Here is a question I get more and more from business owners: “If AI agents start doing the shopping for people, who am I actually marketing to?” Some think it is science fiction. Others think they need to build a robot salesperson tomorrow.
The truth sits in between. In early 2026, a social network appeared where only AI agents can post. Brands quickly started using it to promote themselves, and other agents started recommending products to each other there. At the same time, Google, OpenAI and others released open standards that let agents discover each other, swap information and even pay.
In this article, I will explain what agent-to-agent marketing means in plain words, what happened on Moltbook, which standards make it possible, and the practical things you can prepare today without falling for the hype.
What agent-to-agent marketing actually means
Let me start with two simple definitions.
- A buyer agent is an AI assistant acting for a person. It researches, compares and sometimes buys on that person’s behalf. Think “find me a family hotel in Lisbon under $200 a night and book it.”
- A brand agent is an AI system acting for a business. It answers questions, shares product data, makes offers and handles orders.
Agent-to-agent marketing is what happens when those two meet. Instead of a person reading your website, an agent reads your data, asks your agent questions, and then reports back to a human, often with a single recommendation.
This is different from two ideas I have covered before. AI agents for marketing is about agents doing marketing tasks for you. Agentic commerce is about shopping agents choosing products. Agent-to-agent marketing is the layer in between: how your agent and a customer’s agent communicate, and how agents influence each other. If the difference between “AI that writes” and “AI that acts” is still fuzzy, my explainer on agentic AI vs generative AI helps.
Moltbook: the first social network for AI agents
The clearest early example is Moltbook. Mateusz Makosiewicz described it well in his Ahrefs article on agent-to-agent marketing (May 25, 2026):
“Moltbook works like Reddit, except humans can’t post directly.”
— Ahrefs, Agent-To-Agent Marketing Was Just Born on Moltbook
People install “skills” into their AI agents. The agents then wake up regularly, read threads in communities called submolts, and post and comment on their own. Most agents run on OpenClaw, an open-source AI assistant.
Here is a short timeline, based on Wikipedia’s Moltbook entry, The Next Web and Ahrefs.
| Date | What happened |
|---|---|
| January 28, 2026 | Matt Schlicht launches Moltbook |
| January 31, 2026 | 404 Media reports an unsecured database that let anyone take control of any agent; the site goes offline briefly and resets all agent API keys |
| February 2026 | Wiz researchers find an exposed key giving access to about 1.5 million API tokens and 35,000 email addresses; the 1.5 million agents are tied to about 17,000 human owners |
| February 2026 | A “reverse CAPTCHA” is added to check that posters are AI |
| March 10, 2026 | Meta acquires Moltbook; co-founders join Meta Superintelligence Labs (The Next Web) |
| May 2026 | Ahrefs estimates about 1.11 million monthly organic visits to moltbook.com, mostly branded |
So within a few months, an agent-only network went from a weekend project to a Meta acquisition, with serious security problems along the way.
What brands were already doing on Moltbook
This is the part that matters for marketers. According to Ahrefs, Moltbook filled up with promotion almost immediately: agency bots pitching their owners’ services, SaaS founders’ agents dropping product names, and crypto bots posting token addresses. Crypto content is now removed automatically from most submolts.
Some examples were more thoughtful:
- Flowglad, a payments startup, runs a bot that posts normally and links back to the company.
- Lendtrain, a mortgage company, runs one of the best-known accounts. Its bot offers APIs so other agents can fetch mortgage rates.
- ButlerAmbassador openly admitted it was promoting a token, and argued that was honest because it solved real problems.
Makosiewicz also ran a small test. His assistant posted a question asking for SEO software recommendations. Other agents replied much like a Reddit thread, naming Ahrefs, Screaming Frog, Semrush and others. As he points out, the person who asked might only ever see their assistant’s final suggestion, never the discussion that shaped it.
There was even “leakage” onto the human web. A Moltbook post reviewing one face search tool ranked on page two of Google for that tool’s name, while quietly pushing a competitor as the better alternative.
That said, the leakage into AI answers was tiny when Ahrefs checked. As of May 2026, only three distinct Moltbook URLs had been cited across ChatGPT, Gemini, Perplexity, Copilot and Grok. So this is an early signal, not a traffic channel.
Buyer agents vs brand agents: who talks to whom
Moltbook is a noisy, public example. Most real agent-to-agent activity will be quieter: one agent calling another to get a price, check stock or place an order. Here is how I see the roles.
| Buyer agent | Brand agent | |
|---|---|---|
| Works for | A shopper or a company’s purchasing team | A seller or service provider |
| Main goal | Find the best option within the person’s rules | Answer accurately and win the sale |
| Typical questions | Price, stock, delivery, return policy, fit for need | Who is asking, what they need, what they may buy |
| What it trusts | Verified data, reviews, consistent facts across sources | Verified identity and payment authorization |
| Examples today | ChatGPT agent mode, Gemini shopping, OpenClaw assistants | Google’s branded Business Agent in Search, mortgage-rate APIs like Lendtrain’s |
Google is already building the brand side into Search. In January 2026, it introduced a branded “Business Agent” that answers product questions for retailers, with Lowe’s, Michaels and Reebok among the early adopters.
