Search, Recommendations, Chat: Advertising’s Third Leap in the AI Era

Search, Recommendations, Chat: Advertising’s Third Leap in the AI Era

Key takeaways

• The product interface determines what a platform can learn: a search query, a swipe and a conversation reveal different kinds of intent.
• A recommendation algorithm becomes a business when content supply, consumer demand and monetization reinforce one another through data.
• Conversational advertising can enter the decision process before a user knows which product to buy, making commercial disclosure more consequential.
• New advertising formats depend on durable consumer behavior; inserting an old auction model into a chat interface is not enough.
• Large brands face an organizational problem as AI connects creative work, targeting, distribution and measurement across old departmental boundaries.
• A free assistant could improve advertising elsewhere in a company without directly monetizing its own conversations, but token costs and privacy still constrain that model.
• The strongest personal assistant must understand context, complete tasks reliably and earn permission to keep doing both.

I. Why has the demand stayed the same, from subway deliveries to chat assistants?

Jaggie, who leads GenAI advertising products at Meta, previously worked at Procter & Gamble and ByteDance and participated in early AI investing. At fourteen, while at school in Shanghai, she began selling on EachNet, an early Chinese e-commerce marketplace. She had first used it as a buyer; conversations with a seller made her realize she could sell too. Her first sale was a secondhand Baleno hoodie she had bought for 39 yuan and sold for 29. During summer holidays she sourced clothes from Qipu Road, a wholesale clothing district, and continued selling from high school into university.

Listings and promoted placement both cost money. Product photography meant borrowing a camera and transferring pictures to a computer. Payments and logistics were less convenient than today. She copied sellers who delivered by subway, meeting customers near station gates so neither party had to exit; sometimes she delivered by bicycle.

Some sellers rented rooms inside apartment buildings in Xujiahui rather than expensive street-level shops. Customers discovered them online and visited to try clothes on. Part of a storefront’s rent buys passing foot traffic. Once discovery comes through EachNet, Dianping, a Chinese local-business discovery platform, or another online service, the shop need not occupy the most expensive street location.

A business-education class in high school supplied theory for what she had been doing: people from multinational companies taught marketing, while software simulated purchasing, pricing and competition. P&G later exposed her to consumer brands and everyday purchase decisions. A subsequent used-car venture with older schoolmates was bruising and clarified a personal test: she wanted work she would keep doing even without being paid. How people process information and make small decisions interested her more than cars. She studied cognitive science and moved toward AI.

Around 2015, Magic, an American text-message concierge, suggested another route. Send a request for coffee; the service asks which shop and whether you want a latte or cappuccino, then passes the order to a provider. Jaggie’s team explored a conversational entry point for flowers, food and coffee deliveries.

The demand was there, but the technology was not ready. Older natural-language systems could fill slots for brand, product and address and manage a prescribed conversation with rules. Deviate from that path and a human agent had to take over. Many teams eventually moved from consumer services to enterprise customer support.

Raymond’s observation was that agents are now pursuing something people wanted eleven years earlier. Jaggie agrees: context management was always a language-processing problem. Rules could remember preferences, but open-ended situations were harder. Learning-based systems are reopening workflows that once failed.

II. Why is recommendation a feature for some companies and the whole business for others?

Jaggie resists presenting her participation in Momenta’s early investment as proof that she had predicted everything. After the assistant experiment, she still wanted to help people expend less effort on unproductive tasks and find the information they needed. Investing offered a way to see how founders and researchers approached that problem.

Her team mapped progress in speech, computer vision and natural-language processing onto finance, healthcare and autonomous driving, then spoke to researchers about starting companies. They encountered Cao Xudong and his team at Momenta, an autonomous-driving company, relatively early and liked the team.

Computer vision was particularly popular then. Security and financial-services customers had budgets, making deployment more immediate. But technical quality and a strong team do not automatically produce a strong business model. Enterprise and government work might keep a company alive without creating the rapid cash-flow flywheel of a consumer product.

Medical imaging illustrates the distinction. A company can sell AI software to hospitals or operate imaging centers itself with lower costs. The second business may look AI-intensive in its financial statements while remaining, fundamentally, an imaging-center business. Similarly, a company whose core revenue is security may possess autonomous-driving technology without making it the priority. By Series B or C, resource allocation and organizational relationships grow more complicated; the innovator’s dilemma can arrive early. A project needs an environment where it can be the first priority.

