doubao are different—they can directly pop up service cards that support order placement.

This new traffic distribution model has prompted industry discussions around a core question: If AI is responsible for screening service providers and determining display order in the future, will this mechanism evolve into a new generation of bidding rankings?

Currently, WeChat Agent is still in the beta testing and refinement phase, with some distance to go before its official launch. This article will leverage the clearly defined product architecture and recent industry trends to provide early and reasonable predictions and deductions.

WeChat Bets on Agent Ecosystem, Enabling Direct Invocation of Mini Programs

Many readers are curious whether WeChat Agent will be a standalone app or mini-program like Doubao or WorkBuddy.

Tencent has not yet announced the product form, but industry leaks suggest it is likely embedded within WeChat rather than a standalone application. The rumored scheme involves swiping right on the home page to bring up a conversation interface, where voice commands can be used to invoke various mini-programs and automatically complete tasks like ordering food, hailing a ride, or booking a hotel.

Martin Lau, President of Tencent, stated during the Q2 earnings call that Tencent has paved the way for Agent through WeChat's existing AI capabilities. Leveraging its vast user base and mature ecosystem, the agent is expected to unlock numerous scenarios, benefiting ecosystem partners. However, challenges remain, including high-concurrency inference, privacy and security, and compatibility with WeChat's unique features. No timeline for the project's launch has been announced.

Nonetheless, Tencent's top management has repeatedly and clearly communicated its strategic vision for WeChat Agent. During the Q1 2026 earnings call, Pony Ma, Chairman and CEO of Tencent, discussed the launch timeline for WeChat Agent, stating that there is no need to rush the product to market. On one hand, developers value WeChat's massive traffic Entrance , but on the other, they worry about their capabilities being simply invoked by the platform. Sufficient time is needed for product design refinement, and there is no need to rush the launch.

During the same earnings call, Martin Lau further outlined the long-term development direction of this ecosystem. In his view, WeChat could evolve into an AI-first ecosystem where users can articulate complex needs, and the Agent can complete corresponding tasks on their behalf. In the long run, mini-programs, merchants, and even ordinary users are expected to have their own dedicated agents, with different agents able to communicate and complete transactions with each other. Tencent is currently building the underlying infrastructure required for this system.

At the same time, he mentioned that project implementation requires balancing multiple factors, including cost, privacy protection, and user experience. Only when all conditions are mature can this capability unlock its corresponding commercial value.

To understand the commercial potential of WeChat Agent, one can first observe the mature business model of mini-programs, which have been operating for many years. Mini-programs have established a stable monetization framework: Relying on WeChat Pay qualifications, Tencent earns a 0.6% payment commission on every transaction completed through mini-programs. The larger the transaction volume, the higher this basic revenue.

Various types of advertisements within mini-programs, with revive and level-up incentive video ads in mini-games being the most typical, are split 70-30 between the platform and developers.

Additionally, for offline merchants without independent development capabilities, Tencent offers a SaaS-based mini-program subscription service, charging an annual service fee with additional charges for value-added features.

Combined with the annual 300 RMB certification fee for mini-program entities and technical service fees charged based on usage after exceeding the free quota for interface invocations, multiple revenue channels collectively form the massive commercial foundation of mini-programs.

However, this revenue is not separately broken down in Tencent's financial reports but is instead categorized under various business segments. This mature monetization foundation is the most important development basis for WeChat Agent.

For ordinary users, WeChat Agent represents a completely new product form and operational logic. It is not a standalone app but an AI scheduling system embedded within the WeChat ecosystem.

Users do not need to manually search or individually open mini-programs. Instead, they can simply articulate their needs in natural language, and the system can understand user intent and match mini-programs capable of fulfilling those needs.

According to the Beta access rules released by the WeChat Open Platform, there are two optional paths for mini-programs to be eligible for invocation by the Agent.

Under the automatic mode, developers enable the corresponding switch, and the platform automatically identifies the services the mini-program can provide. The development mode requires developers to write Skill documents, configure business interfaces, and submit them for approval before entering the AI invocation candidate pool.

Regardless of the chosen path, mini-programs must proactively complete ecosystem adaptation to have a chance of being selected by the Agent.

Even if a mini-program can function normally on its own, the agent cannot proactively invoke it without proper adaptation.

The entire invocation chain consists of three layers: A large model interprets user requests, an independent scheduling layer filters available mini-program skills, and finally, service cards are rendered to achieve a complete business closed loop (closed loop) for reservations, orders, and other transactions.

Accompanying this, the success of this ecosystem's commercialization hinges on whether AI agents can drive the overall mini-program transaction volume to grow further.

Based on industry-wide projections, WeChat Agent is more likely to adopt a multifaceted monetization approach in the future: Most lightweight query-based basic capabilities will remain permanently free, using these services to retain users and gradually cultivate the habit of using AI to complete tasks. For complex, high-end tasks that require significant computational resources and intricate business processes, a subscription-based or pay-per-use model will be offered to high-value users.

