Why Every AI-Generated Page Looks Like It Was Designed, Not Assembled
AI website generators have a reputation for producing pages that all look the same, or pages that look like nothing in particular. The fix here wasn't a better prompt — it was taking invention out of the generator's hands.
"AI-generated websites all look the same" and its opposite complaint, "AI-generated websites look like generic slop with no design sense," show up constantly in 2026 coverage of AI website builders — often from the same reviewers, describing different tools, sometimes describing the same tool on different days. Both complaints point at the same underlying failure: a generator that's inventing layout, spacing, and visual choices from scratch on every single request has no reason to be consistent, and no guarantee any individual result is any good.
Piskes' website-builder had this exact problem in an earlier form, and the fix that shipped is specific enough to be worth explaining rather than just claiming "our AI is better."
The failure mode, stated honestly
The theme-kit code that replaced the old approach describes the problem plainly in its own comments: two sites built from the same prompt could differ by a factor of ten in quality, because every generation reinvented the component vocabulary, spacing scale, and color relationships from scratch. There was nothing holding one generation to the same visual standard as the last one — the generator had complete freedom, and complete freedom is exactly what produces inconsistency.
The fix: a fixed vocabulary, not a freer model
The theme kit approach doesn't give the AI generator less power by restricting what it can write about — it restricts what it can invent visually. Instead of generating raw components and styling from nothing, the generator selects from a fixed vocabulary of tokens and components that have already been designed to work together: consistent spacing scale, a constrained set of proven layout patterns, color relationships that are already validated rather than picked fresh each time. The AI's job shifted from "invent a design" to "choose well from a set of designs that already work."
Why this produces the "looks designed" feeling
A page assembled by picking well-matched pieces from a coherent system reads as intentional in a way a page built from scratch, prompt by prompt, rarely does — because it actually is more intentional, just at a different layer. The design decisions (what spacing values are allowed, which layout patterns exist, how colors relate to each other) were made once, carefully, by the people who built the theme kit, rather than being re-decided under time pressure by a generator on every single request. Consistency isn't a side effect of a smarter model; it's a direct consequence of not asking the model to solve a problem that had already been solved.
What this doesn't fix
Constraining the vocabulary doesn't guarantee every generated page is a masterpiece — a coherent system can still produce a page that's a mediocre fit for what you actually wanted, just a mediocre fit that looks put-together rather than mediocre and visibly improvised. If the underlying complaint about AI website builders is "the output doesn't look bespoke to my business," a fixed component vocabulary won't fully answer that. What it does reliably fix is the factor-of-ten variance — you're no longer gambling on whether this particular generation happened to land well.