Marketing to AI Agents: The Second Buying Audience

AI Agents Are a Second Buying Audience. Most B2B Content Is Invisible to Them.

You now market to two audiences at once: the human buyer, and the AI agent that evaluates, shortlists, and increasingly purchases on the buyer’s behalf before a person ever talks to you. Salesforce and Gartner project AI agents will intermediate more than $15 trillion in B2B purchases by 2028, and that 90% of B2B buying will be AI-agent intermediated by then (Gartner via GSPANN, 2026). The content that wins won’t just read well to people. It has to be machine-readable, verifiable, and structured so an agent can cite it, quote it, and act on it. This guide gives you a framework, a decision matrix, and a step-by-step audit to make your content agent-ready.

Why the Human Journey Collapses Into a Single Request

The sharpest mental model for this shift is the User Journey Collapse. A human buyer spends four to eight minutes moving from a homepage to a product page to a case study, and your analytics capture every scroll and click. An AI procurement or coding agent compresses that whole multi-page journey into one or two HTTP GET requests. It fetches the page, strips the HTML, parses the raw data, and leaves in under half a second (Addy Osmani, Agentic Engine Optimization, 2026).

The result is a measurement trap. Your client-side JavaScript sees zero scroll depth, a 100% bounce rate, and zero time on page. A highly qualified buyer just evaluated you, and your dashboard recorded nothing. That is why the second audience is not just a content problem. It is an analytics and data problem too.

What Most Teams Get Wrong About Agent-Ready Content

Five mistakes repeat across B2B teams trying to prepare for AI agents.

1. Optimizing for the wrong agent. Developer-facing agents (Claude Code, Cursor, Copilot) read llms.txt, AGENTS.md, and skill.md. Search-based agents and AI Overviews read your semantic HTML and rely on core Google ranking systems, and Google says it ignores llms.txt entirely (Google Search Central AI Optimization Guide). You can’t build one file and call it done. You need both paths, and which one matters depends on your buyer.
2. Publishing bloated documents. Agents work inside context windows of roughly 100K to 200K tokens. A single oversized manual (the Cisco Firewall Guide hit nearly 193,000 tokens, or about 718,000 characters) threatens to exceed that window, so the agent truncates it, skips it, or falls back to hallucinated knowledge (Addy Osmani, 2026). Word count is no longer a pure win.
3. Leaving product data incomplete. Agents do not call to clarify. A human buyer emails sales when a spec is missing. An agent silently moves to the next supplier. Catalogs with 99.9% attribute completion see three to four times higher visibility in generative recommendations than sparse catalogs (GSPANN, 2026).
4. Tracking only client-side. Standard analytics are blind to server-side agent traffic. If you don’t look at server logs and AI referral paths, you cannot measure your agent-to-human traffic ratio at all.
5. Writing for “exhaustive” instead of “retrievable.” AI-heavy search rewards non-commodity, expert-led content with unique perspectives and clear structure, not 5,000-word puff bylines with no sourceable claim (Google Search Central, 2026).

Meet the Framework: The Two-Audience Compass

The Two-Audience Compass is a four-part test for every piece of B2B content. Run your page, product spec, or blog post through it and ask whether it satisfies both a human reader and a machine reader.

A page passes the Compass when it is readable to people, structured for machines, verifiable by sources, and retrievable in the first screenful. If any corner fails, that is where you lose agent share.

Score every corner 1 to 5 and set a pass bar of 4 before you ship a high-intent page. A developer-tool page will score high on “structured” and “retrievable,” while a brand-awareness page may score higher on “readable” and “verifiable.” That is fine: the Compass is a diagnostic, not a race to 5s. What matters is that you know which corner a given page is weak on and fix the biggest miss first. In practice, most B2B pages fail “verifiable” because their claims are unsourced, or “structured” because specs live only inside paragraphs instead of schema.org and tables.

The Decision Matrix: Which Agent Path Should You Build For?

Not every B2B company needs llms.txt and a developer directory. The right call depends on what you sell. Use the matrix below to decide, then build accordingly.

Step-by-Step: Run the Agent-Ready Content Audit

Here is the workflow I use with lean B2B teams. It takes half a day on your top ten pages and gives you a baseline you can measure month over month.

1. Check your gatekeeper. Open robots.txt and confirm you are not blocking the major AI crawlers (OpenAI, Anthropic, Google-Extended, PerplexityBot, ClaudeBot). This is the cheapest fix and the most skipped.
2. Pull your server logs for agent fingerprints. Look for the user-agents agents leave behind. axios/1.8.4 usually means Claude Code, got means Cursor, colly means Windsurf, and curl/8.4.0 often means Cline or Junie (Addy Osmani, 2026). Count direct server-side hits that never appear in your JavaScript analytics.
3. Segment AI referrals in your analytics. Build a segment for paths like claude.ai/referral, chatgpt.com/organic, and labs.perplexity.ai/referral. This is your agent referral baseline.
4. Run each page against the Two-Audience Compass. Score readability, structure, verifiability, and retrievability on a 1 to 5 scale. Front-load the answer in the first 200 words on every page where the query is a direct search intent.
5. Budget token load. Aim for quick starts under 15,000 tokens, conceptual guides under 20,000, and API reference pages under 25,000 (Addy Osmani, 2026). Split oversized monoliths into chunks that fit an agent’s context window.
6. Complete your product data. Target 99.9% attribute completion on your catalog: price, dimensions, integrations, compatibility, SLA. Move the data out of prose and into structured tables and schema.org.

Worked Example: A Hypothetical Middleware Vendor

Take AcmeGate, a fictional API gateway vendor (not a real company). AcmeGate sells to developer and platform teams. A buyer asks ChatGPT to “find me an API gateway that supports OAuth2, rate limiting, and a free tier.” Here is what happens under the old playbook versus the new one.

