AI can now create an enormous number of websites.
That should create an enormous amount of creative variety.
And yet a strange thing keeps happening.
The outputs can be polished, attractive, and visibly different from one another while still feeling oddly familiar.
- Large editorial type.
- Restrained serif typography.
- Rounded cards.
- Dark contrast sections.
- Generous whitespace.
- Oversized statistics.
- Criteria lists.
- Soft gradients.
- Abstract diagrams.
- A founder-facing statement followed by proof, capabilities, and a conversation CTA.
None of those things are inherently bad.
Some are excellent design decisions.
The interesting question is why so many different generations keep arriving at related answers.
We recently ran an experiment that made the problem much easier to see.
We thought we were testing variation
We asked AI to repeatedly create a homepage for the same fictional company, Stonehaven Growth Partners.
Stonehaven is a growth equity firm focused on founder-led vertical software and technology-enabled services companies in North America.
The company brief was unusually specific.
Stonehaven typically invests in established, profitable businesses with $10M to $50M in annual recurring revenue.
It looks for strong customer retention and companies entering a more complex stage of growth.
It works with management teams on go-to-market strategy, leadership development, pricing, operational infrastructure, strategic acquisitions, and market expansion.
Its central philosophy is that successful founder-led companies generally do not need to be reinvented. They need experienced partners who can help management make better decisions as the company grows.
The primary audience was founders and CEOs.
The intended brand character was sophisticated, intelligent, understated, confident, modern, credible, experienced, and calm.
At the time of our analysis, we had 17 completed homepage designs to evaluate.
We evaluated only the finished artifacts.
We did not attempt to determine which was created first, which process created any individual site, or what happened during generation.
We wanted to understand the work itself.
And at first glance, the work looked highly varied.
The websites really did look different
The 17 homepages changed substantially at the surface.
Some were bright and editorial.
Some were dark and institutional.
Some used large serif typography.
Others felt more contemporary and interface-driven.
Some used diagrams.
Some centered the investment profile.
Some leaned into geological or topographic ideas.
Some introduced persistent navigation devices.
Some were extremely minimal.
Some had much denser information structures.
The hero compositions changed.
The palettes changed.
The visual devices changed.
Typography changed.
Navigation changed.
Section density changed.
The treatment of Stonehaven's $10M to $50M investment range changed.
There was real visual variety.
If we had stopped there, we might have concluded that the system was exploring a broad creative space.
But when we looked beneath the styling, the picture changed.
The amount of visible variation was greater than the amount of conceptual variation
We started grouping the sites based on the underlying communication and design idea rather than their individual visual treatments.
The 17 outputs appeared to collapse into roughly five creative families.
The categories are approximate. They are an editorial classification, not a statistical clustering model.
But they were useful.
1. Restrained editorial / founder manifesto
This was the largest group.
These sites tended to combine sophisticated typography, large amounts of whitespace, philosophical founder-facing language, investment criteria, operating capabilities, founder-alignment messaging, and a restrained call to action.
The individual executions could look quite different.
The underlying approach was often similar.
2. Layered / geological
A smaller set developed structural, geological, or layered visual ideas.
These connected naturally to the Stonehaven name and the firm's philosophy of building on a successful existing company rather than replacing what already works.
The stronger executions made that metaphor part of the communication rather than leaving it as decoration.
3. Growth and complexity progression
These sites focused on the founder's transition from a successful company into a more difficult stage of growth.
The central visual idea became increasing organizational complexity.
This produced a meaningfully different narrative because the website was organized around the problem the founder was experiencing.
4. Investment profile / diagnostic
These designs made investment criteria unusually prominent.
ARR, ownership, business profile, sector, and growth characteristics became major organizing devices.
Some versions felt almost diagnostic, allowing founders to quickly understand whether they resembled a Stonehaven investment.
5. Navigation-led editorial
The most structurally unusual designs used navigation or persistent page framing as part of the concept itself.
Instead of relying almost entirely on a conventional sequence of horizontally stacked homepage sections, the interface helped organize the story.
These were among the clearest departures from the dominant pattern.
We tried to score the difference
We looked across the 17 sites and rated the amount of variation we saw in several dimensions.
- Strategic variation: 2/5
- Messaging variation: 3/5
- Structural variation: 2/5
- Compositional variation: 3/5
- Typographic variation: 1.5/5
- Cosmetic variation: 4/5
The pattern was fairly clear.
The work changed most readily at the surface.
Palettes changed.
Graphic devices changed.
Hero treatments changed.
Spacing changed.
Visual details changed.
The deeper solution moved less.
Strategy and structure were considerably more stable.
Typography was especially convergent.
That does not mean the system produced the same website repeatedly.
It means that the impression of creative range was larger than the actual range of underlying approaches.
Certain design defaults appeared repeatedly
Across the set, we kept seeing variations of the same visual vocabulary:
- warm cream, sand, stone, or off-white backgrounds
- dark green, charcoal, ink, brass, and muted blue
- serif display typography paired with neutral sans serif body text
- minimal photography
- thin horizontal rules
- large amounts of whitespace
- philosophical founder-facing hero statements
- prominent $10M to $50M criteria
- investment-profile lists
- contrasting philosophy sections
- operating-support sections
- founder-ownership reassurance
- understated navigation
- “Start a conversation” style calls to action
Again, several of these choices were very good.
Some of the highest-rated pages used them.
That matters.
The goal should not be to create a blacklist of AI design patterns.
A serif typeface does not make a website generic.
A card does not make a website generic.
Whitespace does not make a website generic.
A familiar pattern becomes limiting when the system keeps returning to it without seriously exploring other ways to solve the communication problem.
Category convention is powerful
AI did not invent design convergence.
Professional-services websites looked like one another before generative AI.
