AI website builders look like magic from the outside: you type a sentence about your business, wait sixty seconds, and a finished website appears. The actual mechanics are more interesting than the marketing copy suggests, and understanding them helps you write better prompts, get better drafts, and know when the AI is going to fail you.
The pipeline, end to end
Every modern AI website builder runs the same four-stage pipeline. Prompt parsing turns your one-sentence description into structured intent (industry, audience, tone, key offering). Content generation produces the actual copy — headline, sub-head, feature bullets, about section, calls-to-action — using a large language model conditioned on that structured intent. Asset selection picks photographs, icons, and colour palettes that match the inferred industry and brand mood. Layout assembly drops the content and assets into a template, applies the colour scheme, and renders the final HTML.
The interesting choice in this stack is whether each stage is one model doing everything or several specialised models. Single-model approaches are simpler to run and faster on first generation but produce more generic output. Multi-stage pipelines (which OneClick uses) take a little longer but produce drafts that feel more deliberately authored because each stage can be tuned independently.
Why three drafts instead of one
Most builders show you a single output. OneClick shows three. The reason is statistical: a language model's first sample is rarely its best — it tends toward the safest, most generic continuation of the prompt. Asking the same model three times with different temperature settings produces three meaningfully different drafts, and the one you actually pick is usually not the first.
The cost of generating three drafts at once is modest (mostly LLM tokens) and the conversion benefit is large enough to be obvious in usage data. Multi-draft is a small technical decision that compounds into materially better outputs.
How the AI knows what your industry looks like
The language model has a rough world-model of what websites in different industries say — dentists talk about cleanings and insurance, restaurants list hours and menus, SaaS startups talk about workflows and integrations. That world-model comes from the model's training data, which included tens of millions of real websites. The output is essentially a statistically average website for your stated industry, which is exactly what most users want for a starting point.
This is also why the AI fails in unusual niches. If your business is genuinely novel (combining two industries, targeting an audience the model has not seen, using terminology that did not exist when the model was trained), the first draft will read as a vague version of the closest familiar niche. Fix this by editing aggressively or by giving the AI more context up front.
Photographs: not generated, selected
Almost no production AI website builder generates photographs from scratch in 2026. Image generation is expensive, slow, prone to anatomical errors, and licence-ambiguous. Instead, builders query a library of licensed stock photography (Pexels, Unsplash, sometimes proprietary collections) with keywords derived from your prompt and pick a small set that match the industry and mood. The result is real photographs of real people that you can legally use, at the cost of some homogeneity across users in the same niche.
If the AI keeps picking generic photos that do not fit your specific business, swap them manually from the editor — every builder lets you do this and it is the single biggest improvement most drafts need.
Tokens, credits, and what you are actually paying for
The variable cost of running an AI website builder is dominated by language-model API calls. Most builders count "AI credits" or "AI words" against your plan; OneClick counts the number of words the LLM actually generates and writes to your draft. A full first-draft generation uses 800 to 1,500 words depending on prompt length and the chosen layout. A subsequent regeneration of a single section uses 100 to 300. Edits you make by hand cost nothing.
This is also why builders generally do not let you regenerate infinitely on the free tier — the underlying API cost is real, and the free tier subsidises generation hoping you upgrade once you commit to the site.
What AI builders cannot do well, yet
Three things. First, factually-grounded copy — the AI will invent statistics, awards, and customer quotes if you ask it to. Always replace these with your real numbers before publishing. Second, brand voice — the AI can be steered toward "playful" or "professional" but cannot capture the specific voice of a real human writer. Third, true originality — the output is, by construction, a recombination of training data, so two users in the same niche with similar prompts will get drafts that share phrases.
For each of these, the right move is: use the AI for the draft, then edit. The edit takes ten minutes and is the difference between a forgettable site and one that converts.