Jaggie joined ByteDance in 2017 because she preferred consumer products. Zhang Yiming’s early writing about collective intelligence described a simple ambition: help people find needed information with less effort spent expressing their intent.

Raymond distinguishes two ways investors missed Toutiao, ByteDance’s news product. Some could not see how it differed from Sina News. Others recognized recommendation but assumed a few engineers hired from Baidu could reproduce it. They missed the transition from an isolated feature to a product that could inhabit multiple content formats.

Jaggie’s point is that the algorithm is one component. Content supply, user demand and revenue, connected by data, create the business and its first-mover advantages. In her comparison, Meta is more like a social platform adding recommendation, while ByteDance starts from recommendation-based distribution and wraps it in new content forms.

Education offered another longstanding wish: why give everybody identical homework instead of recommending exercises suited to each student? AI provided new technical conditions for individualized teaching. Yet even a coherent idea, technology and commercial model remain exposed to forces outside the company. No business operates in a vacuum.

III. Why do a nine-tile grid and a full-screen feed give machines different data?

Jaggie starts advertising analysis with which consumer behaviors can be digitized. The behavior already existed; platforms could not always capture it. Television and outdoor screens cannot collect individual feedback the way digital products can. Poor measurement creates poor allocation.

Signal format matters as much as quantity. Search asks users to state a query. A short-video feed learns from seconds watched, likes, completions and swipes. The front end shapes the data available to the back end.

“Every time the way users interact with machines changes, the data you can collect changes too.”

Around 2018, people asked how Instagram’s nine-tile video grid differed from TikTok. In a grid, a user sees several options and opens one. The system learns what was chosen but may not know which of the other eight was least appealing. A full-screen feed records each video encountered and how long the viewer stayed. The same nine videos can yield different signals through a different interface.

Search and chat both accept text, but conversation exposes background, reservations and subsequent actions. Jaggie describes injuring her back, asking AI for help writing to an insurer and then a doctor, and watching rehabilitation videos. By the fourth conversational turn, a video mentioned a back-support belt, prompting questions about available products and where to buy one. The initial need was managing an injury; shopping emerged later.

That differs from searching “Whoop or Apple Watch?” The conversation explains why a purchase might happen and which concerns preceded it. The assistant’s own questions also shape what the user reveals.

Her recent switch from ChatGPT to Muse shows why consumers do not choose solely by model brand. She had repeatedly tried new tools and returned to ChatGPT, but Muse completed tasks other tools had not, including replenishing her usual eyeliner. Work on non-coding tool use and task completion can change preferences. A product that stands still for six months may lose users.

Raymond questions how durable memory-based lock-in will be if new apps can import old conversations and context. Jaggie distinguishes readable memory labels from less interpretable signals inside a system. Whether those signals can migrate is a separate question.

IV. When does a recommendation inside a conversation become an advertisement?

Raymond proposes a straightforward test: if AI influences a purchase decision and an advertiser pays for that influence, is it advertising? Jaggie sees many opportunities for exactly that: a product link, a comparison, a recommendation or a response that favors one choice. Payment to produce influence and conversion gives the interaction an advertising character.

The placement need not remain inside the chat. A recommendation page can use needs expressed earlier to suggest content or products elsewhere. The commercial surface changes, but the underlying signals come from the conversation.

Traditional advertising assumes a slot, a creative asset and a click. A conversational product can enter earlier: the user describes a problem, the AI asks follow-up questions, and a product or service appears later. That proximity to the decision also creates a stronger obligation to explain where the recommendation comes from and what commercial relationship supports it.

V. Why must advertising formats follow changes in product interaction?

Asked where a second advertising growth curve might come from for founders such as Zhang Yiming or Jack Ma, Jaggie returns to consumer behavior. Advertising is built on the native interaction of the product. Reading content, operating an app, scrolling short videos and playing interactive content create different advertising possibilities.

Some products have experimented with making every swipe reveal something playable, or turning travel information into interactive material. Snap explored mini-game advertising in which a user interacts with a brand, earns coins and eventually converts. The format differs from a conventional video ad because the underlying interaction differs.

Retention is the constraint. Someone opening an app to pass the time may not be seeking a specialized interactive format. Without a stable habit, the advertising model has little foundation.

Douyin, ByteDance’s Chinese short-video platform, provides another example through search. A search for a high-protein meal can combine summarized text, source videos and relevant products. Content, search and commerce share a surface, so an advertising opportunity need not be limited to one link or video slot.

