At a recent guest talk, I filled a few pages with shorthand notes, and one phrase kept pulling me back: "market to agents." At first it sounded almost like a category error. We market to people. Agents are software. But the more I thought about what a personal agent could actually do with enough context, the less strange it seemed.
The drive
Think about this: you're driving somewhere with your family. It's getting close to dinner, your car will need gas soon, and your personal agent already knows your route, your schedule, and what everyone usually likes to eat.
So it suggests two stops. The nearest is a gas station where you can fill up and grab snacks. A few minutes farther along the route, there's a bigger station where you can fill up and also get a dinner everyone will actually eat.
The first stop solves the problem you would have searched for. The second fits the whole situation, and you probably wouldn't have found it yourself, because you were thinking about gas, not dinner. The agent combined those needs before you had even turned them into separate searches.
What changed in that answer
My first reaction was that this was a nice recommendation feature. I now think that undersells it.
When I search for "gas station near me," the results can only respond to what I typed. If I also want food my family will eat, I have to run a second search and then do the combining myself: which of these is on my route, which is open, which is worth a few extra minutes. Most people in a moving car skip that work. They pull into the first station and buy whatever is on the shelf.
The agent does the combining before I see anything, and that changes which options exist for me. The bigger station a little farther away isn't simply ranked higher in a list I was already looking at. In a plain search for gas it wouldn't have looked especially relevant, and in a separate search for food it wouldn't have looked relevant to "on my route, tank almost empty." It only becomes the best answer when everything about my situation is in the same place.
My notes from the talk put it more compactly: "options are given by known context." The context decides which choices show up before any ranking happens. In marketing terms, the agent is building the consideration set.
How a personal agent builds your options
Think about it more generally: you and your family are going out, and your agent already knows your preferences, timing, and constraints. Before you see a single suggestion, this happens.
Your context
- timing and schedule
- where you are
- what everyone likes
- budget and constraints
Business information
- Noodle barhours · menu · price
- Caféhours · menu · price
- Pizzeriahours? · menu is a photo
Personal agent
reads what businesses publish and matches it to you
Consideration set
only what the agent could read and match
Options shown to you
- Caféfits everyone
- Noodle baropen later
If a business can't be read
Does not enter the consideration set
The business loses exposure
You lose a possible option
What I mean by a personal agent
Not every AI agent is a personal agent, and the difference matters here. A chatbot mostly responds to what you type. A task agent can carry out a bounded job, like filling in a form or comparing a few products, and then it's done. A personal agent is meant to stick around: it keeps context about you across time and across the apps you connect, and it can act or go looking for things on your behalf without being walked through every step.
That category is no longer hypothetical. Meta launched Muse in early September and describes it as a personal AI agent that works across a person's apps, remembers what matters to them, keeps working after the app is closed, and asks for approval before it sends an email or makes a purchase. Shopping has quickly become one of its main uses: it browses retail sites, suggests listings, and prepares a checkout for you to confirm, according to CNBC. Google's Gemini Spark, announced in May, is pitched in similar terms: an agent that runs in the cloud around the clock, connects to apps like Gmail, and checks with you before spending money.
I don't know whether either could handle my driving example today, but this is the kind of product where it would live. In this version of agentic commerce, the agent does the looking and the person approves.
The other side of the counter
Now take the gas station's point of view.
The bigger station gets suggested only because the agent can tell that it exists, that it's on the route, and that it sells food this particular family likes. If that information is missing or stale, hard to retrieve or interpret, or simply unavailable to that agent, the station becomes much less likely to make the shortlist. The family drives past. Nobody at the station ever learns it lost a sale, because the customer never knew it was an option.
This is where the notes on that second page sit. Written down in order, they read almost like an argument: stores inject their information into agents so they can participate in the flow; small businesses especially; market to agents; agents are the ones "shopping"; if the agent can't find the option, less exposure. That's my shorthand of what I heard, not the speaker's exact words, but the sequence is clear enough.
