The architecture is the product.
Six layers compose an Intelligent Web build. Most sites have one or two. Ours have all six. Here's what each one is, what we ship for it, and why it matters.
Semantic Foundation
Every product, person, service disambiguated.
Schema.org, JSON-LD, entity graph. The semantic foundation makes the brand readable to crawlers and LLMs as a structured set of entities, not a soup of HTML.
- Nested JSON-LD across pages, types, and relations.
- Entity graph models brand, products, people, places, services, offerings.
- Validated against the latest schema.org spec, monitored for drift.
- Tied to internal linking strategy so the structure reinforces itself.
Content Engineering
Question-first chunks, citation-ready paragraphs.
Pages restructured around how LLMs read. Short, declarative chunks. Clear authorship. FAQ engineering. Canonical answers in machine-extractable form.
- Hero, body, and FAQ sections rebuilt to citation-ready format.
- The question is the H2. The answer is the next paragraph. No filler.
- Pillar-and-cluster topic architecture supports both Google and LLM retrieval.
- Editorial standards documented so new content stays citation-ready.
AI-Discoverable Infra
Crawlers, agents, and LLMs all read the same source.
llms.txt manifest, agent-aware crawl rules, server-rendered HTML, semantic URLs. The infrastructure layer that determines whether the model sees the page at all.
- Server rendering by default. No JavaScript wall between the model and the content.
- llms.txt declares what the site is to agents, name, summary, sitemap, contact.
- Crawl directives respect Googlebot AND modern LLM agents (GPTBot, Claude-Web, Perplexity).
- URLs encode semantics, not categories. /asset-management/multi-strategy-funds, not /products/123.
Agent Layer
An embedded RAG agent. Exposed via MCP.
A retrieval agent trained on the brand's full knowledge base, docs, products, policies. Surfaces as on-site assistant. Exposes as MCP-callable endpoint for external LLMs and agents.
- RAG pipeline indexed against the brand's full content corpus.
- On-site chat surface with citation-aware responses.
- MCP endpoint allows external models (ChatGPT, Claude, Perplexity) to query the brand directly.
- Cost and quality monitored. Prompt tuning baked in.
AI Observability
Citation tracking and prompt monitoring across every model.
Continuous tracking of where the brand is cited, by which models, on which prompts, with what context. Visibility share, prompt coverage, sentiment, sourcing.
- 40+ high-intent prompts monitored across ChatGPT, Perplexity, Gemini, Claude.
- Visibility share trended over time. Alerts when something drops.
- Sample quotes logged so you can see what models actually say about the brand.
- Competitive view: who else gets cited on the prompts you care about.
Continuous AI Ops
The system stays tuned.
Monthly tuning of schema, content, agent prompts, and infra. Model-update regression checks. Competitive citation tracking. The system never goes stale.
- Monthly schema validation, content tuning, entity reinforcement.
- Agent prompt retraining as content changes, costs and quality monitored.
- Quarterly architecture review. Big stack decisions discussed before they're forced.
- On-call when a model update breaks something or a competitor lands a move.
Six layers. One system.
Each layer earns its place. Each one is monitored. None of them go stale.
The Intelligent Web stack isn't a checklist. The layers reinforce each other: the entity graph (L01) feeds the content engineering (L02), which gets surfaced through the AI-discoverable infra (L03), which the agent layer (L04) queries, which observability (L05) tracks, which AI Ops (L06) tunes.
Most sites in 2024 have a partial stack: schema bolted on, a chat widget bolted on, content written for humans. The result: invisible to LLMs and increasingly to the audiences that arrive through them.
Our builds ship all six. Day one.
Let's build the next layer.
Tell us what you're shipping. We'll come back with how we'd ship it. No deck. No pitch. Just the plan.