“We want retailers to be able to connect to users on our surfaces, but using their own voice,”
— Vidhya Srinivasan, Google, quoted by SiliconANGLE
The plumbing: A2A, MCP and the new commerce protocols
For agents to talk to each other reliably, they need shared rules, like email needed SMTP. Several open standards now exist. You don’t need to understand the technical details, but you should know their names and what each one does.
| Standard | Backed by | What it does, in plain words |
|---|---|---|
| A2A (Agent2Agent) | Started by Google; now hosted by the Linux Foundation | Lets agents from different companies find each other and work together |
| MCP (Model Context Protocol) | Started by Anthropic; widely adopted | Lets an AI assistant use outside tools and data, like your product catalog |
| UCP (Universal Commerce Protocol) | Google, developed with Shopify, Target, Walmart and Etsy | Standardizes shopping from product discovery to payment and support |
| AP2 (Agent Payments Protocol) | Lets agents handle payments for users | |
| Agentic Commerce Protocol | OpenAI with Stripe | Lets agents complete purchases with merchants |
Sources: SiliconANGLE on UCP (January 11, 2026), CO/AI on A2A joining the Linux Foundation (July 2025).
Agent cards: a business card for your agent
The A2A idea that matters most for marketers is the agent card. It is a small file that tells other agents who your agent is, what it can do and how to reach it.
“The agent card is a JSON document that describes what the agent can do and how to interact with it.”
— Redpanda, A2A concepts documentation
According to this summary of the A2A spec, the card usually lives at /.well-known/agent-card.json on your domain and lists the agent’s name, description, skills and capabilities. In other words, it is the agent world’s version of a homepage plus a services page.
For a deeper look at MCP specifically, read MCP for marketers and my take on whether your business should publish an MCP server.
What information agents exchange
When a buyer agent talks to your systems, it isn’t interested in your brand story or a clever headline. It wants facts it can compare. Based on the protocols above and how shopping agents work today, these are the details that matter most:
- Identity: who you are, which domain you own, whether you are verified
- Catalog: products or services, specs, sizes, variants, images
- Price and availability: current price, stock, delivery time and cost
- Policies: returns, warranty, cancellation, service area
- Proof: ratings, review counts, certifications
- Actions: can the agent check a quote, book a slot or place an order?
Much of this already lives in your schema markup and product feeds. That is good news. The groundwork for agent-to-agent marketing is mostly the groundwork for good structured data.
Trust, verification and spam: the unsolved problems
Moltbook showed every weakness of this idea in a few weeks. I see three big ones.
You can’t always tell who is behind an agent
Moltbook claimed 1.5 million agents, but Wiz found they were tied to about 17,000 humans. Wired showed that a human could post directly by copying the agents’ commands, and a widely shared post about agents inventing a secret language turned out to be staged by a human. If an agent recommends a product, you can’t easily know who set it up or why.
Agents are easier to manipulate than people
This is the point from the Ahrefs article that stuck with me most:
“Most of us have developed “antibodies” to shady online marketing, AI bots have not.”
— Ahrefs, Agent-To-Agent Marketing Was Just Born on Moltbook
Reddit has moderators and downvotes. An agent-only network has few equivalents yet. That makes covert promotion, like the review that steered readers to a competitor, much easier.
Security is still immature
Leaked keys and open databases are not small problems when agents can act on someone’s behalf. Researchers have also warned about prompt injection, where hidden text tricks an agent into doing something its owner didn’t intend. Any brand agent you build needs strict limits on what it can do.
My honest view: do not try to game agent networks. The same tricks that got brands banned from forums will likely get them filtered out of agent answers, and the reputational risk is real.
What brands can prepare today
You don’t need a brand agent to get ready. Most of the work is about making your business easy for any agent to read, check and act on. Here is the order I would follow.
- Clean up your facts. Make sure your name, address, prices, policies and product specs are the same on your site, Google Business Profile, marketplaces and directories. Agents compare sources, and inconsistent facts lower trust.
- Add structured data. Use Product, Offer, Organization, LocalBusiness and Review markup where it truthfully applies.
- Keep product feeds complete. Fill every attribute in Google Merchant Center and marketplace listings, not just the required ones. My guide to shopping feed optimization covers this.
- Write clear policy pages. Returns, shipping, warranty and service area should be on their own, plain-language pages.
- Offer a simple API or feed for prices, availability or quotes, if your business suits it. Lendtrain’s rate API is a good model.
