Field note · 02
What happens when agents become the internet's users?
Almost every major internet business was designed around one assumption: its user is a human.
- Google gives us a page of search results because we need a way to explore the web.
- Booking.com puts thousands of hotels in one place because we cannot visit each hotel's website.
- Review sites collect other people's experiences because we cannot try every product ourselves.
- Advertisers buy space on those pages because our attention is limited.
Now imagine I ask an agent to find a quiet hotel in Goa next weekend. It checks availability, reads reviews, compares cancellation policies and comes back with three options. I still choose where to stay. I may even enjoy looking through the photos myself. But much of the work that once took me across a dozen websites has been done by software.
That made me wonder what happens to the businesses on those websites.
What does each internet business become when the thing using it is increasingly an agent rather than a human?
The agent is still acting for me. I can ignore it, try another one or book the hotel myself. To become my usual way into the internet, it has to make the whole decision better, not merely sit between me and a website.
Search may become more valuable and less visible
Google Search solves a human problem. There are billions of pages on the web, and a person cannot inspect them all. Google retrieves relevant pages, ranks them and gives us an interface for exploring them.
That interface also became the basis of an enormous advertising business. I search for running shoes, see a sponsored result alongside organic results, open a few pages, and eventually make a decision.
Agents have the same information problem, but very different constraints. An agent can read hundreds of pages for a single search that a person would never have time to read.
Suppose I ask an agent to find the best running shoes for me under ₹8,000. It needs to know what shoes exist, which are available in my size, how they perform in the real world, how other runners rate them and which ones fit my preferences.
But it does not need Google's interface of ten results at a time.
It can search the web, inspect ecommerce listings, read manufacturer specifications, look through running forums and compare reviews. Where I might open three or four results, an agent can examine far more sources before making a recommendation.
This changes what ranking means.
A search engine largely decides which pages deserve a human's limited attention. An agent first has to decide which sources are worth examining, then which pieces of information to trust, then which products satisfy the request and finally which option is best for this particular person.
The obscure running forum I would never have clicked may contain the one detail that changes the recommendation.
Today, the search business largely looks like:
Human → Search engine → Websites
Advertising monetises the search page. A version with an agent could look more like:
Agent → Search provider → Many sources → Recommendation to human
The human may never interact with the search provider at all. Search becomes infrastructure underneath the agent.
That creates a different business layer. Agents need ways to discover the web, crawl pages, retrieve fresh information, access structured data and understand what has changed. Companies providing search APIs, search results data, crawling infrastructure, product feeds and current web data can become important infrastructure even if the people using the agents never know their names.
The business model can change with it. Instead of monetising a human looking at a search page, these providers may charge agents for queries, crawling, structured data, freshness, coverage or access to proprietary information.
Search may therefore become more valuable and less visible. The open question is whether charging machines for access to information can ever generate revenue comparable to charging advertisers for access to human attention.
Personalisation changes what “best” means. Vibes become more important.
If every agent searches for the “best” hotel or shoe, do we all end up choosing the same few businesses?
Only if “best” means the same thing for everyone. I want a quiet room and a good breakfast. Someone else wants to be close to the nightlife. A hotel can be mediocre on a general ranking and unusually good for one of us.
An agent can handle the obvious filters: price, location, availability and ratings. But if it actually knows me, it can also search for the harder-to-describe things I like. It can go through hundreds of reviews, photos and small details to find a place I would never have discovered myself.
For a niche business, the question becomes: which people am I unusually good for, and how do I prove that to their agents?
That could help smaller businesses that are exceptional for a particular kind of customer rather than good enough for everyone.
There is no guarantee it happens. If agents rely on easy signals like popularity and review counts, big brands may win even more often. Better personalisation only works if agents can find good evidence about the smaller options too.
Human evidence(reviews) becomes infrastructure
A content creator can make me want something, tell me whether it is any good or send me somewhere to buy it.
Agents affect each of these differently.
If I see someone wearing a dress on Instagram and decide I want it, the creator created the desire. My agent might then find my size and buy it somewhere else. The creator influenced the purchase, but their affiliate link was never used.
Now imagine a runner on YouTube who has tested a shoe for 500 kilometres and documented how its cushioning held up. That experience is extremely valuable to my agent, even if I never watch the video.
The same is true of reviews. If five washing machines meet my requirements, specifications can tell my agent their capacity and power consumption. They cannot tell it which one starts making a strange noise after eighteen months.
For that, it needs human experience.
It could read Amazon reviews, Reddit discussions, YouTube reviews and specialist tests, combine thousands of experiences and use them to make one recommendation for me.
In an agentic world, human experience becomes infrastructure.
But there is an obvious problem: who pays the humans and platforms producing it?
Today, my visit can create an ad view, subscription or affiliate payment. An agent could read all of that evidence, give me a three-sentence answer and send me nowhere.
We are already starting to see this tension. Reddit is restricting automated access to its data, while Amazon has fought to stop outside shopping agents from accessing its store without permission.
So agents may not be able to simply crawl everything for free. Reviews, communities and creators could eventually be accessed through licensing, APIs or paid evidence queries.
