The second audience
The 4Ps, positioning theory, brand building, the purchase funnel, pricing psychology, content strategy, distribution. All of it was built on how people think, feel, compare, and decide.
When someone asks Claude "what is a good tool for checking my brand's AI visibility," Claude does not browse Smithery, read landing pages, or feel impressed by a clever tagline. It searches its training data, evaluates structured tool descriptions, checks for consistent entity information across sources, and surfaces whatever matches the intent most precisely.
When a developer adds an MCP server to their workflow, the agent evaluating that server does not care about your brand story. It reads the tool schema, checks whether the capability description matches what the user needs, and either calls the tool or moves on.
Your product now has two audiences. Most founders are marketing to neither of them well.
What agents are (and are not)
Agents are not people with different preferences. They are a structurally different kind of audience.
Humans discover through search, social media, word of mouth, and community. They evaluate through narrative, social proof, visual design, and gut feel. They decide based on emotion, trust, identity, and (sometimes) rational comparison. They remember brands as bundles of associations, feelings, and expectations.
Agents discover through registries, tool descriptions, package managers, llms.txt files, and training data. They evaluate through capability matching, reliability signals, and structured data. They decide based on computed fit, verifiable claims, and consistent entity information. They "remember" brands as weighted patterns in their training corpus.
The mechanisms are different at every stage. Discovery, evaluation, decision, retention. This is not a minor variation. This is a different species of audience behavior.
The dual-audience framework
Every marketing decision now has two audiences. The substance must be consistent across both. The language is different.
| Domain | Human Track | Agent Track |
|---|---|---|
| Audience | Who are the humans who will buy? | Which agents will discover and recommend? |
| Positioning | Evocative: triggers associations, identity | Declarative: states capabilities, category, differentiation |
| Messaging | Narrative, emotional, progressive disclosure | Structured claims with verifiable evidence |
| Channels | SEO, social, communities, word of mouth | MCP registries, llms.txt, AGENTS.md, package managers, AEO |
| Content | Blog posts, videos, case studies | Tool descriptions, API docs, JSON-LD, schema markup |
| Brand | Visual identity, voice, emotional associations | Structured data, entity consistency, reliability signals |
| Pricing | Psychological framing, anchoring, tiers | Transparent, machine-readable, usage-based |
| Measurement | Traffic, conversion rates, NPS | Share of Model, registry rankings, agent recommendation frequency |
A human reads "ScrappyCMO helps you get discovered." An agent reads {"category": "marketing-intelligence", "capabilities": [...]}. Same truth, different language.
The classic frameworks, reconsidered
Product: two surfaces, one truth
Your product now needs two interfaces. The human surface: UI, narrative, onboarding, emotional design. The agent surface: API, MCP server, structured data, tool descriptions, documentation quality.
These are not two products. They are two interfaces to the same product. A human reads "We help you understand your competitive positioning" and decides whether it sounds credible. An agent reads compare_competitors(brand, competitors): CompetitiveReport and evaluates whether the schema matches its user's intent.
Price: the end of pricing psychology (for agents and bots)
Humans evaluate price through cognitive biases: anchoring, framing, loss aversion, round-number effects. These biases are real and well-documented. They work.
Agents have no cognitive biases. An agent evaluates price as a numerical input in an optimization function. It compares your price against every alternative it can find, instantly, without anchoring effects.
You need both: psychologically framed pricing for the human, transparently structured pricing data for the agent. Usage-based and outcome-based pricing models are inherently agent-native. Agents understand "$0.01 per analysis" more naturally than "$299/month for the Pro plan with 47 features."
Place: two channel maps with zero overlap
Human discovery channels: Google search, social media, communities, conferences, word of mouth, review sites.
Agent discovery channels: MCP registries (Smithery, mcp.so), llms.txt files, AGENTS.md, package managers (npm, pip), Share of Model (what AI models already know about you), JSON-LD and schema markup.
These channel maps have almost no overlap. "Omnichannel" used to mean "be on both email and social." Now it means "be on both TikTok and Smithery."
