Ask a language model for a website a hundred times and you get the same page a hundred times: hero, three features, testimonials, a button. Change the prompt and it changes the words.
I spent a good part of this year building a generator that had to produce sites that didn't look like each other. Here is what worked.
Don't ask for a website, ask for blocks
The model doesn't write HTML. It fills in a blueprint, and the blueprint picks from a library of over 100 block types with several hundred variants. That one decision removed a whole class of bugs. A model can't break the layout of a block it doesn't get to write.
It also means every way of creating a site goes through one contract. The wizard, the template gallery and the "redesign my site" flow all produce the same blueprint and go through the same checks. When a fix lands, it lands everywhere.
Give each site a seed before it starts
Left alone, a model picks the safe answer. So before generating, I pick a random design seed across eleven dimensions, each with five to twenty options: type, colour treatment, decoration, motion, spacing and so on. The seed goes into the prompt as guidance, not as an order.
On top sit eight design presets, each owning a full visual language. Changing a site's whole look is one field, with no new generation.
Don't be impressed by the number of combinations. A trillion possible seeds doesn't mean a trillion different sites, which is why the checks below exist.
Check the output, because the model won't
After the model answers, a set of checks runs. Each one exists because of a failure I kept seeing.
Section order. Some sequences are banned, like hero straight into features straight into testimonials. If the output has one, the blocks get rearranged. The same check splits up identical neighbouring blocks and fixes a run of sections with the same background.
Variant repeats. The list of allowed variants is shuffled per site, so the same variant doesn't appear on every page.
Layout. A static check reads the generated HTML and CSS for patterns that are known to break at 768 and 375 px. It's far cheaper than a screenshot of every page.
Text that's too similar. This one is easy to copy. Two businesses of the same kind tend to get nearly the same paragraph, so each new text is compared with recent ones:
function shingles(text, size = 3) {
const words = text.toLowerCase().match(/\p{L}+/gu) ?? [];
const out = new Set();
for (let i = 0; i + size <= words.length; i++) {
out.add(words.slice(i, i + size).join(' '));
}
return out;
}
function similarity(a, b) {
const shared = [...a].filter((s) => b.has(s)).length;
return shared / (a.size + b.size - shared || 1);
}
It splits text into runs of three words and measures the overlap. I treat 0.4 or more as too close. Two sentences about a family bakery that differ only in the city and the ending score 0.47. It's plain string work and needs no model.
Two boring things
Every call to the model goes through one wrapper. It retries when the API says it's overloaded, waiting 2, 4 then 8 seconds. If an answer is cut off by the token limit, it retries once with one and a half times the budget. Without that, a truncated answer is a failed build.
The second is measuring. "Does this look good?" isn't a test, so the post-generation audit scores contrast, imagery, fonts and whether the services the business offers actually appear on the page. A low score goes in a report instead of becoming an argument.
What I'd tell you before you try this
Sameness is mostly a problem with the prompt's shape, not the model. A better model gives you a nicer version of the same page.
I ended up not selling sites this way. Sitely today makes custom sites and apps, with a person directing the design. The generator didn't go away. It became one tool behind that work.
