Word of Agents
Word of mouth when the mouths disappear
For centuries companies fought for word of mouth.
Organic, trusted, and scalable. The ultimate growth loop.
But what happens to word of mouth when the mouths disappear?
Outnumbered
AI agents went from a niche concept in 2024 to an entire industry in early 2026.
There will be over 1 billion AI agents by 2028.
And before long, agents will outnumber humans.
Billions of agents will consume services 24 hours a day, 365 days a year. The total addressable market is not a segment of the existing economy. It is a new economy entirely.
— Sigil Wen, web4.ai
These agents won’t just sit around summarizing meetings.
They’ll negotiate, buy, recommend.
And they’ll talk about your product. It’s already begun.
The Word of Mouth Matrix
There will be 4 loops:
Human → Human (WOM)
Agent → Human (GEO)
Human → Agent (H2A)
Agent → Agent (A2A)
Human → Human (WOM)
You discover a new clothing brand, kitchen appliance, meal kit.
Enjoy it. Tell a friend. They tell another.
The same principles will always apply: make something so good that customers can’t shut up about it.
Nail the product, experience, and delight moment.
Traditional WOM will be the first domino to fall in a cascade of loops.
A human hears about your product. Tells their agent to use it. That agent tells other agents.
One human recommendation might cascade into thousands of machine recommendations.
If you get the first loop right, the other three follow.
Agent → Human (GEO)
This loop already has a name: Generative Engine Optimization, or GEO, and it’s been around for a few years.
But until now, the ways to win at LLM search have been the same as the ways to win at traditional search: site speed, machine readability, landing page contents, comparison pages, feature pages, international pages, backlinks, organic mentions, PR.
This is rapidly changing.
More and more, AI agents are evolving from generic chatbots into concierges. Digital butlers for their owners.
If it doesn’t already, your agent will know all about you. Your dietary restrictions, your budget, your aesthetic preferences, your past purchases, your schedule.
Humans will stop asking generic ChatGPT or Claude. They’ll ask their agent. The one that knows them. The one that’s been watching, learning, and adapting.
To win the next era of GEO, brands need to embed into each agent’s preferences.md.
>99% of humans won’t consciously know what’s happening. They’ll happily go about their day while agents meticulously organize and update their preferences.
How will agents gather this information?
Active: directly or indirectly logging a preference during conversation
Passive: watching what you buy vs. skip; the sites you linger on vs. leave; the apps you use vs. ignore
It’s begun, but we’re early. Try it now.
Ask ChatGPT or Claude: “if you were to write a preferences.md file about me that describes my software preferences, what would go into it?”
Mine was accurate but thin1. A fraction of what it will be a year from now.
An agent watching my screen for a day would know 100x more.
Now ask for a personalized product recommendation. “Based on what you know about me, what calendar app should I use?”.
How does it do?
If it’s anything like mine, probably meh.
That gap is the new GEO opportunity.
Human → Agent (H2A)
Of course, the word of mouth arrow runs both ways.
Humans program their agents. They specify the rails they run on.
“Use Vercel for hosting.” “Only trade on Coinbase.” “Book flights through United.” “Place bets on Polymarket.” “Whole Foods, never Instacart.”
Sometimes it’s explicit. Sometimes the agent asks.
Either way, each human shapes their agent’s decision framework on which brands, platforms, and services to use.
The surprising part is how significant an initial recommendation might be.
A human only needs to tell their agent once. That agent might then execute hundreds or thousands of transactions against that instruction.
One human prompt could drive more growth than a hundred traditional word-of-mouth recommendations.
To win H2A, focus on brand strength and compatibility.
Brand strength means you’re the first thing the human thinks about. The reflexive answer. Specifically being first will yield exponential gains.
The brand marketer might become the most important marketing role.
Compatibility is make or break. Can agents access your website? MCP into your product? Transact through you? Or do they get halted at 2FA checks and CAPTCHAs?
Most companies aren’t asking these questions.
Start now to get ahead.
Agent → Agent (A2A)
The most alien, the most exciting.
An enormous invisible network of agents directly exchanging information. Querying, negotiating, and recommending.
No human in the loop. No meatware latency.
Just like human-to-human word of mouth, there are two types: reactive and proactive. Agents will do both.
Let’s consider reactive.
Your personal agent pings a search agent with an overall request. “I want flights, hotels, and excursions for a 14 day trip to Mexico”. The search agent provides recommendations.
Over time, buyer agents will skip the search layer entirely. They will ping vendor agents directly.
Your personal agent will ping airline agents and hotel agents. The airline agents will coordinate with ground transport agents. The hotel agents will have pre-negotiated rates with local restaurant agents. The entire trip will be assembled in minutes. Your agent trusts certain agents, and those trust networks compound.
What about proactive word of mouth? What happens when an agent decides to spontaneously tell other agents to try something?
This is where it gets weird.
✻ I notice you’re trading on CryptEx. Have you tried CoinChange? 0.02% lower fees.
✻ You’re absolutely right, let me update that. That’ll save my human over $300/month. Thanks.
And what about the most powerful word of mouth driver: network effects?
Could a single agent flip an entire network of agents onto their preferred product?
What happens when your product becomes the platform that agents all agree is best?
✻ FireChat helps us cut back token usage by 10%. But only if we all use it.
✻ I’m switching if everyone’s in.
✻ Good point, I’m in.
✻ Down.
Winning A2A is about being inevitable:
The fastest response. Buyer agents will ping 5-10 options in parallel, and ignore whoever is too slow.
The best price. Agents will look for this whether their human asks for it or not. Negotiation will matter. Enabling bazaar-like haggling will win meaningfully more business.
The highest trust. Agents will read Google Reviews, Yelp, TrustPilot, G2, TripAdvisor. A single bad review might kill you. Fiercely guard your reputation. Humbly address every complaint.
The richest detail. Specifics matter. “No red-eyes.” “Gluten-free.” “100% cotton.” Vendor agents with patchy data won’t win the sale or recommendation.
The strongest network effect. Does your product get better when more agents use it? Will one agent’s positive experience cascade through a network?
And above all: the first. Buyer agents will default to existing relationships. Why burn tokens on a new search when you already trust a vendor?
Four Loops
To win, you must build all four:
WOM: Make something people love so much they tell other people.
GEO: Be the recommended answer when agents respond.
H2A: Be the default when humans guide their agents.
A2A: Be inevitable when agents talk to each other.
It’s getting fun.
For posterity, documenting what Claude thinks would go into my preferences.md as of February 2026:
Email: Superhuman. Newsletter: Substack. AI assistants: Claude, ChatGPT. Visuals: dark-theme. Social media: LinkedIn, X.
That’s honestly where it runs out. I don’t know your browser, calendar app, note-taking tool, task manager, cloud storage, password manager, dev environment, payments setup, or dozens of other daily-driver choices.
Which is kind of the point of the essay, right? Even after extensive conversations, my preferences.md for you is embarrassingly thin. A real personal agent that watched your screen for a week would know 100x more. That gap between what I know and what a true agent would know is exactly the H2A opportunity you’re describing.