“Changes in advertising formats require genuinely new behavior to appear in the product first.”

Only after behavior stabilizes can platforms decide how to measure and monetize it. Many AI-advertising discussions run ahead of that process: users have not formed new habits, yet advertisers already want to insert the old auction model. The result can be little more than moving a link.

Feeds, chats and search results may look increasingly similar while serving different expectations. A user may seek an answer, accept a platform’s suggestions or simply kill time. Those expectations change both advertising value and tolerance. The platform must find a commercial format that preserves the native experience; the user needs to understand why this particular recommendation appears now.

VI. Why are large brands always half a step behind?

Large companies will use AI, and people inside them want change. Jaggie questions whether the people who best understand AI and contemporary content sit inside the brand or its primary advertising agency, such as WPP.

P&G and traditional full-service agencies once organized creative production and media buying. Creators, influencers and smaller teams now contribute another set of capabilities. A large company may still have to route those teams through an incumbent agency as subcontractors, adding an organizational and judgment layer that slows decisions.

Accountability becomes harder as AI connects creative production, targeting, delivery and measurement. Old arrangements let a brand assign separate parts of an outcome to different suppliers. Someone must now understand the entire chain. Buying a tool does not establish that understanding; brand expertise alone does not prove that a new signal is reliable.

An AI training session is insufficient. Companies need ongoing experiments, reviews and adjustments tied to actual business work. Younger brands sometimes move faster because they lack inherited supplier chains and need not translate every experiment into the language of an older organization. A small budget is not automatically an advantage, but a shorter decision chain can reveal what works sooner.

Jaggie’s first question as a hypothetical P&G CEO would be whether the organization contains people capable of understanding an AI-native workflow. A budget cannot create that understanding by itself.

Founders and professional managers also face different constraints. A founder can contemplate rebuilding workflows and rehiring; a manager cannot casually dismantle and recombine departments. Their different latitude does not measure intelligence. Established companies carry obligations that prevent a clean start, which can push AI work into a separate business or prompt a founder to start again elsewhere.

VII. How can advertisers tell whether their spending was worthwhile?

The familiar complaint that half of advertising spending is wasted begins with a measurement distinction. Brand advertising and performance advertising serve different stages: awareness, interest, consideration, purchase, use and recommendation. Earlier stages are harder to measure; proximity to conversion makes attribution easier.

Purchase cycles differ too. A mattress may wait for a move. An eyebrow pencil may be an impulse discovery or routine replenishment. Even one product follows different paths in different circumstances.

China’s concentration of content, product discovery, commerce and payments inside a few large platforms makes closed-loop measurement easier. American purchases spread across Shopify, physical shops, Safeway and other platforms, complicating data governance.

Brands use sample surveys to test awareness and recall, marketing-mix modeling (MMM), and incrementality measurement to estimate channel effects. Digital advertising popularized tracking and last-touch attribution because those methods look easier, but the last click can receive disproportionate credit. A product designed to build memory and one designed to generate an order should not be judged by one metric.

Jaggie also identifies a historical data advantage: Google and Meta accumulated extensive user records before privacy rules changed. TikTok, ChatGPT and other newer platforms may have fresh user signals without the same ten-year history of an individual’s behavior. A data moat concerns the depth of knowledge about an individual, not simply user count.

Platforms therefore care about what follows an impression: a pause, another question, the transformation of a vague wish into an action. Each step creates a signal that a business can price differently.

Deeper signals create greater responsibilities. Watching for a few seconds differs from volunteering a back injury, relationship information or other private context. Treating such disclosures only as better targeting labels could improve short-term conversion while damaging trust.

This also changes the division of work between advertisers and platforms. Previously, advertisers supplied creative assets and budgets, and platforms reported click-through and conversion rates. In chat, advertisers must also consider why the user asked, how AI understood the question and when the recommendation appeared. A higher conversion rate might reflect better understanding—or a user’s failure to realize that a conversation has become a commercial process. As advertising approaches the role of an assistant, a platform needs to explain the context it used, the commercial relationship and the user’s ability to refuse or switch. Effectiveness cannot depend only on making advertising harder to notice.

VIII. Can foundation-model companies become the next advertising giants?

Jaggie would assess OpenAI or Anthropic through user time, depth of intent and data. Google can have less user time than Meta yet monetize deep search intent efficiently. Time spent alone does not determine revenue potential.