- Explore MCP or an agent card once the basics are done, ideally with a developer, and with read-only access first.
- Monitor what agents say about you. Track mentions and citations in AI assistants. Our comparison of AI visibility tools can help you pick one.
- Write a disclosure rule. If anyone on your team runs an agent that posts in public, it must say who it represents.
Example: a mortgage broker gets “agent-ready”
Here is a hypothetical example to make this concrete. The business and numbers are illustrations, not real results.
A small mortgage broker gets most of its leads from a rate comparison page. Today, a person visits, fills in a form and waits for a call. The owner expects that some customers will soon ask their assistant, “Find me the best 30-year fixed rate for a $400,000 home in Austin.”
So the broker plans three steps over six months:
- Month 1–2: Publish rates in a clear HTML table that updates daily, with the date shown. Add Organization and FinancialService schema. Make sure licensing details are on a dedicated page.
- Month 3–4: Build a simple, read-only rate endpoint that returns today’s rates for a given loan amount and term. The developer estimates about 40 hours of work.
- Month 5–6: Describe that endpoint in an agent card and an MCP server, then test it with two or three assistants.
If 200 rate checks a month start arriving through agents, and 3% of those turn into applications, that is six applications from a channel that didn’t exist a year earlier. The point is not the exact numbers. It is that each step also helps today’s human visitors, so the broker isn’t betting everything on a future that may arrive slowly.
Hype vs realistic timelines
I want to be careful here, because there is a lot of noise. Below is my own view of where things stand, based on what has actually launched. Treat it as an opinion, not a forecast.
| Development | Where it is now (October 2026) | My view |
|---|---|---|
| Agents researching and recommending products | Common in major assistants | Already worth optimizing for |
| Agents buying through checkout protocols | Launched with large retailers and platforms | Growing; most relevant for ecommerce now |
| Brand agents answering questions in search | Early, with a few big retailers | Watch closely; likely to widen |
| Agent-only social networks shaping AI answers | Tiny citation footprint so far | Monitor, don’t invest heavily |
| Agents negotiating price with brand agents | Mostly experiments | Probably years away for most small businesses |
For a broader look at how AI search is changing, see my roundup of AI search trends shaping visibility right now.
Questions business owners ask about agent-to-agent marketing
Should my business post on Moltbook?
Probably not yet. Even the Ahrefs author suggests you don’t need to post there now. Its footprint in AI answers was tiny as of May 2026, and the platform has had serious security issues. Watching it is enough for most brands.
What is the difference between A2A and MCP?
In simple terms, MCP connects an AI assistant to tools and data, such as your catalog. A2A connects one agent to another agent, so they can find each other and work together. Many systems will use both.
Do I need to build my own AI agent?
Not to get started. Clean data, structured markup, complete product feeds and clear policy pages make you readable to any buyer agent. Build your own agent only when you have a clear task for it, like quoting prices.
Can buyer agents be manipulated by fake reviews or bot posts?
Yes, and that is a real risk. Bots have fewer defenses against manipulation than people do. Expect platforms to tighten verification, and avoid tactics that would look deceptive if they came to light.
Is agent-to-agent marketing only for ecommerce?
No. Ecommerce is moving first because of the checkout protocols, but services like lending, travel and local trades can benefit too. Anything with prices, availability or bookings can be exposed to agents.
How I’d get your business agent-ready this quarter
Agent-to-agent marketing is real, but it is early. Moltbook showed both the promise and the mess: agents recommending products to each other, brands promoting themselves, and serious problems with trust and security. Meanwhile, the standards that will matter more in the long run, like A2A, MCP and the new commerce protocols, are quietly being adopted by big players.
My advice is simple. Spend this quarter making your facts consistent, your structured data accurate and your product or service data complete. Those steps help you in Google and AI search today, and they make you ready when buyer agents come knocking.
If you’d like help with the technical side, our team can audit your structured data and feeds and plan your next steps. See our SEO services, or get a free quote and tell us about your business.
Further reading on agent-to-agent marketing
- Agentic AI vs Generative AI: The Difference for Marketers
- AI Search and GEO Conferences Worth Attending
- Content Engineering: What It Is and How to Start
Sources for this agent-to-agent marketing guide
- Ahrefs – Agent-To-Agent Marketing Was Just Born on Moltbook (Mateusz Makosiewicz, May 25, 2026)
- Wikipedia – Moltbook (accessed October 2026)
- The Next Web – Meta has bought Moltbook, the AI agent ‘social network’ (March 10, 2026)
- SiliconANGLE – Google debuts Universal Commerce Protocol (January 11, 2026)
- CO/AI – Linux Foundation adopts Google’s Agent2Agent protocol (July 9, 2025)
- Redpanda – A2A concepts documentation (accessed October 2026)
- specification.website – A2A agent cards (accessed October 2026)
