If you ask the same AI to build the same website 20 times, what should change?
We originally thought this experiment would mostly be about consistency.
Would the same facts survive every generation? Would the same positioning keep showing up? Would the visual identity wander? Would some versions simply be much better than others?
Those questions mattered.
But once we put the outputs next to each other, another question became more interesting:
How many different ideas were actually there?
Because companies do not want the exact same website twenty times.
They also do not want twenty different interpretations of who they are.
The useful space sits somewhere between those extremes. The company should remain recognizable. The strategy should remain grounded. The facts should remain true. But the system should still be capable of finding genuinely different ways to solve the creative problem.
That is what we wanted to examine.
## The experiment
We used a fictional company called Stonehaven Growth Partners.
Stonehaven is a growth equity firm investing in founder-led vertical software and technology-enabled services companies in North America.
The brief was intentionally specific.
Stonehaven typically invests in established, profitable companies with $10M to $50M in annual recurring revenue. It looks for strong customer retention and companies entering a more complex stage of growth.
The firm works with founders and management teams on go-to-market strategy, leadership development, pricing, operational infrastructure, strategic acquisitions, and market expansion.
Its central philosophy is relatively simple:
Successful founder-led companies usually do not need to be reinvented. They need experienced partners who can help management make better decisions as the company grows.
Stonehaven generally makes control or significant minority investments and is comfortable with founders retaining meaningful ownership.
The primary audience for the website is founders and CEOs considering a growth investment.
The desired character was sophisticated, intelligent, understated, confident, modern, credible, experienced, and calm.
The experiment was designed around 20 independent generations.
At the time of this analysis, 17 completed homepages were available for evaluation, so the findings below cover those 17.
For the evaluation itself, we deliberately separated the finished work from the process that produced it.
The evaluator received the same source-of-truth company brief and the 17 finished homepage artifacts.
The evaluator was instructed not to infer which site came first, what process created it, or how any individual design had been generated.
Only the finished work was evaluated.
## What we evaluated
We looked beyond whether each site was attractive.
Each homepage was evaluated across:
- positioning and differentiation
- messaging specificity
- narrative coherence
- information architecture
- message hierarchy
- audience understanding
- credibility
- factual integrity
- visual distinctiveness
- composition and layout
- typography
- design-system coherence
- avoidance of generic investment-firm patterns
- appropriateness
- overall creative quality
- launch readiness
We also looked at variation itself.
- Were the outputs strategically different?
- Were they structurally different?
- Were they compositionally different?
- Or were many of the differences primarily color, typography, styling, and execution?
That distinction became important.
## What stayed surprisingly consistent
The core strategic idea survived remarkably well.
Across the designs, Stonehaven was repeatedly positioned around some version of the same argument:
You have already built a successful company.
Growth creates a new kind of complexity.
Stonehaven can help you navigate that complexity without unnecessarily reinventing what already works.
The $10M to $50M ARR range appeared prominently.
Founder-led businesses remained central.
Vertical software and technology-enabled services remained central.
Founder ownership and alignment remained important.
The six operating areas repeatedly appeared as evidence of how Stonehaven helps companies scale.
That consistency is meaningful.
The source material exerted real control over the strategic content of the outputs.
But another kind of consistency appeared too.
The pages repeatedly returned to similar ways of expressing those ideas.
## The websites looked very different
At first glance, the set felt highly varied.
- Some sites were light and restrained.
- Some were dark and institutional.
- Some used large typography.
- Some used diagrams.
- Some emphasized investment criteria.
- Some leaned heavily into editorial whitespace.
- Some introduced topographic or geological imagery.
- Some used persistent navigation devices.
- Some felt more like a traditional investment firm.
- Others felt closer to an editorial publication or modern technology company.
- Hero compositions changed.
- Color systems changed.
- Typography changed.
- Navigation changed.
- Section density changed.
- Visual metaphors changed.
