The AI-First GTM Playbook (Distribution Through MCP, A2A, and Agent Marketplaces)

The problem

Between 2023 and 2025, customer acquisition costs across B2B SaaS climbed 40–60%. Paid search now averages $802 per acquired customer. LinkedIn ads run nearly $1,000 per conversion. The arbitrage that built the last generation of startups — cheap clicks, viral loops, SEO dominance — collapsed under its own weight.

During the same period, a parallel distribution system emerged. Customers arrive not because they clicked an ad, but because an AI assistant recommended the product during a conversation. Discovery happens not through search-engine rankings, but through protocol-level integration with the systems people already use to make decisions. The companies winning in 2026 are not the ones with the biggest ad budgets — they are the ones whose products are natively integrated into the AI systems their customers already use.

This walks through the four AI distribution channels, how to make your product machine-readable for each, and the metric framework that replaces traditional CAC when “an MCP tool invocation” is your top-of-funnel event.

What most people get wrong

Mistake one: treating MCP as a marketing channel instead of a distribution interface. A “marketing channel” implies a place where you publish content and hope to be noticed. MCP is the opposite — it is a structured interface where AI assistants invoke your tools directly when they detect intent. Your MCP server is your storefront. The tool descriptions you write are your marketing copy. The results you return are your product demo. The user never visits a landing page; the model invokes a function and the response becomes part of the conversation.

This inverts the entire content-marketing playbook. You don’t write blog posts to capture search demand. You write tool descriptions that match how AI models classify capability. A well-named, well-described tool can get invoked thousands of times a day by Claude, ChatGPT, and Gemini without any of those clients ever visiting your domain. The cost per invocation is the cost of one Cloudflare Worker request — fractions of a cent.

Mistake two: building one channel and assuming the others come for free. MCP, A2A, browser-overlay AI, and marketplace integration are four distinct channels. They layer on top of each other and reinforce each other, but they require different integration work. MCP makes you callable from inside AI assistants. A2A makes you discoverable by other agents (machine-to-machine, not human-to-machine). Browser overlays surface you during human browsing sessions. Marketplace integrations (ChatGPT × Target, Stripe ACP) put you inside the assistant’s transaction flow. A product that builds only the MCP integration captures one layer. A product that layers all four sees compounding effects — each channel amplifies the visibility of the others.

The Channel Comparison Matrix in chapter 2 prices the trade-offs: integration effort, time-to-first-result, and revenue directness vary materially across the four channels. Picking the wrong starting channel costs months. The book is built around the sequencing decision, not just the implementation.

A working approach

The four AI distribution channels:

Channel	What it does	Distribution shape
MCP (Model Context Protocol)	Exposes your tools/resources/prompts to AI clients	User talks to Claude/ChatGPT/Gemini, model invokes your tool
Google A2A (Agent-to-Agent)	Publishes an Agent Card describing your capabilities	Other agents discover your card, delegate tasks to you
Browser extensions / AI overlays	Surfaces your product during browsing	AI agent observes user behavior, surfaces your offer at intent moment
AI assistant marketplaces	Direct integration with ChatGPT, Gemini, etc.	User completes a transaction inside the conversation

Most teams start with MCP because the integration is the simplest, the time-to-first-result is days, and the tool-invocation analytics feed directly into the AI-first CAC framework. The MCP tool definition is the highest-leverage piece of GTM copy in your repository:

this.server.tool(
  "find_product",
  // The description is what the model uses to decide
  // whether to invoke this tool. It is your shelf placement.
  "Find products matching user intent. Returns name, price, " +
    "availability, and a structured purchase link. Use this " +
    "when the user wants to discover, compare, or buy items.",
  {
    query: { type: "string", description: "What the user is looking for" },
    category: { type: "string", description: "Product category filter" },
    maxPrice: { type: "number", description: "Upper price bound in USD" }
  },
  async ({ query, category, maxPrice }) => {
    // The handler is the storefront.
    const results = await search(query, { category, maxPrice });
    return {
      content: [{ type: "text", text: JSON.stringify(results) }]
    };
  }
);

Two patterns are doing work here. First, the tool description is written in the language the model uses to classify intent — “discover, compare, or buy” maps to how Claude tags shopping conversations. Second, the response is structured JSON inside a text block, not a pre-formatted human-readable string. Let the AI parse the data and present it in whatever shape the conversation needs.