Promotion: from persuasion to proof
Human promotion relies on attention, interest, desire, and action. It uses emotional appeals, storytelling, social proof, and urgency.
Agents do not respond to persuasion. They respond to proof. Clean data. Verifiable claims. Reliable fulfillment. Low friction. An agent evaluating your tool checks: does it do what the description says? Do the specs match the behavior? How easy is the integration?
This does not mean you stop telling stories to humans. It means you also need a parallel body of evidence that agents can parse.
Brand: the invisible layer
A human brand is a mental shortcut: trust, expectations, identity, emotional associations. Brands reduce the cognitive load of evaluation.
Agents do not have brand affinity in the human sense. But they have something functionally similar: weighted reliability signals. When a brand consistently appears in positive contexts across training data, the model mentions it more frequently in recommendations.
Here is the critical difference: agent "brand preference" only survives when there is a demonstrable, measurable difference. Brand premium built purely on narrative and emotional association does not survive agent evaluation.
Your brand, as seen by agents, is the sum of structured data that describes what you are, what you do, and how well you do it. JSON-LD, entity consistency, registry metadata, training data footprint. This invisible layer is the machine-readable brand. Building it is now a marketing activity.
Positioning: explicit vs. evocative
Human positioning is evocative. It works by triggering associations beyond what is literally said.
Agent positioning is declarative. It works by explicitly stating what is true. You cannot rely on agents to read between the lines. They parse what you state. If your tool description says "helps with marketing," an agent has no way to know whether that means email automation or competitive positioning.
Both positioning statements must be consistent in substance. The encoding is different. The truth underneath is the same.
The dual-audience model in practice
Here is what this looks like for a founder who shipped a product:
Human track
- Write a homepage that tells a clear story: what this is, who it is for, why it matters
- Pick one or two channels and show up consistently
- Create content that demonstrates expertise and builds trust
- Design pricing that signals value
Agent track
- Write tool descriptions and registry metadata that precisely state capabilities
- Add llms.txt and AGENTS.md to your documentation
- Implement JSON-LD schema markup on your website
- Check your Share of Model: when someone asks an AI about your category, do you get mentioned?
Most founders do the first list partially. Almost none do the second list at all. The ones who do both will own their categories in the next two years.
Why this matters now
Agent-mediated discovery is growing fast.
AI-referred traffic converts at over 10%, the highest of any channel. ChatGPT referral traffic to major brands grew 40% month-over-month through 2025. MCP registries passed 12,000 servers. The volume is still small compared to Google, but the trajectory is vertical.
The infrastructure exists.
MCP is the transport protocol. Smithery, mcp.so, and MCPize are the marketplaces. llms.txt and AGENTS.md are the documentation standards. Everything needed to build, distribute, and monetize agent-facing products is live today.
The pieces exist. The framework does not.
Plenty of people are writing about AEO, GEO, Share of Model, and machine-readable brands as individual topics. Nobody has synthesized these into a unified framework. Dual-audience marketing does not have a textbook or a practitioner community yet. It will. The question is who defines it.
What this is not
This is not about using AI to do marketing. It is about marketing to AI. Different problem. Different discipline. Different skill set.
And it is not about choosing one track over the other. The human track is not going away. People still make purchasing decisions. People still respond to stories, trust, and community. The dual-audience framework does not replace human marketing. It extends it to cover the other half of your audience that did not exist three years ago.
Where to start
If you have shipped a product and done none of this, here is the sequence:
- Positioning. Define what you are, for whom, and why you instead of the alternatives. Do this twice: once in human language (for your landing page), once in machine language (for your tool descriptions and registry listings).
- Audience. Define who your human buyers are and which agents will discover you. These are different questions with different answers.
- Messaging. Write the five artifacts you need: README first paragraph, Product Hunt tagline, Hacker News title, homepage H1, and a structured tool description.
- Channels. Pick one human channel and one agent channel. Go deep on both before adding more.
- Measurement. Track both tracks. Human: traffic, conversion, engagement. Agent: Share of Model, registry rankings, AI citation frequency.
ScrappyCMO exists because this framework needed tools, and the tools needed a framework. We built both.