Ads, Agents, and What Comes Next for AI
When problems can be said in full, people say them to their assistant, not a search box. What that means for ads, and for agents.
A year ago I wrote that AI would move the power in advertising. This is where my thinking has gone since.
1. Sponsored ads on shopping sites
Sponsored ads on a shopping site are a simple business. Sellers bid on keywords. When you search, the platform mixes a few high-bidding, still-relevant products into the normal results. Amazon has run this for years. It brings in tens of billions of dollars a year and is widely seen as its most profitable business after AWS.
But it stands on one thing: people's limits. Search "air purifier" and you get dozens or hundreds of pages. I look at the first one or two. I don't have time to open each one and compare. Those first pages are what sellers fight over. What ads sell has never been the product. It's the user's limited time and attention.
Large language models loosen that limit. Muse, ChatGPT — they know who you are, how many people live in your home, what you bought last month. They don't get tired, and they can flip through every page for you in the background. More importantly, you no longer have to squeeze your problem into a keyword.
An example. A new dad hasn't slept well for weeks and has no energy during the day. In front of a search box, he first has to diagnose himself: is it the pillow, a mattress that's too hard, or should he try melatonin? Once he decides, he squeezes it into a keyword like "pillow for better sleep" and pages through results. But his real problem may not be the pillow at all. Maybe the baby needs feeding at night, work is stressful, and he scrolls his phone before bed. A search box can't hold any of that. It only knows keywords. Zeng Ming makes this point in his book Intelligence: in the past, consumers did most of the work of defining their own problem.
With an agent, he can just say: "I've been tired lately and I can't sleep." The agent knows he just had a baby and has been working late. It helps him sort out the problem first, then decides what to suggest. Maybe a pillow. Maybe a sleep mask. Maybe just taking turns with his partner for night feedings. Search records where your attention went and answers "what kind of person are you." An agent cares about "what problem are you facing."
Once a problem can be said in full, people will say it where they're understood best. Nobody explains why they can't sleep inside a shopping app. They tell the assistant that's been with them all year. Amazon did build Alexa Shopping, but it's hard to use, and it's in the wrong place: it waits at the shelf, while intent gets spoken one level up. Whoever catches the intent can make better recommendations — and an ad placed there makes more sense too, because it answers a specific problem, not a guessed label.
Amazon's choice has been to defend. It blocks ChatGPT and Muse from its site. That's a kind of protectionism: the side that builds a tariff wall is usually the side no longer sure it can win in open competition. The wall can keep agents out. It can't stop users from speaking somewhere else. I raised this internally twice — once last year, and again this year when Muse came out — and nothing moved in that direction. People are still fighting for fractions of a percent in the next experiment, inside those one or two pages of results.
Without a big change, here's the ending I see: Amazon retail turns into the warehouse behind someone else's assistant — it has the goods, the shipping, the prices, but no front door for traffic. Advertisers follow intent and go bid at the level above. Zeng Ming said on a podcast that old platforms rarely become new platforms. I hope that isn't true for Amazon. But guarding the gate isn't the answer.
2. What comes next for AI
So what grows at that level above? I used to think this wave of AI was nearly over. After Zhang Xiaojun's podcast interview with Zeng Ming, I realized we're far from the end. Large models are the steam engine, the electricity, the bottom layer of the internet. What gets built on top — the light bulb, the fridge, the washing machine; the browser, the search engine; let alone truly native apps like TikTok — hasn't arrived yet.
Most of what we do now is take an existing workflow, remove the human, and put AI in. That isn't AI-native thinking. AI-native means building the scaffolding around AI, so that a whole workflow runs on it end to end.
Zeng Ming also said agents are about to explode in number — so many we can't count them. Then we'll need a recommendation system for agents. And recommendation needs benchmarking first: run agents across all kinds of tasks to learn which tasks belong to which agent. It's a bit like the App Store: once there were too many apps, people needed charts and recommendations to find the one they wanted. Agents are harder to try, though. You only learn whether one is good by actually handing it a task and seeing the result. Users don't have the time or money to try them one by one. In the end, it's the same reason ads exist: resources are limited, so someone has to pick for you by some standard. Ads were always a kind of recommendation. Once shopping-site ads drift off true north, the next star to steer by may be agent recommendation.
Even picking the right agent only gets you halfway. Today's model plus harness plus agent has friction everywhere. Autonomy and accuracy aren't good enough, so a human has to keep watching. I call it botsitting. Zeng Ming writes in Intelligence that until AI can take responsibility for results, humans have to decide, confirm, and catch what falls. Parallel work can't scale forever; the bottleneck is the human. The more you do, the more tired you get.
To let the human step away, you have to use the data each task leaves behind to train the agent better. Collect traces while the agent runs, clean them, study them, and use them to improve it. Once it improves, rerun the benchmark to check. Once it passes, start the next round. That's how an agent gets smarter, one vertical at a time. And when it's time to recommend across agents, the benchmark gets called back as the basis for the recommendation. Once that loop works, humans can truly leave the loop: agents do the well-defined work, and people go pick the next problem worth solving.
I believe that over the next two or three years, nobody gets around this.
Originally written in Chinese on November 1, 2026. The earlier piece: From Fooling People to Fooling AI.