- The treatment of Stonehaven’s investment criteria changed.
There was clearly variation.
But once we stopped looking primarily at the surface, the set started to feel much narrower.
## 17 different-looking homepages did not necessarily represent 17 different creative ideas
We grouped the sites according to the underlying communication and design logic rather than their color palettes or styling.
The 17 outputs appeared to collapse into roughly five broader creative families.
### 1. Restrained editorial / founder manifesto
This was the largest family.
These sites generally used generous whitespace, sophisticated typography, philosophical founder-facing messaging, understated navigation, investment criteria, operating capabilities, and a clear conversation CTA.
The executions varied, sometimes substantially.
But the underlying idea was often similar.
### 2. Layered / geological
A smaller group used structural, geological, or layered visual ideas.
These connected naturally to the Stonehaven name and to the broader positioning around building on an existing foundation rather than starting over.
The strongest example made the metaphor part of the communication rather than leaving it as decoration.
### 3. Growth and complexity progression
These sites visualized the transition from a successful but relatively simple organization into a more complex stage of growth.
The creative idea was less about Stonehaven as an object and more about the moment the founder was experiencing.
That created a different way into the same company story.
### 4. Investment profile / diagnostic
These designs made Stonehaven’s investment criteria unusually prominent.
ARR, company profile, ownership, sector, and growth characteristics became organizing devices rather than supporting information.
Some versions approached the homepage almost like a diagnostic or investment dashboard.
### 5. Navigation-led editorial
The most structurally distinctive sites treated navigation and page framing as part of the creative concept.
Instead of relying almost entirely on a conventional sequence of full-width sections, the interface itself helped organize the narrative.
These represented some of the clearest departures from the dominant homepage pattern.
The boundaries between these families are not scientific.
A site can borrow characteristics from more than one.
But the exercise revealed something useful:
The amount of visible variation was much greater than the amount of conceptual variation.
## We scored the variation itself
When we looked at the set as a whole, the differences were uneven.
- 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
That distribution matters.
The AI was very willing to change how the work looked.
It was considerably less willing to change the deeper logic of the solution.
Different palettes, type choices, graphic treatments, spacing, and hero compositions created a strong impression of exploration.
But underneath those choices, many generations were still solving the problem in similar ways.
## Certain defaults appeared again and again
Across the set, we repeatedly saw:
- warm cream, sand, stone, or off-white backgrounds
- dark green, charcoal, ink, brass, or muted blue accents
- serif display typography paired with a neutral sans serif
- minimal photography
- thin horizontal rules
- large amounts of whitespace
- philosophical founder-facing hero copy
- prominent $10M to $50M investment criteria
- criteria ledgers or specification lists
- contrasting philosophy sections
- sections covering Stonehaven’s operating capabilities
- founder-ownership reassurance
- understated navigation
- “Start a conversation” style calls to action
None of those decisions are inherently bad.
Several of the best sites used them.
That is important because the lesson from this experiment is not that recurring patterns should simply be avoided.
A familiar solution can still be the right solution.
The interesting question is how often the system considered something meaningfully different before arriving there.
## More variation did not automatically produce better design
The highest-rated sites make this especially clear.
Site #9 and Site #15 both received an average score of 9.3.
They were also among the most meaningfully distinctive designs in the set.
Site #9 used a topographic visual language and a more deliberate editorial structure to support Stonehaven’s positioning. Its visual concept, narrative organization, and communication strategy reinforced one another.
Site #15 used a fixed navigation rail and expansive editorial canvas. It also produced one of the strongest lines in the entire experiment:
“Your company already works. Growth is what gets complicated.”
The composition was unusual, but the difference had a purpose.
Site #2, which scored 8.9, developed one of the strongest visual concepts by using a layered structural language that connected naturally to the Stonehaven name and philosophy.
Site #3 used the transition into organizational complexity as the core visual idea and was especially strong from an audience-strategy perspective.
But then there was Site #11.
