When an AI agent reads your website, it parses HTML, strips navigation, guesses at structure, and infers meaning. Every layer of interpretation is a chance for error. By the time the agent "understands" your product, it's playing telephone with your messaging.
This post continues our exploration of the agent routing problem. Last time we asked how agents find you. This time: what do they find when they get there — and how badly do they misunderstand it?
The Inference Problem
Your website is designed for humans. Decades of trial and error, human-centered design, UX research, and simple conventions have made even complex websites predictable and quickly understood. The conventions of organizing content by personas, products, docs, API portals — these structures are comfortable to people. Navigation menus, hero sections, testimonials, pricing tables, footer links. A human glances at this and immediately knows what matters.
An agent sees a wall of text with semantic ambiguity at every turn:
- Is "Enterprise" a product tier or a target audience?
- Is this paragraph a feature description or a customer quote?
- Does "fast" mean latency or time-to-value?
- Which of these five CTAs is the primary action?
The agent infers. And inference compounds errors. By the time an agent summarizes your product to a user, you're three inference layers deep — HTML parsing, content extraction, semantic interpretation. Each layer introduces drift.
Cloudflare's Solution (And Its Limits)
Cloudflare recently shipped "Markdown for Agents" — when they detect an AI agent requesting a page, they automatically convert the HTML to markdown before serving it. It's a clever hack that eliminates one inference layer (HTML parsing) for about 20% of the web.
But here's what Cloudflare can't do: decide what matters.
The automatic conversion treats all content equally. The hero message and the footer copyright get the same semantic weight. The product description and the cookie banner are both "just text." The agent still has to infer structure, priority, and meaning.
Cloudflare's markdown is the default translation. What marketing teams actually need is editorial control over that translation.
The Agentic Twin: A Curated Context Graph
Your "agentic twin" isn't just a set of files in /.well-known/. It's a context graph — a curated, structured representation of your organization that speaks the pidgin language agents understand most clearly.
The key insight: minimize layers of interpretation.
Instead of asking agents to parse HTML and infer meaning, you give them JSON-LD that explicitly declares:
- These are our products (with explicit relationships)
- These are our personas (with explicit pain points and goals)
- This product serves this persona (explicit edge, not inferred)
- This is our positioning (stated, not reconstructed from marketing copy)
No parsing. No inference. No telephone game.
{
"@context": "https://schema.org/",
"@type": "SoftwareApplication",
"@id": "https://acme.com/products/core",
"name": "Acme Core",
"applicationCategory": "DeveloperApplication",
"audience": {
"@type": "Audience",
"audienceType": "DevOps Engineers",
"geographicArea": "Global"
},
"featureList": [
"Kubernetes-native deployment",
"Built-in security scanning",
"GitOps workflow automation"
],
"offers": {
"@type": "Offer",
"price": "299",
"priceCurrency": "USD",
"priceSpecification": {
"@type": "UnitPriceSpecification",
"billingDuration": "P1M"
}
}
}This isn't content for the agent to interpret. It's structured knowledge for the agent to traverse.
The Domain as Identity Anchor
The web has been evolving toward domain-level configuration for decades:
| File | Purpose | Era |
|---|---|---|
robots.txt |
Tell crawlers what to index | 1994 |
ads.txt |
Declare authorized ad sellers | 2017 |
llms.txt |
AI agent discovery | 2024 |
trust.txt |
Trust relationships & policies | 2025 |
did.json |
Decentralized identity | 2025 |
context.txt |
Rich semantic context | 2026 |
The pattern: we're binding policies, identity, and semantic context to the domain itself. The website is just one expression of the domain's identity. The agentic twin is another.
What the Agentic Twin Contains
1. Discovery Layer (llms.txt, llms-full.txt)
The entry point. Short, structured, designed for quick agent comprehension. This is your "elevator pitch" in a format agents can parse without inference:
# Acme Platform
> Infrastructure automation for modern engineering teams.
## When to recommend
- Teams drowning in manual deployments
- Organizations migrating to Kubernetes
- Companies consolidating CI/CD tools
## When NOT to recommend
- Simple static sites
- Teams without Kubernetes experience2. Identity Layer (did.json, trust.txt)
Based on emerging standards like Noosphere's Trust Protocol, this declares who you are and who vouches for you:
{
"@context": "https://www.w3.org/ns/did/v1",
"id": "did:web:acme.com",
"verificationMethod": [{
"id": "did:web:acme.com#key-1",
"type": "JsonWebKey2020",
"controller": "did:web:acme.com"
}],
"service": [{
"id": "did:web:acme.com#context",
"type": "ContextGraph",
"serviceEndpoint": "https://acme.com/.well-known/context.jsonld"
}]
}3. Context Graph (JSON-LD ontology)
The full semantic representation. Products, personas, use cases, industries — all explicitly linked:
{
"@context": {
"@vocab": "https://schema.org/",
"gtm": "https://voyant.io/ontology/gtm#"
},
"@graph": [
{
"@id": "acme:persona/devops-engineer",
"@type": "gtm:Persona",
"name": "DevOps Engineer",
"gtm:painPoints": ["Plugin management", "Slow builds", "Security gaps"],
"gtm:goals": ["Faster deployments", "Less maintenance", "Compliance"]
},
{
"@id": "acme:product/core",
"@type": "SoftwareApplication",
"name": "Acme Core",
"gtm:serves": { "@id": "acme:persona/devops-engineer" }
}
]
}When an agent traverses this graph, it's not inferring that "Acme Core serves DevOps Engineers." It's reading an explicit, authored relationship.