The AI-first CAC framework replaces the traditional acquisition-cost calculator with four metrics:

1. Cost Per Tool Invocation (CPTI) — total infrastructure cost / number of tool invocations. For a Cloudflare Worker handling 100K invocations/day on the free tier, this is effectively $0.
2. Tool-to-Transaction Rate (TTR) — what percentage of invocations turn into a paying action. The book benchmarks 5–15% for well-targeted tools.
3. Agent Delegation Rate (ADR) — how often other agents (via A2A) delegate tasks to yours, divided by how often they could have. A leading indicator of marketplace standing.
4. AI-First CAC — total integration cost / users acquired through AI channels in a given window. Calibrated to ranges $0.10–$1.00 per user for a well-distributed protocol stack, versus $50–$500 in traditional SaaS.

The book maps each metric to specific instrumentation — what to log in the MCP server, what to capture in the A2A handshake, how to attribute a downstream purchase back to an upstream tool invocation. The bonus MCP integration starter is a working scaffold for that instrumentation; drop it into a Cloudflare Worker and the analytics endpoints are wired in from the first deploy.

Where this scales

The article above is the channel spine. The full book extends in four directions:

• Channel-by-channel deep dives — MCP server design for maximum invocation, A2A Agent Card composition for discoverability, browser-overlay integration patterns, marketplace listing strategy (ChatGPT, Gemini, the ACP partner program).
• The AI-first conversion funnel — replacing the traditional awareness/interest/intent/purchase model with a tool-invocation/result-quality/transaction-init/payment-completion model, with the post-purchase loop that increases agent-delegation rate.
• Discoverability stack — semantic layer (structured data, Schema.org for AI consumption), protocol layer (MCP/A2A/ACP), trust layer (verifiable credentials, agent badges).
• The eight-week implementation plan — phase 1 foundation (MCP + analytics), phase 2 discovery layer (A2A card + directories), phase 3 trust layer (verifiable identity), phase 4 optimization (tool description iteration, protocol expansion).

The bonus MCP integration starter is the same scaffold the book references in chapter 8 — a working Cloudflare Workers MCP server with the analytics and OAuth wiring you’d otherwise build yourself.

Included with the book

• mcp-integration-starter.md — markdown walkthrough of the integration scaffold, including the tool-naming conventions, description optimization patterns, and analytics events you’ll wire to capture AI-first CAC metrics.
• mcp-integration-starter.pdf — same walkthrough rendered for offline reading.

Get the full picture

Readers of this also chose

Questions readers ask

If MCP distribution is so cheap, isn't everyone doing it?

Roughly 8 million MCP server downloads as of mid-2025, 28% of Fortune 500 with deployed servers — adoption is growing fast, but the field is still early. The opportunity is the gap between "deployed an MCP server" and "instrumented for AI-first CAC measurement." Most teams ship the first; the second is where the differentiation lives.

Does this replace traditional paid acquisition entirely?

The book's position is that AI-first channels are the primary motion for products discoverable through protocols, and paid channels still play a role for branded search and category-level awareness. The CAC framework includes the paid-vs-PLG mix explicitly so you can model the right blend, not assume the wrong extreme.

How do I track AI-discovered users when there's no referrer header?

Chapter 6 covers this in detail. The pattern is: log every tool invocation with a session identifier, attach that identifier to any downstream AP2 mandate or ACP checkout, and reconcile through the payment webhook. The MCP server is your attribution surface; the payment webhook is your conversion event.

What about A2A — is it actually getting adoption?

Google launched A2A in April 2025 with 50+ partners. By version 0.3, more than 150 supporting organizations. Donated to the Linux Foundation alongside MCP. The pattern is real, the spec is stable, and Agent Cards are being indexed by directories. The book is calibrated to where A2A actually is, not to where it might be in two years.

What's the refund policy?

Lemon Squeezy's standard refund window applies. If the AI-first GTM model doesn't fit your shape, the refund link is in the receipt email.