The phrase I keep coming back to is "legible to agents." A business can be excellent in person, with good food and friendly staff, and still be close to invisible inside an agent's decision space. The way I'd put the chain now: the quality of a business's information, whether an agent can retrieve and interpret it, and whether the agent has access at all together decide whether the business can be considered. That decides whether it enters the consideration set, and the consideration set decides its exposure.
Access is the part I hadn't thought about at first, and it means legible to agents isn't always the same as open to agents. About two weeks after Muse launched, Amazon started blocking it from shopping on Amazon.com, after asking Meta to leave Amazon out. People who tried saw a notice that access by an unauthorized AI agent violates Amazon's Conditions of Use. Amazon has no shortage of structured product data. In that case the AI shopping agent could have read the store fine; the store chose not to let it in. A small restaurant with a photo for a menu has the opposite problem. Either way, the shopper ends up with one less option and probably never notices.
Why small businesses may feel this first
The talk flagged small businesses specifically. The reasoning that follows is my own.
Many larger retailers already maintain structured catalogs, inventory systems, product feeds, and people whose job is to keep that data clean. A lot of small businesses keep their most important facts somewhere else. The menu is a photo on Instagram. The hours are on a listing nobody has touched in a while. What's actually on the shelves this week lives in the owner's head.
People are surprisingly good at piecing those fragments together. We squint at the menu photo, skim a few reviews, and make a guess. I'm much less sure an agent will do that reliably, or bother, when a competitor's information is already clean and complete. If agents favor what they can verify, being good may not be enough to make the shortlist.
That bothers me a little. It's a new kind of unfairness, one that rewards the most readable station, which isn't always the best one.
Is this just SEO again?
Partly. Businesses already put a lot of effort into being found by search engines, and more recently into showing up in AI-generated answers, which people have started calling GEO. Some of that work will carry over. Clear, consistent, machine-readable business information helps whether the reader is a search crawler, a language model writing an answer, or an agent putting together a shortlist.
The difference I notice is in what the customer sees. With search, a business that ranks seventh is still on the page, and a customer who scrolls can find it. An agent that hands you two options in a moving car gives you no page to scroll. If an agent returns only a few options, a business outside that consideration set is effectively invisible, and the customer never sees what was left out.
So I think the job moves from ranking well toward being included at all. For agent discoverability, that means making sure an agent can tell what you sell, whether it's available, what it costs, and who it's for, and that it can actually reach you to check.
What I'm not sure about
It's easy to take one vivid example and build a whole future on it, so I want to be careful.
The driving example is my own hypothetical, built from what I took away from the talk. The talk was looking ahead, and even with products like Muse and Gemini Spark now available, I don't know how much of this agents can do reliably today. The context that makes the agent useful (location, family preferences, schedule) is exactly the data people are nervous about handing over, and privacy was in my notes too. I don't know whether people will trust a shortlist they can't see past. And "injecting information into agents" could mean many different mechanisms, none of which I could pin down from my notes.
The Amazon dispute also left me with a question I didn't have before. If an agent decides what makes my shortlist, I also want to know whose interests shaped that shortlist.
None of that makes the basic chain less interesting to me. If agents start doing more of the searching, the businesses they can't read, or can't reach, will get less exposure. The ones that lose most may be small businesses that never had a reason to structure their data.
A note on sources. The ideas about discovery moving to AI, personal agents that draw on a person's wider context, "market to agents," and reduced exposure for businesses an agent can't find come from my handwritten notes of a recent guest talk. Short phrases in quotation marks are from my notes, not verbatim quotes. The driving example is my own hypothetical. Details about Muse, Gemini Spark and Amazon's block are linked where they appear and come from the companies' own announcements and news reporting. The framing around consideration sets, legibility and access, the small-business reasoning, the comparison with SEO, and the conclusions are mine.