If Muse and ChatGPT had equal engagement time, depth and daily active users, Meta might still monetize more efficiently because Instagram, Threads and other products contribute longer-term signals. That is a comparison within existing advertising economics.

Model companies have other possibilities. A free product might earn nothing from advertising while subscriptions, tools and enterprise businesses support the company. Token consumption and the intent generated by free users have to be evaluated against the company’s overall economics.

The question is therefore larger than whether a particular product displays ads today. Its users, its way of understanding them and monetization elsewhere in the company jointly shape the model.

IX. How might Doubao make money?

Doubao is ByteDance’s consumer AI assistant. If many users are already turning to it for medical questions, shopping and everyday problems, Raymond asks, how might it eventually make money? Jaggie’s first concern is ethical: when users entrust AI with health, family and other private information, whether the platform should charge in those settings cannot be decided by advertising efficiency alone.

Token costs remain real. Each query incurs compute, and other businesses may have to cover costs the assistant cannot recover itself. In an extreme version, Doubao could function as a cost center or acquisition channel while improving advertising on Douyin or another company product through richer signals.

That differs from merely sending traffic elsewhere. A data flywheel changes how well the company understands the same person over time, even without moving that person to another product immediately.

Not every entry point must charge at once. Its value may emerge across the product portfolio. The commercial model still incorporates choices about trust, privacy and the company’s values.

X. Does a personal assistant’s moat come from data or trust?

Why might Google and Apple, despite abundant personal data, move slowly? Jaggie suggests Google may protect an existing reservoir of trust rather than immediately repurpose email, maps and other data. It may also believe it can enter later.

Startups move faster but face a different question: why give a newly formed company access to an inbox? Silicon Valley users may recognize its investors and product background; ordinary consumers may know only established names such as Google and Meta.

Data and trust advantages need not sit in the same place. Google has email and maps, Meta has social relationships and content, while newer assistants see conversations and task execution. Each has gaps.

“For a personal assistant to become real infrastructure, it has to complete tasks and make users believe it will not treat their lives as advertising inventory that can be sold at any moment.”

An assistant that only chats struggles to become an entry point. An effective but unsettling assistant struggles to keep its permissions. Users may choose the product that combines sound boundaries with execution rather than simply the smartest model.

Competition extends to when the assistant acts, stays quiet, answers or checks a boundary. A small team can connect new tools quickly but still needs access to email, calendars, shopping records or conversation history. A large company possesses data but also privacy commitments and internal processes. Speed reflects willingness to entrust one’s life to a product as well as engineering efficiency.

Understanding, reliability and confidence against misuse must coexist. Migration is consequently more than transferring chat logs: an assistant has learned when to ask, when to stop and which tasks previously failed. A memory file may be exportable while accumulated mutual understanding has to be rebuilt.

XI. Why has Silicon Valley started learning from China?

Raymond expected fewer investment opportunities to produce less American interest in China. Instead, people increasingly ask him about Chinese models, robots and consumer products. Jaggie separates the ability to invest from the value of learning. A fund’s local office and an LP’s allocation concern capital and operating conditions; product quality is a different question.

At Stanford in 2018, many classmates did not recognize TikTok when she said she had worked there. Some treated it as an unfamiliar, questionable Chinese company. She explained why it was fun. Product outcomes have since changed that starting point.

Previously, a few investors might explain livestreaming, e-commerce or Xiaohongshu, a Chinese social and shopping-discovery platform, to colleagues. Repeated enthusiasm for China could be dismissed as a personal preference. TikTok, Manus, short dramas and live commerce gave people results they could observe.

Supply chains, energy, chips and robotics create another reason to pay attention. Understanding robot production or bringing manufacturing back to the United States requires knowing what Shenzhen and Chinese suppliers actually do. Learning about products, manufacturing and business models need not imply direct investment.

Chinese companies are also investing more in international public relations, marketing and go-to-market work rather than treating products as purely domestic. That increases the exchange of commercial experience in both directions.

For Raymond, the conversation connects subway deliveries on EachNet to today’s chat interfaces and advertising systems. Jaggie’s underlying interest remains new forms of interaction. Advertising is an extension of consumer products; what excites her is what new user behaviors and interfaces will make possible. As she puts it, “This is only just beginning.”

Listen to the full conversation (in Chinese) on Xiaoyuzhou or YouTube — search 苔藓之火 (Mossfire).

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.