The first wave of AI website tools asked a remarkable question:
What if you could describe a website and have it appear?
That was a real breakthrough.
For years, creating a website meant coordinating strategy, copy, design, development, CMS setup, implementation, and revision.
Generative AI compressed huge parts of that process into minutes.
It can now write copy, propose layouts, generate code, create imagery, build interactions, adapt responsive behavior, and revise the work through natural language.
That changes the economics of production.
It also makes the remaining problems much easier to see.
A website can appear almost instantly.
The harder question is whether it is the right website.
Generation solved an important problem
We should not minimize what has happened.
The ability to turn an idea into a polished creative artifact quickly is enormously useful.
Execution used to be one of the biggest constraints on creative work. You could know what you wanted and still need a team, budget, software, and significant time to produce it.
AI is removing more of that constraint every month.
But once generation becomes abundant, producing another option becomes less valuable by itself.
The quality of the decisions around generation starts to matter more.
What should the website say?
What should it emphasize?
What should it leave out?
What should it feel like?
Which creative direction is strongest?
Are the alternatives actually different?
Does the finished result work as one coherent object?
Is it factually grounded?
What should the system learn from the outcome?
Those are different problems from generating the page.
The system has to understand the company first
Before creating anything, a useful creative system needs a working model of the company.
Who is the audience?
What does the company actually do?
What does it believe?
What differentiates it?
Which claims are approved facts and which are assumptions?
What evidence matters?
What should the audience understand or do?
What has already been decided about the brand?
Without that context, generation is forced to fill in too much of the problem from broad patterns and category conventions.
Company understanding creates constraints.
It tells the system what should remain true.
But our recent testing suggests that context alone does not solve the entire creative problem.
Understanding the company does not guarantee meaningful exploration
The company brief was unusually specific. It included the audience, positioning, investment criteria, philosophy, operating capabilities, ownership approach, and intended brand character.
The context clearly affected the work.
Across the designs, important facts and strategic ideas remained surprisingly stable.
But the amount of creative variation depended heavily on what level we examined.
The sites looked substantially different in color, graphics, hero composition, section treatment, navigation, and other visible details.
When we grouped them by the underlying communication and design logic, the 17 outputs appeared to collapse into roughly five broader creative families.
Our qualitative variation scores showed the same pattern:
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 finished work contained much more surface variation than strategic or structural variation.
That does not tell us why the outputs converged. The evaluation intentionally looked only at the finished artifacts and did not infer the generation process.
But it does reveal another requirement for a mature creative system:
It needs to know whether the alternatives it is generating are actually different.
Exploration needs to become an explicit stage
A system can understand a company very well and still settle too quickly on one way of expressing it.
Consider Stonehaven.
The same company context could support several legitimate directions.
One homepage could be organized around the founder's decision about taking outside investment.
Another could revolve around the complexity that appears when a successful company reaches its next stage of growth.
Another could behave like an investment diagnostic, helping founders recognize whether they fit Stonehaven's profile.
Another could make the firm's operating capabilities the organizing framework.
Another could build the entire experience around the idea of preserving what already works.
Those are different ways of understanding and communicating the same company.
They are meaningfully different before anyone chooses a color palette or typeface.
A stronger creative system should deliberately develop alternatives at that level.
Then it should evaluate them before committing to one.
Are these actually different directions?
Does each follow from the company context?
What communication problem does each solve differently?
Which assumptions are shared across all of them?
Which direction gives the audience the clearest or most memorable understanding of the company?
That is a different job from asking for five visual variations.
Planning still matters, but the plan also needs judgment
In a separate experiment, we compared a homepage designed immediately with one that was planned before it was designed.
Planning improved several important things. It helped with strategic discipline, factual integrity, proof continuity, and narrative coherence.
It did not improve every creative decision.
Some individual decisions in the unplanned design were stronger.
The important lesson was that a plan generated by AI is still a set of decisions generated by AI.
Some parts of the plan are facts or constraints that should be protected.
Other parts are strategic judgments or creative hypotheses that should remain open to challenge.
A plan should therefore be evaluated before it becomes the blueprint for the work.
Combined with the Stonehaven variation experiment, the workflow becomes clearer.
The system should not rush from company understanding into one plan and one design.
It should explore, compare, choose, and then plan.
The workflow changes
The first generation of AI creative software largely optimized:
A stronger system needs something closer to:
Understand → Diagnose → Explore Directions → Evaluate Directions → Plan → Create → Evaluate → Refine → Remember
Each stage solves a different problem.
