How the AI writes a page
A page is the output of a pipeline. It is grounded in your business data, structured by a design system and checked before you see it. Each step is below, in plain terms.
Several models, each doing one job
Any AI tool can produce a paragraph of marketing copy. A complete page that uses real facts from your business, has a deliberate visual structure, matches your brand and catches its own mistakes before you read it is a different job, and that job is the pipeline.
It runs on Google's Gemini family today, with the model chosen per step. A cheap one classifies, a stronger one generates and a vision model checks the rendered page. The routing layer is not tied to one provider and can adopt other models as they become viable. When you ask for a pricing page, none of the steps below are visible. You get a draft in the queue with a preview attached.
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 checked against your knowledge base. Context is the background it reads but does not quote. A claim it cannot trace to source material is dropped before it reaches the output. The rule for testimonials, prices, statistics, team members and case studies is that they come from your data or they are dropped.
The pipeline, stage by stage
A full page with images and quality checks typically runs 30 to 90 seconds. Most of it is invisible.
- Step 1
Grounding
The pipeline pulls your relevant documents, past content, branding and CRM data, and extracts the facts the draft must include.
- Step 2
Structured design
It plans the page against a catalogue of layout patterns rather than inventing one, and assigns visual tiers so bold, accent and plain sections alternate. No two adjacent sections share the same shape.
- Step 3
Parallel generation
Sections are generated in parallel for speed, and each section sees only its own data, so a pricing block cannot repeat the hero headline and a quote cannot be invented. The output must match a strict schema. Unknown fields are rejected.
- Step 4
Visual QA
The pipeline renders the complete page, screenshots it and evaluates the result. Layout and spacing problems are detected and fixed, starting with small adjustments and escalating to broader regeneration only if needed.
- Step 5
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.
- Step 6
Your review
The draft reaches your queue with a rendered preview. You accept, reject or ask for changes. Nothing publishes without you unless you switch publishing on for that content type.
Four switches, per content type
Each step the AI takes on its own is a setting you turn on. Start with approvals everywhere and hand over publishing when you trust it.
Ideas
Mind reads your project and proposes what to publish next. The cheapest step and the most reversible. Dismiss an idea and nothing happens.
Plans
Each accepted idea gets a plan before anything is written. Approve the plan yourself, or let the switch pass it through to a draft.
Drafts
A complete, structured draft written without you. Switch it on per content type once the drafts have shown they are worth reading.
Publish
A finished draft goes live on its own. The last switch, per content type, and you can turn it off at any time.
Translation on accept and where the pipeline stops
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 step. The same pipeline handles edits. It reads the current page, decides what to keep, change, add or remove, and regenerates only the sections that change.
The limits. AI output is probabilistic and quality varies run to run even with the same inputs, so the quality checks catch most issues, not all. Grounding cuts hallucination a long way down and does not eliminate it, so claim-sensitive content still needs your eyes. Legal documents, academic papers and technical manuals are not reliable outputs. That is why approval 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.