Has OpenAI's 100-billion advertising narrative suffered a 90% shrinkage?

OpenAI's advertising business has recently hit a major roadblock. According to its original expectations, this business was supposed to deliver considerable performance, but in reality, its long-term revenue forecast has been drastically cut, with the reduction potentially reaching as high as 90%.

This judgment comes from the latest data forecast for the U.S. chatbot advertising market released by third-party research firm eMarketer.

The trillion-dollar narrative is more like an impossible mission

To understand why the figure of "90% reduction" is so impactful, we first need to compare the "blueprint" OpenAI presented to its investors with the much less optimistic reality seen by third-party institutions.

In April this year, OpenAI put forward an extremely ambitious goal during its roadshow: to reach $2.5 billion in advertising revenue in 2026, and this figure will surge to $100 billion by 2030.

Just a few months later, eMarketer released another set of forecasts, stating that the size of the entire U.S. chatbot advertising market will be around $10 billion by 2030. Even if OpenAI takes the entire market share, it will still be 90% short of the trillion-dollar promise it made.

Over the past two years, ChatGPT has indeed attracted hundreds of millions of users, and AI chat has gradually become a new internet entry point. However, when compared to the entire advertising market, there is still a significant gap between AI chat products and the traditional search system represented by Google, whether in terms of user base, usage frequency, or the traffic supply that can carry advertisements.

More critically, the statistics from eMarketer do not only cover ChatGPT, but the entire U.S. chatbot advertising market, including major players such as Microsoft Copilot, Google Gemini, and Amazon's shopping assistant.

In other words, even if OpenAI becomes the biggest winner in the AI chat market in the future, it will only compete for the AI chat advertising cake, not the entire digital advertising market.

If we continue the deduction, to achieve this goal, OpenAI needs to meet at least three conditions that have barely appeared in the history of internet advertising at the same time.

First, massively shift search and display advertising budgets from Google and Meta.

Second, build a near-monopoly market share in the highly competitive AI chat track.

Third, AI conversational advertising must create higher commercial efficiency than search advertising, feed advertising, and even short video advertising.

If any of these three conditions cannot be met, the trillion-dollar growth story will be difficult to sustain. Beyond these problems, returning to ChatGPT itself, there is a more inherent dilemma waiting for it.

The inherent contradictions of ChatGPT

Internet advertising over the past two decades has been mainly built on two mature models.

One is the "intentional advertising" represented by Google. When a user searches for "which noise-canceling headphone is good", there is a clear purchase intention behind the behavior, and the advertisement that appears there essentially helps users make decisions. The other is the "attention advertising" represented by Meta and TikTok. Users scroll through videos, and the platform algorithm continuously pushes products according to their interests, creating demand in the process of entertainment.

But ChatGPT fits neither of the two models. Most people who open ChatGPT type their first query as "help me write an email", "explain this piece of code" or "summarize this paper". These demands point to productivity: users come here to get an answer and complete a task, not to browse around or compare prices.

It is naturally short of the consumption signals that search advertising relies on to survive. Just imagine: when someone is modifying a PPT, an advertisement for office software suddenly pops up from AI; or when debugging code, a server advertisement is inserted into the conversation. The commercial value may not be improved, but the experience is damaged first. For a product that emphasizes efficiency, any information unrelated to the task may become a distraction.

More troublesome is that what ChatGPT sells is not just answers, but also trust. Google has been a search engine from the very beginning, and users have long expected to see advertisements in search results. But ChatGPT wants to be a personal AI assistant, and users default that the answers it gives are the optimal solutions based on content quality, not based on who pays more.

Once this boundary becomes blurred, problems will arise.

For example, if a user asks: "Recommend the most cost-effective noise-canceling headphone." If ChatGPT prioritizes answering about paid brands instead of the product it truly considers the most suitable, users will question not just this recommendation, but whether the entire AI assistant is still trustworthy. For AI, trust itself is part of the product. And advertising can easily become a variable that undermines this trust.

There is also a third hurdle: brand safety. Global advertisers such as Procter & Gamble, Unilever, and Disney care most about a controllable and safe communication environment when buying traffic.

The biggest feature of generative AI is that its output is probabilistic. No brand wants to see its advertisement followed immediately by an AI response that contains factual errors, value deviations, or even illegal content.

In the final analysis, what OpenAI faces is not just the design challenge of advertising products, but a more underlying paradox: the harder it works to become a trustworthy AI assistant that can help improve people's efficiency, the harder it is to mechanically apply the advertising logic of the internet over the past two decades. At the same time, in addition to the contradictions at the product level, there is also a huge mountain that is difficult to cross.

The advertising system itself is a high wall

Many people think that advertising is nothing more than inserting a sponsored content into the response. If it were really that simple, OpenAI's advertising would have been launched long ago. The real difficulty is that advertising is a whole set of complex infrastructure. What is truly difficult for Google and Meta to be replicated is the advertising system they have built little by little over many years.