But even paying for access does not solve the deeper problem. Reddit is valuable because humans keep asking questions, disagreeing, testing things and coming back six months later to tell everyone what happened.
If agents consume that evidence without supporting the people and communities creating it, eventually there is less evidence to consume.
An internet navigated by agents still needs humans to experience things and tell one another what they learned.
Aggregators can gain usage while losing power
Booking.com does more than list hotels. It aggregates rooms, keeps prices and availability current, collects reviews and handles bookings. Hotels pay because it brings travellers.
Now imagine I ask my agent: Find us a quiet boutique hotel in Goa next weekend, under ₹15,000, near the beach and refundable.
The agent could use Booking.com to discover hotels, check reviews elsewhere, compare the hotel’s own offer and book wherever the terms are best. Booking.com’s information may become even more useful to agents, while it controls fewer of the resulting transactions.
But discovery is only part of what aggregators sell. Booking.com also handles payments and problems. Amazon goes further: catalogue, reviews, checkout, delivery and returns all work together. An agent can compare sellers, but that does not recreate the entire experience.
Agents can therefore unbundle aggregators before they replace them. Discovery and comparison may become easier to bypass; trust, payments, logistics and guarantees are harder.
More usage does not necessarily mean more power. As The Agentic Economy suggests, agents can make it easier to switch between sellers. How much power an aggregator loses depends on which of its jobs the agent can actually do.
Advertising may buy “consideration” instead of attention
An agent does not get bored and click a beautiful banner. So a shiny new thing may not attract an agent at all, but a new business still needs a way to be found.
Suppose I launch a running-shoe company with eighty customers. Nike has decades of reputation and far more independent evidence. If an agent only recommends shoes with a long public record, how does mine get a fair look, or any look at all?
So I could pay for one thing: consider my shoe too. Perhaps ₹20 for the agent to evaluate it.
The agent evaluates it alongside Nike and other candidates. It checks the price, specifications, reviews, return policy and whatever other evidence it trusts. I pay whether it recommends my shoe or rejects it. I am buying consideration, not the recommendation.
Today, advertising often looks like:
Brand → Pays for human attention → Human considers product
A version with an agent could look like:
Brand → Pays to be evaluated → Agent considers product → Human gets recommendation
This only works if payment cannot buy the answer. Samsung could pay my agent to evaluate its television, but if another television is better for me, the agent should recommend the other one. Samsung still pays for having been considered.
That creates a new question: who gets that money? The company providing my agent could keep it. A personal agent I pay for could pass some of it back to me. Or it could refuse paid consideration entirely.
Attention is paying to get in front of me. Consideration is paying to get in front of my agent.
Microsoft Advertising has described the importance of an AI shopper's shortlist, and Pierre Subeh has proposed paid consideration rather than paid selection.
Can CAC convert into customer discounts?
An agent with access to my plans could know what I intend to do before I search, scroll or buy. Side note: please listen to this comedy from Ronny Chieng. It’s actually becoming real now :)
If agents make it cheaper for brands to reach the right customers, what happens to all the money they currently spend on customer acquisition?
Suppose a television brand is willing to spend ₹5,000 to acquire me as a customer. Today, that money might go to Google, Meta, influencers, affiliates or retailers before I eventually buy the television.
But my agent already knows I want a 55-inch television under ₹60,000. The brand does not need to spend as much money finding me. It could use some of that ₹5,000 to give me a better offer instead.
Give me ₹3,000 off. Add a longer warranty. Give me free installation or a 30-day trial.
Now imagine this at scale. An agent platform knows that 40,000 people want a television. Seven thousand need one this week. Four thousand care heavily about gaming.
It can go to Samsung, LG or a new television brand and say: I have 40,000 people looking to buy. What can you offer them?
The direction flips.
Today: I have a product. Help me find people who might buy it.
With agents: I have people who want to buy. What will you offer them?
Not all customer acquisition spending will turn into discounts. Brands will still spend money to get discovered and build trust. But some of the money currently spent finding and reaching customers could instead be used to win customers who are already there.
The interesting question is who captures that money: the agent platform or the customer?
Do I pay for the agent, or does the agent pay for itself by selling to me?
If I ask an agent to find the best television within my budget for the way I actually use it, I want it to look at reliable evidence, find the best options and tell me why it chose one.
But who pays the agent for doing all that work?
A free agent could make money from merchants paying commissions when I buy something, or from brands paying to be considered. That does not automatically make its recommendations bad. Google Search can show ads and still give me useful organic results.
But there is always a tension. The agent is supposed to find what is best for me while also making money from companies that want to sell to me. If it pushes that too far and starts recommending worse products, I can leave. So it has to constantly balance monetisation with trust.
A subscription changes that relationship.
If I pay my agent, its business model can be much simpler: work for me. It can ask sellers to compete for my purchase, negotiate harder on my behalf and pass the savings back to me instead of earning a commission from the seller.
Of course, people have repeatedly chosen free products over paid ones on the internet. Agents may be no different.
But after living through the last few decades of the internet economy, which one would you rather have: an agent you pay to represent you, or a free agent that pays for itself by selling access to you?