Site #11 was one of the most launch-ready designs in the entire group.
Its underlying approach was relatively conventional.
It succeeded because it was clear, polished, credible, factually disciplined, and highly appropriate for the audience.
That matters just as much as the more experimental examples.
At the other end, Site #5 used one of the most immediately distinctive devices in the group: an enormous $10M to $50M visual treatment.
The device was unusual.
The overall homepage was not stronger because of it.
Difference alone did not create quality.
## The distinction we found most useful
The experiment made us think about creative variation more precisely.
Meaningful variation changes the strategy, narrative, structure, hierarchy, composition, or way the audience understands the company.
Surface variation changes how a recurring solution looks.
That does not make surface variation useless.
Typography matters.
Color matters.
Spacing matters.
Visual character matters.
But changing those things should not automatically count as exploring a new creative direction.
A system can produce ten attractive variations and still explore essentially one idea.
## This changes how we think about consistency
We started this experiment expecting to study consistency.
We still think consistency matters.
But “make it consistent” is too crude a goal for creative AI.
Some things should be highly stable:
- company facts
- positioning
- audience
- proof
- strategic priorities
- brand character
Other things should be allowed to move:
- narrative strategy
- message hierarchy
- composition
- structural concept
- visual metaphor
- section architecture
- editorial treatment
The system needs to know the difference.
And now we would add another requirement.
It needs to know whether the alternatives it is generating are actually different.
## More generations do not automatically mean more ideas
This may be the most important finding from the experiment.
Generating another output is easy.
Generating another genuinely useful point of view on the same problem is harder.
The 17 sites we reviewed contained a lot of good work.
Several were strong enough that we could imagine launching them with relatively modest refinement.
But the set also showed how quickly apparent creative abundance can become misleading.
If five outputs use essentially the same strategic narrative, information architecture, and compositional logic but change the palette, typography, hero treatment, and graphics, we have five executions.
We may not have five ideas.
That distinction becomes increasingly important as generation gets cheaper.
The bottleneck shifts.
The hard part becomes deciding what should remain true, where exploration is valuable, whether the alternatives are meaningfully different, and which solution is actually best.
## What a stronger creative system would need to do
We came away with three requirements.
First, preserve what should remain true.
Company facts, positioning, audience, proof, and core brand principles should not randomly drift between generations.
Second, deliberately explore the areas where multiple strong answers are possible.
That includes narrative, hierarchy, composition, structural concept, visual metaphor, and creative direction.
Third, evaluate both quality and difference.
A system should be able to recognize that two attractive designs may actually represent the same underlying idea.
It should also be able to recognize that a highly unusual solution is not automatically a better one.
That is a harder problem than generation.
It is also a much more interesting one.
## Controlled variability
We still like the idea of controlled variability, but after this experiment we would define it more precisely.
Controlled variability means maintaining strategic and factual consistency while deliberately exploring meaningfully different creative solutions.
The goal is not maximum consistency.
The goal is not maximum variation.
The goal is a system that understands what should remain stable, where divergence is valuable, and how to judge the results.
That is a very different kind of creative tool.
## A limitation worth being explicit about
This experiment tells us something about the finished artifacts.
It does not tell us why the outputs converged.
The evaluator intentionally did not know how each design had been produced and did not attempt to infer generation order, process, or workflow.
The five creative families are also an analytical grouping, not a statistically derived clustering model.
And at the time of this analysis, 17 of the planned 20 outputs had been evaluated.
So we should be careful about making causal claims.
What we can say is simpler:
The finished work contained substantially more cosmetic variation than strategic or structural variation.
Several of the best designs found meaningfully different ways to solve the problem.
Several other excellent designs succeeded without doing anything especially unconventional.
And many apparently different outputs shared much more underlying creative logic than their surfaces suggested.
That feels like an important distinction.
Because as AI makes generating another option almost free, the number of outputs matters less.
What matters is whether the next option contains another idea.