This comes from taking apart our own page, decision by decision. MOTO Mini is a free Japanese-style Framer template we built; the full MOTO template is sold on Framer's official marketplace, and Mini is the one-page cut of the same system. Framer users keep asking why everything the AI Agent builds looks the same. Every value below comes from the live page, not theory.
| Fix | What to change | Exact value | Why it works |
|---|---|---|---|
| Paper, not white | Background + body text | bg #F3EEE4, ink #262320 | Not #FFFFFF/#000000; paper reads as a decision |
| One typeface, one voice | Which font performs, which stays quiet | Shippori Mincho on H1, Inter elsewhere | Hierarchy is voice, not size |
| Accent, countable jobs | Where the accent is allowed | Matcha #2E4033 on 5 jobs: kanji, labels, prices, "view" marks, footer link | Unlimited use isn't an accent, it's wallpaper |
| Signature, repeated softer | How loud your motif is, each time | Kanji: large in hero, small once at grid, gone | Falling volume reads as system, not decoration |
Why AI builds all look the same
Ask a website generator, Framer's AI Agent included, for a landing page and it reaches for the same four defaults: a pure white background, one bold sans-serif carrying the whole page, an accent color sprayed across every button and icon, and a hero with a headline stacked left over two lines. None of this is a bug. The model is predicting the statistically safest layout in its training data, and safest is also average. We wrote up that mechanism separately: the reason AI websites converge. This is the fix, not the diagnosis.
Fix 1: Replace pure white with a paper tone
Change the background from pure white to a warm paper tone, and body text from pure black to a soft near-black. On our page the background sits at #F3EEE4, and the ink sits at #262320. Not #FFFFFF. Not #000000. Two hex values. Zero redesign.
Pure white next to pure black is the strongest "nobody chose this" signal a page can send: those are the values nobody has to choose, and every design tool opens with them by default. A warm paper field with a soft ink reads as a decision, because getting there required one.
Tell the Agent the exact values:
Fix 2: Let exactly one typeface have personality
Give one typeface, in exactly one place, all the personality on the page; a quiet sans carries everything else. Ours is Shippori Mincho, a serif with real stroke contrast, appearing once on the H1 headline. Everything else, subheads, body copy, labels, prices, footer, runs on Inter, a quiet grotesque sans.
When every block shouts in the same bold sans, nothing outranks anything else, which is the generator default. Picking a second typeface, and deciding exactly where it stops, is a decision the model has no reason to make on its own. Type hierarchy is voice, not size.
Tell the Agent where the personality typeface starts and stops:
Fix 3: Give your accent color a countable job list
Pick one accent color, then write down every place it's allowed to appear. If the list runs past five items, or you can't finish it without opening the page, the color isn't an accent yet.
Ours is a matcha green, #2E4033, doing exactly five jobs: the vertical kanji, the numbered product labels (01, 02, 03), the prices, the small "view" mark on each card, and one footer link. Nothing else.
The test is simple: if you can't list where your accent appears without hunting for it, you don't have an accent, you have weather. Generators spray it everywhere because "use this as the primary color" is the instruction they're given.
Give the Agent the list, not just the color:
Fix 4: Repeat your signature once, softer
Take whatever makes your page recognizably yours, a repeated mark like a rule line or a stamp, and use it exactly twice: full size once, then a smaller, quieter echo. Then stop.
On our page that's vertical kanji. 発酵美容 (fermented beauty) runs large next to the hero headline. 季節肌 (seasonal skin) runs again at the top of the product grid, smaller, set to the side. After that it's gone. No third appearance, no footer repeat.
Falling volume is what makes a signature read as a decision instead of a texture: the first appearance announces it, the second confirms it wasn't an accident, and a third would just be wallpaper. The rule works the same for any repeated mark, not just vertical type.
The instruction to give the Agent:
How to check your own page
Open your own page and run four checks before changing anything.
- Is the background pure white, #FFFFFF? Check the hex value, not your memory of it.
- How many typefaces are actually performing, not just existing? More than two is a problem. So is one doing every job.
- Can you list your accent color's jobs out loud, without opening the page? If you have to go look, the list doesn't exist yet.
- Does your signature element show up at the same size everywhere, or does it fall in volume after the first use?
Two or more failures and the page reads as generated, regardless of the content underneath it. We packaged the full walkthrough as a free teardown kit, the AI × Framer Teardown Kit: a 22-page annotated PDF plus the build-along prompts we used on our own page.
FAQ
Why do all AI-generated websites look the same?
Generators produce the statistically safest output: a pure white background, one bold sans-serif doing every job, and an accent color used everywhere instead of somewhere specific. That output is also the most average one. None of this is a bug. The tool predicts the most common pattern in its training data, not the most interesting one.
What colors stop a website from looking AI-generated?
Two changes matter most. Replace pure white (#FFFFFF) and pure black (#000000) with a warm paper tone and a soft near-black: ours are #F3EEE4 and #262320. Then pick one accent color, ours is matcha green, #2E4033, and restrict it to a specific, countable list of elements instead of spraying it everywhere. A color used everywhere signals no decision was made.
Can Framer's AI Agent produce non-generic design?
Yes, but only with real constraints in the prompt. Left alone, the Agent defaults to the generic layout it knows best. What works is a two-pass approach: structure first, sections and hierarchy, no color or font yet, then style second, exact hex values, named typefaces, and the accent's job list. Asking for structure and style in one vague prompt produces the generic result.
How do I test if my site looks AI-made?
Run four checks. Is your background literally #FFFFFF? How many typefaces are carrying real meaning, rather than just existing? More than two, or only one doing every job, is a problem. Can you say, without looking, every place your accent color appears? Does your signature element repeat at full volume everywhere, instead of once big and once smaller? Two or more failures means the page reads as generated no matter what the content says.