When advertisers want to run ads, they need a back-end management system; when they want to improve ROI, they need precise targeting; when they want to continuously increase their budget, they need to rely on real-time bidding, performance attribution, automatic optimization, payment and settlement, plus a bidding market formed by the participation of millions of advertisers.

All these elements combined constitute the moat of Google Ads and Meta Ads today.

There is even a saying circulating in the industry: "What is really difficult to do is not AI, but AdTech". In technical communities such as Hacker News and Reddit, many ad tech practitioners have put forward a point of view: "OpenAI has the world's top AI scientists, but it has underestimated another thing: ad tech is an 'old wine' that takes ten or twenty years to brew slowly".

For Google's current advertising system, well-written code is only one aspect, and the more critical point is that enough people are using it. Millions of advertisers bid non-stop every day, tens of billions of ad requests continuously feed the training model, and tens of billions of clicks correct the conversion rate again and again. Every delivery optimizes the entire system in reverse: the more it is used, the more accurate it becomes; the more accurate it is, the more people use it.

However, OpenAI does not have such capabilities at present. In terms of advertiser scale, Google and Meta have millions of small and medium-sized enterprise advertisers, and most of the budgets are automatically delivered through self-service platforms. OpenAI currently mainly relies on a small number of brand customers, as well as partners such as agencies and DSP platforms to test advertising products. With limited ad inventory and a limited number of advertisers, it is naturally difficult to form a mature bidding market.

In terms of performance attribution, more and more brands today do not just want exposure when they run ads, but to prove how much sales every penny of the budget has brought. Google knows whether a user clicks on an ad, visits the website, and finally makes a purchase. Meta also knows whether a user converts after seeing an ad.

But most of the behaviors in ChatGPT take place in a closed conversational environment, it is very difficult for advertisers to track whether a user really goes to Taobao or Amazon to place an order after reading the AI recommendation.

No attribution means no ROI. No ROI means it is difficult to form a continuous budget.

OpenAI is starting to make money in a different way

In response to this, OpenAI is also gradually adjusting its direction. In the past six months, it has begun to shift its focus from the "single advertising model" to the two-wheel drive of "advertising + transactions".

On the one hand, the advertising business is indeed being implemented rapidly. It has successively recruited the head of programmatic advertising from The Trade Desk, and also introduced the former CMO of ServiceNow to complement the capabilities of brand cooperation and business teams. It launched ChatGPT ad testing in the United States in January 2026; in May, it officially opened the self-service ad management platform to enterprises; in June, ads were fully launched for global free users.

On the other hand, it is laying out "transactions" and putting more bets on Agents.

Compared to awkwardly inserting an ad content in the chat box, the business model of Agent follows a different logic. When a user says "Help me book the cheapest flight to Chicago next week, and then book a Marriott hotel", the AI in the future will not throw out several links, but directly complete the entire transaction process. For OpenAI, the source of revenue will also change accordingly: it will no longer charge by exposure, but charge a commission on transactions. Advertising makes money from attention, while Agent makes money from transactions.

This also represents a more imaginative future. In fact, as early as the end of 2025, OpenAI launched the "Instant Checkout" function for a trial run. Although it was terminated for some reason, its intention to explore the transaction model is very obvious.

Not long ago, it expanded its strategic cooperation with Visa, embedding Visa's payment network into the OpenAI platform, so that AI agents can directly complete shopping and payment; it also jointly developed the "Agentic Commerce Protocol" with Stripe, issuing one-time payment tokens to allow AI agents to safely complete transactions on behalf of users.

Once this model works, OpenAI's role will change from an "ad distribution platform" to a "transaction matching and execution platform". In this regard, domestic AI manufacturers have gone further. Qwen and Doubao are both trying to bind the ecosystem of "AI entry + transaction scenarios", directly converting conversational demands into e-commerce or local life orders.

Qwen is connected to Taobao, Doubao is connected to Douyin, and Yuanbao is connected to JD.com. In July this year, JD AI Agent and Yuanbao completed the interconnection of the mini-program ecosystem. When users consult products in Yuanbao, the system will directly pop up JD product cards, and they can click to jump to the mini-program to place an order. Recently, Qwen also launched a new round of testing: brands such as Luckin Coffee and Mixue Ice Cream & Tea began to cultivate users' usage habits again in the form of "coupons".

Conclusion

So, is OpenAI's trillion-dollar advertising story over?

Not necessarily. But the eMarketer report gives the market a reminder: user growth does not necessarily bring advertising growth automatically. Advertising is only one path for AI commercialization, far from the only answer.

For OpenAI, this means that the business model must be rebalanced. On the one hand, it will continue to increase subscription revenue such as Plus and Pro, and accelerate the expansion of enterprise services and API business. On the other hand, it will put more commercial bets on new models that are more in line with the logic of AI products, such as Agents, e-commerce, and enterprise automation.

This article is from the WeChat official account "wj00816" (ID: Morketing), author: Alan Wang Jingxing, published with authorization from 36Kr.