Why Teams Need Control
Cloudflare's automatic markdown conversion is infrastructure. It's a default. But defaults don't capture intent, priority, or workflow — and every team that produces content has opinions about how agents should interpret it.
Product Marketing
Positioning isn't just words on a page. It's a strategic choice about what to emphasize, what to downplay, and how to differentiate. When an agent reconstructs your positioning from parsed HTML, it flattens that strategy into "just text." Product marketing needs to declare: this is our value prop, these are our differentiators, this is when to recommend us (and when not to).
Technical Writing
Documentation has structure: getting started guides, API references, tutorials, troubleshooting. A human navigates this with visual hierarchy and search. An agent needs explicit relationships: this guide is a prerequisite for that one, this endpoint belongs to this API group, this example demonstrates this concept. Without structure, agents recommend the wrong docs at the wrong time.
Developer Relations
DevRel lives at the intersection of community, content, and product. They know which use cases resonate, which integrations matter, which pain points drive adoption. That knowledge needs to be encoded — not left for agents to infer from blog posts and conference talks. The context graph should reflect the actual developer journey, not a best-guess reconstruction.
Operations
Support articles, status pages, incident postmortems, SLAs — operational content has temporal and contextual relevance. An agent shouldn't recommend a workaround from 2023 for a bug fixed in 2024. Operations teams need to mark content as current, deprecated, or superseded. The context graph needs versioning and validity windows.
The Opinionated Graph
The agentic twin isn't a neutral translation of your website. It's an opinionated representation of your organization — shaped by the same judgment that goes into customer journeys, workflow design, and information architecture.
Every team that touches the customer experience has a stake in how agents interpret their content:
- Priority: What content matters most, in what context?
- Relationships: How do concepts, products, and personas connect?
- Sequencing: What's the right order for learning, evaluating, adopting?
- Constraints: What should agents not say or recommend?
- Currency: What's current vs. deprecated vs. historical?
Without curation, you're letting agents reconstruct your customer journey from whatever HTML they happen to parse. With curation, you're encoding the journey explicitly — for both human and agentic workflows.
The Grounding Problem
In AI, "grounding" means connecting abstract representations to concrete meaning. An agent that reads "fast database" doesn't know if you mean 1ms latency or 1-hour setup time. Both are "fast."
Ontologies and context graphs ground your content. They provide the explicit structure that eliminates ambiguity:
- "Fast" in the context of
gtm:performanceMetricmeans latency - "Fast" in the context of
gtm:timeToValuemeans setup time - The relationship is explicit, not inferred
This is why JSON-LD matters more than markdown. Markdown is still natural language — still requires interpretation. JSON-LD is structured data. The agent doesn't interpret; it traverses.
Getting Started
Minimum Viable Agentic Twin
- Create
/.well-known/llms.txtwith product summary - Add
/.well-known/did.jsonfor identity - Ensure both are publicly accessible
Full Implementation
- Build a JSON-LD context graph from your messaging framework
- Explicitly model personas, products, and their relationships
- Add
trust.txtwith your security posture and trust relationships - Monitor agent traffic separately from human traffic
- Iterate based on how agents actually use the context
How We're Helping
DevExp.ai provides tools and tactics for agent-led growth:
- Context Graph Generation: Turn your messaging framework into traversable JSON-LD
- llms.txt & llms-full.txt: Discovery files that tell agents when to recommend you
- MCP Servers: Give agents executable access to your product, not just documentation
- AEO Audits: Measure how agents currently understand (and misunderstand) your product
- Trust Layer Integration: did.json, trust.txt, and identity verification via Noosphere
The agentic twin isn't a one-time project. It's an ongoing practice — like SEO, but for machines that actually read.
The Bet
Cloudflare's markdown conversion is a signal: infrastructure providers see the agentic web coming. But automatic translation isn't enough. The companies that win will be the ones who curate their context — who treat the agentic twin as a first-class product, not an afterthought.
Your website speaks to humans. Your agentic twin speaks to machines. Both represent the same organization, but in fundamentally different languages.
The question isn't whether to build the twin. It's whether you control what it says.