When Bad Taste Converts: Why LLMs May Outperform Designers on Conversion Creative
Why do ugly websites like Booking.com convert better?
There is older work suggesting that Comic Sans MS is one of the greatest fonts for online conversions, and that ugly websites actually convert better than pretty ones.
Booking.com is the canonical example — the interface is not pretty, but it works.
As one operator put it, "bad taste is good taste" in the sense that the only thing that matters is whether the page converts.
Taste and conversion are not the same variable, and the answer depends heavily on which industry you operate in.
There is also a deliberate version of this: some sites lean into bad copy because nostalgia is a recurring trend.
Going back thirty years and reusing that aesthetic is itself a strategy.
Can LLMs identify what creative actually converts?
Possibly better than humans.
Because taste does not equal conversion, LLMs trained on outcome data may be better positioned than designers to identify what works and what doesn't.
Tools like GoMarble act as a performance marketing optimization engine — looking at campaigns and their associated assets, then making recommendations on text and creative.
Feed those winning concepts into an image orchestrator like Weavee , or an image generation LLM like DALL-E 3 , and you can auto-generate visuals from what's already winning.
The argument is blunt: a human should not be the one going through Excel sheets or campaign reports to figure out which creatives are the most successful.
That's pattern-matching work, and machines do it faster.
“Comic Sans MS is actually one of the greatest fonts to get conversions online or that ugly websites actually convert better than pretty websites. So actually it could be that bad taste is good taste.”
When should you override AI creative recommendations?
When the AI is stuck in a local maximum.
LLMs are good at identifying what already works, but sometimes you need to "pull it out of its local maxima to try more creative approaches." This is the human's job: deciding when to break from what the data says converts and push toward something the model wouldn't have proposed on its own.
Industry context matters here — what works for Booking.com is not what works for a luxury brand, and the LLM's recommendations only reflect the patterns in front of it.
“maybe the LLMs are better than us at identifying what works, what doesn't work. But sometimes you do need to pull it out of the sort of pull it out of its local maxima to try more creative approaches.”
What needs to be in place before AI can pick winning creative?
Two years ago, prompting itself was a competitive advantage.
Today it's table stakes.
The advantage now belongs to teams with clean data, a strong semantic layer on top, and the discipline to delegate answers to machines the way they previously delegated to humans.
On the measurement side, the question is whether you're treating conversions correctly at all.
The most sophisticated setup today combines multi-touch attribution and marketing mix modeling — unified, this is called unified marketing modeling .
Worst case, you should at least have last-touch correct.
Then you can feed that data through APIs into a central engine that optimizes your campaigns for you.
“I don't think a human should be doing that. I don't think a human should be going through Excel sheets or campaign reports to find out what are the most successf”
How does the two-internet future change creative strategy?
One 2026 prediction from the AI advisory brain trust: we're moving toward two internets — one for humans, one for machines.
In practice, we already have them.
“Two years ago, like, prompting was a competitive advantage. Today, it's become kind of table.”
Frequently asked questions.
- Do ugly websites really convert better?
- There's prior work suggesting that ugly websites convert better than pretty ones, and that Comic Sans MS is among the strongest fonts for online conversion. Booking.com is the live example — the UI isn't pretty, but it works. The caveat is that this is industry-dependent: what converts in travel marketplaces is not what converts everywhere.
- Are LLMs actually better than designers at picking creative?
- Because taste and conversion are not the same thing, LLMs can be better than humans at identifying what works and what doesn't. The practical workflow is to let a tool like GoMarble analyze campaigns and assets, then use an orchestrator like Weavee or an image model like DALL-E 3 to auto-generate visuals from winning concepts.
- When should a human override the AI's creative recommendation?
- When the model is stuck in a local maximum. LLMs optimize against what they can see, so sometimes you have to pull the system out of that local maximum to try more creative approaches. The human's role is judgment about when the proven-converting pattern is no longer the right one — often driven by industry context.
- What is unified marketing modeling?
- Multi-touch attribution and marketing mix modeling, done correctly and combined, is called unified marketing modeling. It's the most sophisticated way today to define what a conversion actually is. If you can't reach that bar, getting last-touch correct is the worst-case baseline before feeding data through APIs into a central optimization engine.
- Is prompting still a competitive advantage in AI marketing?
- Two years ago prompting was a competitive advantage. Today it's table stakes. The advantage has moved to teams with clean data, a strong semantic layer, and the organizational muscle to delegate answers — the same teams that previously learned to delegate to humans are now applying the same skill to machines.
- What tools make up an AI-driven performance creative stack?
- GoMarble is a performance marketing optimization engine that looks at campaigns and their assets, then recommends optimizations on text and creative. Weavee is an image orchestrator for building image-generation workflows. DALL-E 3 is the image generation LLM itself. Together they auto-generate visuals from winning concepts without a human combing through campaign reports.