The old playbook: AcmeGate’s homepage is a marketing landing page with a hero, a form, and a 3,000-word blog about “why API gateways matter.” There is no llms.txt, no AGENTS.md, and the pricing page is a rendered image. The coding agent opens the repo docs, finds an oversized manual that exceeds its context window, truncates it, and cannot confirm the free tier. It recommends a competitor.

The new playbook: AcmeGate adds a root llms.txt that lists “API Gateway Documentation” with one-line outcomes (“OAuth2, JWT, rate limiting, quotas, self-hosted free tier”). It adds AGENTS.md to its repo so a coding agent instantly maps the endpoints and rate limits. The quick start is under 12,000 tokens. Pricing is a clean, structured table with schema.org markup. The agent cites AcmeGate, verifies the free tier, and shortlists it. The human never saw a “landing page,” but AcmeGate was in the recommendation.

WordLift’s published agent-SEO results show the payoff pattern: one e-commerce client saw a 30% sales boost and an 80% traffic surge after agent optimization, and a legal-services client reported 50% more visibility and triple lead growth (WordLift, 2025). The second audience is measurable once you build for it.

What the Numbers Say: The Data Behind the Shift

Here are the figures that should anchor your next strategy meeting to the same sources I use.

Chart sources: Gartner via GSPANN 2026 (90%, 33%), Gartner March 2026 survey via GSPANN (45%), Forrester via GSPANN (20%).

Token Economics Is Now a Content Metric

Treat token count like you treat word count, load time, and readability. An agent that cannot fit your content in its context window is a prospect you never reach. The practical limits are real and published.

On-device context windows run about 100K to 200K tokens, so a single overloaded document is a real blocker. The Cisco Firewall Guide, at nearly 718,000 characters, is the cautionary tale: it exceeded that window and forced agents to truncate or skip it (Addy Osmani, Agentic Engine Optimization, 2026). Budget your pages the way you budget page weight, and surface the answer to your buyer’s core question in the first 200 words.

Agent-to-Agent Is Where the Deal Happens Now

The most disorienting change is that the purchase decision no longer happens on your website. More B2B transactions are moving to an automated handshake between a buyer’s procurement agent and your pricing or inventory system. Forrester predicts 20% of B2B sellers will be forced into agent-to-agent quote negotiations in 2026 (via GSPANN). If your commerce stack relies on nightly batch syncs and cannot expose real-time pricing and availability programmatically, you are locked out of the deal before a human salesperson ever gets an email.

Mastercard and McKinsey describe the infrastructure forming around this: Model Context Protocol (MCP) to share structured context, A2A protocols for machine-to-machine coordination, and payment rails that support “Know Your Agent” and cryptographic purchase mandates. You do not need to build all of it, but you do need your pricing and product data available through an API your systems can serve without human input, and you need it accurate enough for an agent to quote on without calling to clarify.

How to Track the Second Audience (When Your Dashboard Goes Blind)

Because agents skip JavaScript, your normal analytics underestimate your real reach. Move three things into your measurement stack.

1. Server logs. Aggregate user-agent strings and IP patterns that look like machine crawlers. Log not just requests but the response code and time, so you can tell if agents can actually read your pages.
2. AI referral segments. Build GA4 (or your tool of choice) segments for the well-known AI referral domains, and treat them as their own channel.
3. Answer coverage. Track which of your priority questions an LLM does and does not cite you on, on a rolling monthly basis. This is the machine equivalent of “rankings.”

What To Do Next

Start small and prove the loop in one week. Pick your three highest-intent pages, run the Agent-Ready Content Audit above, add llms.txt and AGENTS.md if you sell developer tools, and stand up the server-log and AI-referral segments. Set a monthly cadence to re-check your answer coverage on the five questions your buyers ask first. The teams that ship this in the next three quarters will own the reference answers before their competitors even know the question changed.

Related reading on this site: What B2B Buyers Actually Do With AI (and How to Get Cited), Trust Architecture: How B2B Brands Win Credibility in the AI Era, and AI Agents for B2B Content Marketing: A Lean Team Guide.

Frequently Asked Questions

What does “marketing to AI agents” actually mean?

It means preparing your content and data so that AI systems can find, verify, cite, and act on your product information on a buyer’s behalf. The human buyer still exists, but an agent often picks and shortlists your solution before the human sees it.

Should I create an llms.txt file?

If you sell developer tools, APIs, or open-source software, yes. Coding agents read llms.txt as a low-token sitemap. If you sell to a broad enterprise audience, Google Search ignores it for AI Overviews, so focus your energy on semantic HTML and strong content instead.

Does SEO still matter if agents drive traffic?

Yes. Generative search features are rooted in Google’s core ranking and indexing systems using retrieval-augmented generation (Google Search Central, 2026). Optimizing for discoverability and being cited by AI systems are complementary, not competing, goals.

How do I measure traffic from AI agents?

The reliable path is server-log analysis. Look for agent user-agent fingerprints and direct server-side HTTP hits, then build AI-referral segments in your analytics. Client-side tools alone undercount machine traffic badly.

Is this a consumer-retail problem or does it apply to B2B?

It applies to both, but the B2B stakes are larger. Gartner projects $15 trillion in agent-intermediated B2B purchases by 2028, versus roughly $1 trillion in US B2C orchestrated revenue by 2030 (McKinsey, 2025). B2B is the bigger opportunity.

What is the cheapest first step?

Check your robots.txt to make sure you are not blocking the major AI crawlers, then pull a week of server logs for agent user-agents. Both take under an hour and cost nothing but time.