How the AI actually writes a page.
It isn't a single prompt fired at a language model. It's a pipeline — grounded in your real business data, structured by a tested design system, and quality-checked before you ever see it. Here's each step, in plain terms.
Not a wrapper around one model
Any AI tool can produce a paragraph of marketing copy. That's the easy 20%. The hard part is a complete page that uses real facts from your business, has an intentional visual structure, matches your brand, and catches its own mistakes before you read it. That last 80% is the pipeline.
It runs on Google's Gemini family today, with the model chosen per step — a cheap one to classify, a stronger one to generate, a vision model to check the rendered page. The routing layer is provider-agnostic and can adopt other models as they become viable. When you say build me a pricing page, none of the steps below are visible. You see: draft ready, preview attached.
Grounding: facts versus context
It quotes what it can prove, and reads the rest
Before generating anything, the pipeline assembles your business context: brand voice, design brief, knowledge base articles, uploaded documents, website pages, and CRM data. Two kinds of grounding do different jobs. Facts are the exact values the AI uses — a version number, a price, a real detail — extracted from your prompt and verified against your knowledge base. Context is the background it reads but doesn't quote. Claims that can't be traced to real source material are dropped before they reach the output. Testimonials, pricing, statistics, team members, and case studies are never fabricated — they must exist in your data.
The pipeline, stage by stage
A full page with images and quality checks typically runs 30 to 90 seconds. Most of it is invisible.
Grounding
The pipeline pulls your relevant documents, past content, branding, and CRM data, and extracts the facts the draft must include.
Structured design
It plans the page against a tested catalog of professional layout patterns — not free-form invention — and assigns visual tiers so bold, accent, and quiet sections create rhythm. No two adjacent sections share the same shape.
Parallel generation
Sections are generated in parallel for speed, but each section only sees its own data, so a pricing block can't repeat the hero headline and a quote can't be invented. The output must match a strict schema; unknown fields are rejected.
Visual QA
The pipeline renders the complete page, screenshots it, and evaluates the visual output. Layout problems and spacing issues are detected and fixed — small adjustments first, broader regeneration only if needed.
Content QA
Every text field is checked for garbage text, placeholder content, and AI artifacts, and fixed at the field level without regenerating whole sections.
Your review
The draft reaches your queue with a rendered preview. You accept, reject, or ask for changes. Nothing publishes without you.
The autonomy ladder
Four gates you open one at a time. Start with approvals — hand over publishing when you trust it.
Ideas
Mind watches your project and proposes what to publish next. On by default — thinking costs little and stays fully reversible.
Briefs
Each idea gets a plan before anything is written. Also on by default. You can review and approve a brief before it becomes a draft.
Drafts
A complete, structured draft, written without you. Off by default — turn it on per content type when the AI has earned it.
Publish
A ready draft goes live on its own. The last gate you open, per content type — and you can take it back at any time.
Translate on accept, and the honest limits
When you accept a draft, the AI fills every language your project has enabled — the English you approved becomes French, or any other locale you turned on, at the moment you accept. Single-language projects skip this entirely. The same pipeline also handles edits: it reads the current page, decides what to keep, change, add, or remove, and regenerates only the sections that change.
And the honest limits. AI output is probabilistic — quality varies run to run even with the same inputs, and the quality checks catch most issues, not all. Grounding reduces hallucinations significantly but does not eliminate them, so claim-sensitive content still needs your eyes. Legal documents, academic papers, and highly technical manuals are not reliable outputs. Review is required. That's why the approval gate is the default, and why nothing ships without you until you decide it can.
Try it on one draft
You shouldn't trust an AI with your voice at first. Ask it for one page, read what comes back, and reject it if the voice is wrong and tell it why. You earn trust by watching it hit the mark a few times. Nothing asks you to trust it before you've seen what it can do.