What a Restaurant-Search Experiment Reveals About Rebuilding Advertising and Commerce Around AI
Weisha Zhu | Scott Shields October 6, 2026
OpenAI has begun introducing advertising into ChatGPT using pricing methods long familiar to internet advertising: charging for impressions or clicks. The logic is easy to understand. Google, Facebook, and many other internet platforms have used these models for years, and advertisers know how to buy them.
But this raises a simple question: Why should the user have to click an ad?
A large language model is not a search engine. Its most important difference is that it can have a conversation with the user.
A Simple Restaurant Experiment
I tested this with an ordinary restaurant search.
I live in Hong Kong. I asked ChatGPT, “I want a Beijing-style restaurant nearby. What would you recommend?”
It quickly suggested several places. The restaurants were real, and the reasons for recommending them were not obviously wrong. But I immediately saw the problem: there were too few. Because I know the area, I knew there were other choices.
If I had been in a city I did not know, ChatGPT could have given me three real, well-reviewed restaurants and I might have thought the answer was excellent. I would not have known what was missing.
So I widened the search and asked for Shanghainese restaurants across Hong Kong. This time ChatGPT produced more than a dozen.
Then I asked another question: “Which of these are in the Michelin Guide?”
The problem appeared again. Hong Kong has notable Shanghainese restaurants listed by Michelin that had not appeared in the original recommendations. Only after that question did the model identify some of what it had missed.
This reveals an important limitation. Large models are very good at expanding from a starting point. Once an option is in view, they can rapidly compare price, location, cuisine, reviews, and atmosphere. But they do not necessarily know whether the starting point itself is right, or whether the candidate set is sufficiently complete.
That may be one of the most important current limits of large models.
The Solution Is Conversation
Yet large models already possess a capability that traditional search does not: conversation.
If I simply say, “Find me a Shanghainese restaurant in Hong Kong,” crucial information is missing. Where am I? How far am I willing to travel? How many people are dining? What is the budget? Do I want traditional or contemporary Shanghainese food? Is this a casual meal or an important dinner? Does Michelin recognition matter?
Each answer can change the recommendation.
A traditional search engine has to infer what I want from keywords and other data. A conversational AI can ask me directly.
It does not need to ask everything. It only needs to identify the unknown variables that could materially change the result. If two questions are enough, ask two.
After a few exchanges, a vague request can become something much more precise: dinner tonight for four people on Hong Kong Island, around HK$500 per person, preferably traditional Shanghainese food, with a reasonably good setting.
Now the AI can search again, check for important omissions, compare the restaurants that actually fit, and leave the final decision to me.
I say, “That one.”
At that point, the economics change.
Why Is an Ad Click Still Necessary?
Suppose that after I choose, ChatGPT can book the table for me. What has the restaurant received? Not a click, but a real reservation for four people that evening.
If a merchant receives a confirmed order through the AI, it has a clear reason to pay for that result. In situations where the transaction outcome can be directly verified, the unit of commercial value can move beyond the click and toward the result.
Pay-per-click made sense in the traditional internet advertising model. An advertising platform could send a potentially interested person to a merchant, but often could not know whether that person ultimately bought anything. The click therefore became a convenient intermediate measure.
Conversational AI may be able to extend the chain:
Need → Dialogue → Clarification → Search → Comparison → Human Judgment → Execution → Order.
If that chain can be completed, the unit of value can move further downstream. A restaurant can pay for a confirmed reservation. A hotel can pay for a completed booking. E-commerce can pay on an actual sale. Other services can pay for a verifiable appointment or transaction.
This does not mean every form of advertising should be priced by outcomes, or that attribution, cancellations, refunds, and other operational problems disappear. It means something narrower: when AI can connect a real need directly to a verifiable result, the click should no longer be assumed to be the most important unit of value.
One Line Cannot Be Crossed
There is, however, one essential condition: judgment cannot be for sale.
Which restaurants enter the candidate set, which one best matches the user’s needs, and how their strengths and weaknesses are explained cannot depend on which merchant pays the highest fee. If payment determines judgment, the most valuable part of the AI system is compromised.
Users are willing to tell an AI where they are, what they can spend, and what they genuinely prefer because they expect that information to improve the decision—not to help the platform sell them the product carrying the highest commission. Once commercial incentives contaminate judgment, users will have less reason to reveal their real needs, and the value of the conversation itself begins to collapse.
Judgment and transaction therefore need to be separated.
The judgment layer should remain independent. The transaction layer can be commercialized.
The AI can clarify the need, search as broadly as possible, and explain the differences. The human makes the final choice. After that choice, the AI can book, buy, schedule, or pay, and the merchant can compensate the platform for delivering real business.
That structure can align the interests of the user, the AI platform, and the merchant more closely than traditional advertising does.
We May Still Be Using an Old Framework to Understand a New Technology
This small restaurant experiment points to a larger problem.
When a new technology appears, the easiest thing to do is place it inside an old framework.
The search era gave us impressions, clicks, and CPC. So when large models arrive, it feels natural to ask where ads should appear, how click-through rates can be improved, and how much a click should cost.
But perhaps the question should be reversed: once large models exist, which parts of the old process are no longer necessary?
If AI can use dialogue to understand a real need, help the user search, compare, and judge, and then execute the transaction, clicking an advertisement is no longer inherently a required step. At least where outcomes can be directly verified, we should ask whether the click is still the value itself—or merely an intermediate metric inherited from the old internet.
This is also why conversation may be one of the most important and still underused capabilities of large models. Conversation does more than make a machine easier to talk to. When key conditions are unclear, the AI can ask. When its own starting point may be wrong, it can recalibrate.
The Hong Kong restaurant experiment shows that a model can produce a plausible answer without knowing whether its starting point is correct or its candidate set is complete. Simply searching more does not fully solve that problem. Future AI applications need to embed dialogue more deeply into the process of judgment, so human-machine interaction happens not only after an answer is produced, but before and during the formation of the judgment itself.
The real innovation in AI may not be making the old process faster. It may be looking at the entire process again: what should AI do, what judgments should remain with humans, and which old steps no longer need to exist?
When AI can already deliver a real order, we should at least ask again: Why should a click remain the primary unit of value?
Sometimes the real constraint on a new technology is not the technology itself, but the old framework still embedded in our thinking. The deepest change created by a new technology is often not that it makes the old world work better. It is that some steps we once took for granted suddenly become unnecessary.
