Architecture
Model-Agnostic Website Architecture
The site should speak HTTP to a model. It should not import a vendor’s entire product surface into the frontend.
A thin adapter
Live features go through a small server-side function with a shared request shape: instruction, approved context, and a length limit. GPT, Gemini, Grok, or Claude can sit behind that function.
The browser never holds the provider key, and the page template does not import a chat SDK.
Context comes from the site, not the open web
The model should answer from pages, fields, and documents the client approved. Open-web free association is how a lodge website starts recommending a competitor’s season.
Retrieval, when needed, is a list of known records. It is not an excuse to scrape the internet at request time.
Evaluate with fixtures
We keep a short set of real questions and the answers a human approved. When a model or prompt changes, those fixtures get run again.
That is cheaper than discovering a bad answer in a customer email.
Continue
Using AI in the Website Build Workflow
Where GPT, Gemini, Grok, and Claude help during discovery, design, and engineering, and where they stay out of the decision.
Read →Answer Engines and Website Structure
How to structure a site so people and answer engines can use it, without writing for bots or inventing rankings.
Read →Privacy and AI Website Integrations
What should never be pasted into GPT, Gemini, Grok, or Claude, and how live AI features stay limited to necessary data.
Read →