Understand
Build a working model of the company, audience, brand, goals, facts, proof, priorities, and context.
Diagnose
Identify what is missing, unclear, inconsistent, strategically unresolved, or likely to weaken the work.
Explore Directions
Develop genuinely different ways to solve the communication and creative problem. Vary the strategic framing, narrative, hierarchy, information architecture, structural concept, composition, and visual idea where multiple strong answers are possible.
Evaluate Directions
Determine whether the alternatives are meaningfully different, grounded in the company, appropriate for the audience, and worth developing. Reject directions that are simply restyled versions of the same idea.
Plan
Turn the selected direction into a coherent plan for the page or artifact. Define the narrative sequence, proof requirements, hierarchy, section roles, constraints, and creative hypotheses.
Create
Use generative models to execute at speed within the chosen direction and plan.
Evaluate
Review the finished work against the company context, strategy, factual requirements, audience needs, brand, composition, usability, and quality standards.
Refine
Improve weak decisions based on the evaluation rather than relying on endless unguided tweaking.
Remember
Preserve approved decisions, rejected ideas, successful patterns, changing company context, feedback, and useful precedents so future work does not start from zero.
Evaluation cannot be an afterthought
Generative systems can produce work that looks finished before it has been seriously judged.
That creates a dangerous shortcut.
A polished homepage can still have generic positioning.
A beautiful deck can bury its strongest proof.
A sophisticated one-pager can invent a fact.
A visually consistent site can repeat the same section pattern until the experience becomes monotonous.
A highly distinctive design can simply be strange.
The system needs standards beyond whether the output looks plausible.
Evaluation should ask whether the work is strategically correct, factually grounded, coherent, useful, appropriate, distinctive where it matters, and strong enough to ship.
It should also evaluate the alternatives against one another.
The best direction is not necessarily the most unusual one.
In the Stonehaven experiment, several of the most distinctive designs were among the strongest. One of the most launch-ready sites was comparatively conventional and succeeded through clarity, credibility, factual discipline, and execution.
That is why creative intelligence needs both exploration and judgment.
Variation without evaluation creates noise.
Evaluation without exploration can simply optimize the first plausible idea.
The whole site is a systems problem
A website is not one generation.
It is a collection of pages, sections, messages, proof points, components, images, interactions, and decisions that need to work together.
The system has to know what the homepage already said before writing the About page.
It needs to recognize when a layout pattern has been overused.
It needs to preserve the brand without making every page identical.
It needs to understand when a new page should follow precedent and when the content calls for a new solution.
That requires awareness beyond the current prompt.
It requires a model of the work itself.
Creative memory changes the value over time
Every project creates information.
A message was approved.
A visual direction was rejected.
A proof point turned out to be especially important.
A particular composition worked.
A pattern became repetitive.
The company's positioning changed.
An executive gave feedback that should affect future work.
If all of that disappears when the session ends, the next project starts unnecessarily close to zero.
A useful creative system should become better at working with a company as it is used.
That is creative memory.
Memory should not blindly preserve everything. It should distinguish approved precedent from experiments, stale decisions, rejected directions, and superseded information.
The objective is continuity without rigidity.
This becomes bigger than a website generator
Once a system understands the company, explores and evaluates creative directions, plans work, evaluates output, and remembers what it learns, the website becomes one expression of a larger intelligence.
The same foundation can support:
decks
one-pagers
campaigns
sales materials
social content
announcements
reports
new website pages
brand assets
other high-stakes communication
The unit of value changes.
The question becomes less about whether the system can generate a webpage.
The more interesting question is whether it can help the company make better creative decisions repeatedly over time.
Creative intelligence
We use the phrase creative intelligence to describe a system that combines persistent company understanding, creative knowledge, deliberate exploration, planning, judgment, generation, evaluation, refinement, and memory.
That definition has become more specific as we have tested the individual parts.
Company understanding helps preserve what should remain true.
Exploration helps prevent the first plausible solution from becoming the only solution.
Direction evaluation tests whether alternatives are actually different and useful.
Planning creates coherence.
Plan evaluation keeps creative hypotheses from becoming accidental facts.
Generation makes execution fast.
Output evaluation applies standards.
Refinement improves the result.
Memory lets the intelligence compound.
The models themselves will continue becoming more capable.
That makes the systems around them more interesting, not less.
A raw model can generate almost anything.
A creative intelligence system should know what is worth exploring, what should remain true, which direction is strongest, how to execute it coherently, whether the result is good, and what should be remembered afterward.
That is the transition we think is coming.
From AI website generation to creative intelligence.