The primary revenue at the commercial level will still come from the B-side, relying on indirect channels such as interface invocation service fees and transaction volume commissions to complete the commercial closed loop (closed loop).

Of course, this commercialization path is not guaranteed. Actual revenue will ultimately be constrained by two key variables: user usage frequency and business conversion efficiency.

If most users only use the agent for non-transactional needs like checking the weather or generating text, and rarely proceed to place orders, the expectation of boosting mini-program GMV will be significantly diminished.

The Dilemma: Decentralized Ecosystem or Concentration Toward Top Players

This architecture faces an unavoidable issue: When multiple similar service providers are connected, the scheduling algorithm directly determines the display order of each service.

A deeper question for Tencent is whether this new distribution system will ultimately lead to a Matthew effect, with traffic flooding toward top players, or whether it will maintain the decentralized nature that has characterized mini-programs.

AI agents bypass traditional search and directly output service results based on user needs. This interaction method naturally narrows the range of services available to users.

Take ordering coffee, a high-frequency scenario, as an example. When a user instructs WeChat Agent to "help me order a cup of coffee," the product can handle the request in three ways.

The first is a fully automated closed-loop ordering process: The system retrieves the user's past consumption records and preferences to directly select a brand and complete the entire ordering process without user intervention.

The second is a single-output mode, where the AI, based on its judgment, provides only one optimal solution, displaying a single merchant to the user.

The third approach presents multiple candidate solutions, aggregating multiple merchants into service cards and returning the decision-making power to the user.

It is evident that the first two interaction logics will further amplify the traffic advantages of top brands, reinforcing the Matthew effect. Only the third approach provides a traffic window for small and medium-sized merchants.

The industry is already discussing how to counteract this tendency through algorithmic mechanisms. On the algorithmic side, weighted logic based on geography and localized reputation can be introduced to prioritize displaying high-quality local shops with good evaluations nearby, rather than solely favoring national chain brands.

A traffic exploration mechanism can also be implemented, allocating a fixed proportion of request traffic—such as 10% of invocation quotas—to non-top merchants, providing exposure opportunities for new stores and niche businesses.

At the same time, giving users control by offering preference-switching options in product settings allows them to decide whether to prioritize processing speed or explore new merchants.

On the merchant side, lightweight marketing channels can be established, allowing small and medium-sized merchants to avoid costly bidding wars and instead target potential customers nearby with promotions, gaining exposure opportunities with user consent.

However, implementing this solution is highly complex. The ranking system would need to incorporate multi-dimensional scoring items, such as diversity metrics and weights supporting new merchants.

Even if product design can rely on mechanisms like multi-option displays and diversity weights to create buffer space, the industry still debates a practical issue: Even if the platform aims to maintain a decentralized ecosystem, it is difficult to prevent traffic from objectively concentrating toward top players.

Objectively speaking, from the perspective of platform revenue structure, WeChat's commercial foundation is built on transactions generated by a vast number of merchants, not relying on a few top players. Therefore, the platform itself lacks a strong subjective incentive to deliberately direct most traffic to top players.

However, intention does not equal the final outcome. Traffic stratification may not stem from active platform regulation but rather from the implementation capabilities of merchants themselves. To join the Agent ecosystem, merchants must complete a series of adaptation tasks, including skill development and interface debugging.

Top enterprises, with mature technical teams and sufficient operational budgets, can complete integration faster and continuously iterate to optimize adaptation results.

In contrast, many local small and medium-sized merchants lack dedicated technical personnel. The threshold of adaptation development alone will screen out a significant portion of participants, creating traffic disparities even before the algorithm begins distributing traffic.

Another layer of uncertainty comes from the algorithm's inherent operational logic. Even if the platform completely eliminates paid bidding channels, the algorithm naturally favors merchants with stable transaction performance and better conversion rates when making selections.

If weight settings overly favor order volume and conversion efficiency, traffic will naturally concentrate toward top players. Only by consistently assigning sufficient weight to local supply, new store growth, and merchant diversity can the decentralized nature be preserved.

Market concerns are not unfounded. Mature aggregated food delivery platforms like Meituan have already set a precedent: Platform rankings are heavily influenced by commercial behaviors such as merchant partnership packages and traffic purchases. Paid placements directly determine merchants' exposure frequency and ranking order, giving the platform significant control over traffic.

For WeChat Agent, there are similar concerns about whether this new AI scheduling system will gradually evolve a similar commercial ranking logic.

The real test for WeChat Agent lies in finding a long-term, stable balance between operational efficiency and ecosystem richness.

Merchant ranking has always been a contentious issue, as demonstrated by the Doubao incident, where commercial interests make the public highly sensitive to ranking fairness.