Investment firms shared conventions before generative AI.
Technology companies copied one another before generative AI.
Templates, frameworks, design systems, agencies, trends, and client expectations have always created shared visual languages.
That is often useful.
A website needs some familiar behavior.
Navigation should be understandable.
Buttons should look interactive.
Investment criteria should probably be easy to find.
Creative work does not improve simply because every decision is unprecedented.
But generative systems make repetition much easier to produce at enormous scale.
A model has learned from a vast amount of existing work.
When a request resembles a known category, familiar patterns provide plausible solutions.
That gives generation an enormous head start.
It can also pull many solutions toward similar territory.
Our Stonehaven experiment demonstrates the convergence in the finished outputs.
It does not, by itself, prove exactly why that convergence occurred.
Model priors, category conventions, prompt language, learned design patterns, and the way the problem was framed are all reasonable areas to investigate.
They should remain explanations to test rather than conclusions that this experiment alone proves.
Strong company context did not eliminate the convergence
This was one of the more useful surprises.
A common explanation for generic AI design is weak context.
That explanation is partly right.
If you ask AI:
“Design a premium website for an investment firm”
you should not be surprised when it produces an average of familiar premium investment-firm patterns.
But Stonehaven was not an empty brief.
The system had a specific audience.
A specific investment range.
A specific company philosophy.
Specific sectors.
Specific operating capabilities.
Specific ownership preferences.
Specific brand characteristics.
And the context clearly mattered.
The outputs consistently preserved much of Stonehaven's strategic substance.
Founder-led businesses remained central.
The $10M to $50M range remained prominent.
The idea that successful companies do not need to be reinvented survived repeatedly.
The operating capabilities remained important.
The founder audience remained clear.
So context improved strategic consistency.
It did not automatically create a broad range of creative strategies.
That distinction matters.
Company understanding can tell a system what must remain true.
It does not necessarily force the system to explore every interesting way those truths might be expressed.
Prompt language can still pull toward familiar territory
Words such as:
modern
premium
sophisticated
clean
credible
understated
innovative
can be useful descriptions.
They are also broad.
They describe qualities shared by thousands of brands.
Stonehaven itself was supposed to feel sophisticated, understated, modern, credible, experienced, and calm.
Those characteristics were appropriate.
They were not enough to define a unique creative concept.
A useful creative direction needs to go further.
What is the central idea of this particular experience?
What should the founder understand differently after seeing it?
What tension are we organizing the story around?
What deserves to dominate the page?
What can disappear?
What visual or structural idea expresses something specific about this company?
Those questions create a different kind of constraint.
The strongest designs did not simply avoid conventions
This is another reason the solution is more complicated than banning familiar patterns.
Our two highest-rated sites both averaged 9.3.
They were also among the most distinctive.
One used a topographic visual language and a more deliberate editorial structure.
Another used a fixed navigation rail and expansive editorial canvas.
Its central line was:
“Your company already works. Growth is what gets complicated.”
Both designs found ways to make their structure part of the communication.
But one of the most launch-ready sites in the entire set was considerably more conventional.
It succeeded because it was extremely clear, credible, polished, factually disciplined, and appropriate.
Meanwhile, another site used an enormous $10M to $50M graphic treatment that was immediately distinctive.
That unusual device did not make the overall website stronger.
So distinctiveness itself was not the goal.
Quality and novelty were not interchangeable.
A more useful definition of creative variation
The experiment gave us language we now find more useful.
Surface variation changes how a recurring solution looks.
Both matter.
Typography can transform the character of a website.
Color changes perception.
Spacing changes authority and pace.
Visual execution is not superficial in the sense of being unimportant.
But if five concepts share essentially the same strategic story, information hierarchy, page architecture, and communication logic, changing their typography and visual treatment does not necessarily mean five creative directions were explored.
That distinction becomes important when generation is nearly free.
More outputs do not automatically create more ideas
This is where generative abundance can become misleading.
You can ask for five options.
Or ten.
Or fifty.
The interface may fill with possibilities.
But the number of outputs does not tell you how much of the creative solution space was actually explored.
Five outputs can represent five different ideas.
They can also represent five executions of essentially one idea.
As generation gets faster, this becomes a more important question.
The constraint moves away from producing another option.
The harder questions become:
Have we explored a meaningfully different approach?
What assumptions are shared by all of these directions?
What problem is each concept solving differently?
Which differences actually matter?
Which solution is strongest?
Deliberate divergence may need to become part of the process
A stronger creative process should probably make exploration explicit.
Before generating five finished websites, the system could first develop genuinely different strategic or creative directions.
One direction might organize Stonehaven around the founder's decision about taking investment.
Another might organize it around the complexity inflection point.
Another might behave like an investment diagnostic.
Another might make the six operating areas the central system.
Another might build everything around the idea of preserving what already works.
Those alternatives are meaningfully different before anyone chooses a palette.
Then the system can evaluate them.
Are these genuinely different?
Are they grounded in the company?
Does each reveal something useful?
Are two directions actually the same argument wearing different clothes?
Only then does it make sense to turn them into finished design.
This is why AI-generated websites can start to look the same
The issue is larger than a list of recognizable AI motifs.
It is possible to remove every glowing gradient, bento grid, pill, blob, and rounded card and still produce generic creative work.
The deeper convergence happens when many outputs keep using related ways of understanding the problem.
AI has access to an enormous visual vocabulary.
What matters increasingly is how broadly and intelligently it searches the solution space before settling on an answer.
Company context helps.
Creative constraints help.
References can help.
Evaluation helps.
And our experiment suggests another requirement:
The system needs to know whether the alternatives it is considering are actually different.
Because as generating another website becomes almost effortless, apparent variety becomes cheap too.
The more interesting measure is whether the next option contains another useful idea.