Real-World Precedent: The Doubao Commission Controversy

The issue of ranking fairness discussed earlier is not merely theoretical.

The Doubao hotel channel fee adjustment incident that surfaced in late July 2024 brought the real commercial game theory (game) behind AI service distribution into the public eye for the first time, providing the industry with an observable real-world example.

On July 27, an internal channel notice from Douyin's Local Services leaked, stating that starting August 10, Doubao would be classified as an independent transaction channel, with the comprehensive fee rate for hotel orders in this channel increasing to 12%. Previously, hotel orders generated through Doubao were categorized under Douyin's organic traffic and subject to a uniform 8% fee rate.

Once the news spread, "Doubao-recommended hotels now charge fees" quickly became a hot topic. The public's primary concern was not the channel fee rate itself but whether the rate change would indirectly influence recommendation results.

Many worried that even if the platform did not explicitly sell rankings, higher commissions could still become an implicit consideration in the algorithm's traffic allocation.

In response to various external speculations, Doubao issued an official statement on the evening of August 10. The platform clarified that its Local Services segment had not yet opened paid promotion entries, preventing merchants from influencing recommendation rankings through payments. Channel service fees are only collected from merchants after a successful order. Hotel recommendation results are primarily generated based on dimensions such as geographical location, user demand match, merchant ratings, and fulfillment stability. This business is still in its early trial stage.

While both Doubao and WeChat Agent rely on AI for service scheduling, they differ fundamentally in their supply structures.

Doubao's hotel and travel transaction closed loop (closed loop) is rooted in Douyin's proprietary ecosystem, with all merchants invoked being those registered through Douyin LaiKe. WeChat Agent, on the other hand, follows an open approach, theoretically capable of simultaneously integrating multiple third-party OTA platforms like Ctrip, Fliggy, and Tongcheng in the future.

As of now, WeChat Agent, still in beta testing, has not announced any commercialization plans, nor has it disclosed details about its ranking rules. The WeChat Open Platform has only stated that mini-program developers can autonomously choose whether to join the Agent ecosystem, and participation will not affect the mini-program's original operational status.

However, following the Doubao incident, the market has begun asking in advance: After multiple OTAs complete adaptation in the future, when a user requests to book a hotel, what criteria will WeChat Agent use to arrange the order of service providers?

The Hidden Risks Behind Scheduling Authority: Avoiding a Repeat of Past Mistakes

Beyond the ecosystem fairness issues arising from ranking, the industry's greater concern stems from the painful lessons learned from Baidu's bidding ranking scandal.

The core controversy back then was not merely that merchants could pay to secure top positions. The bigger issue was the blurred boundary between advertisements and organic search results, with most users assuming that top-ranked information was more credible, ultimately leading to a series of severe problems.

Such risks do not automatically disappear with the end of the search engine era. In the context of Agent's new architecture, some contradictions may even be further amplified.

The first issue is the narrowing of user options. Traditional search engines return a vast number of webpage results. Even if ads appear at the top, users can still scroll down to view other results.

AI agents operate under a fundamentally different product logic, typically providing only a limited number of candidate service providers—or, in extreme cases, directly pushing a single service card.

It is difficult for ordinary users to know the complete list of service providers or to ascertain which options have been filtered out by the system. The information asymmetry caused by the algorithmic black box is far more pronounced than in traditional search. Even if the platform displays multiple options by default, the challenge of how to screen the candidate list remains—if the candidate pool only references global indicators such as sales volume and ratings, traffic will still continue to concentrate toward top merchants.

Figure: A screenshot of Pony Ma's latest statement

The second contradiction arises from the inherent conflicts in platform monetization. Platforms have multiple monetization paths available, including transaction commissions, service provider entry fees, and paid weighting.

Even if platforms explicitly reject the straightforward highest-bidder auction model, real-world challenges persist: platform revenue is deeply tied to transaction volume, and algorithms are more prone to favor merchants with higher conversion rates; eliminating indirect commercial influence on ranking over the long term is highly challenging to implement.

Once high-risk categories such as medical services and aesthetic medicine are integrated into the Agent system, if preliminary qualification reviews fail to keep pace, the pitfalls previously encountered by search engines are likely to reappear.

An even greater shortcoming is that industry regulations cannot keep up with the speed of product iteration. To this day, there is still no unified standard in China to define which results in Agent recommendations are natural outputs and which are commercial promotions; there are no clear, established guidelines on how commercial weighted content should be labeled or which information in ranking factors can be appropriately disclosed to the public.

Traditional search has established a mature advertising labeling mechanism, but this set of standards has not yet been successfully transferred to the new AI agent system.

AI agents have reconfigured the rules of traffic distribution, but the fairness challenges of the old era have not disappeared. Where WeChat Agent ultimately heads will test not only product capabilities but also the platform's ability to balance commercial and public interests.