A Guide to the Web Design Process for People Whose Last Designer Was an AI Agent

The pitch for AI design tools is genuinely exciting. Type a prompt. Get a website. The whole thing in forty seconds, looking cleaner than anything you could have cobbled together yourself at 11pm on a Tuesday. Professional typography. Confident layout. A hero section that knows exactly where it should be.

What arrives is usually all of that. It also usually says nothing.

Last month I had a call with a SaaS founder who'd done exactly this. Framer AI, a free weekend, a site he'd built himself and was proud of. He shared his screen. It looked good – genuinely good, better than a lot of sites I audited in 2022 that someone charged actual money for. I told him that. He smiled.

Then I asked him who the site was for.

He described four different people. Three of them contradicted each other. The site, to its credit, had managed to speak to all four of them equally – which is to say, it had spoken to none of them at all. It was positioned for a demographic that doesn't exist: everyone with a vague interest in his general category.

He wanted "just a few tweaks." Specifically, he thought the CTA button needed to be more prominent.

The button was fine.

This guide is about what wasn't – and about the web design process that got skipped to get there.

AI design tools are genuinely impressive. I'm not going to pretend otherwise, and this isn't a piece about protecting the profession. v0, Framer AI, Webflow AI, whatever launched in the last six months that I haven't used yet – they can produce something that passes for a real website faster than any human can open a new Figma file.

That's not the problem.

The problem is what gets generated alongside the layout. AI tools are excellent at visual pattern matching – they've ingested enough well-designed websites to know how a homepage should be structured. Hero section, value proposition, social proof, CTA. Correct. What they can't do is figure out what your value proposition actually is, who it's for, or why anyone should believe it. So they generate something that fits the pattern. It sounds like a website. It has all the right components in all the right places.

It just says nothing. The headline is something like "Empowering teams to achieve more through intelligent solutions." Grammatically intact. Visually prominent. Meaning: zero. Every competitor in the category could use that line. Most of them already do.

And here's the part nobody warns you about: a site that says nothing in a very professional font is harder to fix than a site that looks bad. At least when it looks bad, you know something's wrong. When it looks polished and still doesn't work, you try other things first. You change the hero image. You run an A/B test on the headline – between two versions that are both wrong. You move the CTA above the fold. You pick a more confident font. You ask someone junior to "freshen up the copy." You consider whether the colour palette is the issue.

It is not the colour palette.

At some point, usually after six months and a Hotjar subscription, you book a call with someone like me. You share your screen. I tell you the site looks good. And then I ask you who it's for.

Before getting into steps, one reframe worth making: the web design process is not primarily visual. I know that sounds backwards coming from a designer, but the visual part is the last thing that happens, and it's also the fastest. If you want the broader context on how the discipline actually works versus how it gets taught, UX design 101: the official curriculum and the 2026 translation covers that in more depth.

The process exists because a website is an argument. It's making a case – for a product, a service, a person, a company – to someone who didn't ask to hear it and will leave in seconds if it doesn't land. Every structural decision on the page either supports that argument or undermines it. The process is how you figure out what the argument is before the first pixel moves.

Skip the web design process, generate the layout, and you get something that looks like an argument but isn't one. Which is where most AI-built sites end up.

The steps below are the process. Not every project needs all of them in equal depth. But every project needs all of them in some form, even if "in some form" means twenty minutes of honest thinking before opening any tool.

Discovery is the part where you establish what the site needs to do, who it's for, what makes the thing being offered different, and what success actually means. Sounds obvious. Takes longer than anyone budgets for. Almost always the first thing cut when the timeline gets tight.

What it involves, in practice, is asking uncomfortable questions until the answers stop being PR and start being honest. The questions aren't complicated – they're just hard to answer well:

• What do you do that a direct competitor genuinely can't claim?
• Who specifically is this for – and who is it not for?
• What does someone need to believe to hire you?
• What does a good outcome from this site mean, in a number?

"More clients" is not a number. Neither is "better visibility." Both get typed into briefs every day.

The reason these take time isn't that the answers are complicated. It's that the first answer is almost always the polished one – the version refined through pitch decks and investor calls. Getting past that to something true usually takes thirty minutes and at least one awkward silence. The silence is the useful bit. It means someone's actually thinking rather than reciting.

What good discovery produces is a short document – one page is fine, two is plenty – that answers those four questions specifically. Not a mood board. Not a competitor analysis. Not a Notion doc with forty-seven bullet points that somehow says nothing. One page. Four answers. Precise enough that two different designers would build the same site from it.

In terms of time: two to four hours of conversation for a straightforward project. Up to two days for anything complex. If someone tells you they did it in twenty minutes, they did a different thing and called it discovery.

Discovery in 2026: Most founders now arrive with what feels like a completed discovery document. It's usually a ChatGPT conversation that produced a positioning statement in four seconds, a Notion doc assembled from competitor websites, and a brief written for the AI tool rather than for a human being. None of these are wrong to have. None of them are discovery. A four-second answer to a thirty-minute question is a guess with good grammar.

AI treats the prompt as the brief. Whatever the founder typed – "I run a B2B SaaS for logistics companies, make it look professional" – becomes the entirety of the discovery process. The site gets built around the pitch. The pitch is usually the most optimistic version of the truth. Optimistic positioning and accurate positioning are not the same document.

This is the one step a founder can lead themselves, provided they're willing to be honest rather than aspirational. A designer's job here is to ask the questions and push back on the answers – not to accept the first version of the story. If you're working alone, find someone who'll ask the uncomfortable questions rather than nodding along. A co-founder works. A trusted client works better.

What skipping it costs: everything downstream. Copy that doesn't ring true. Positioning that appeals to everyone and convinces nobody. The wrong clients arriving consistently, close enough to take the work, wrong enough that every project is harder than it should be.

Discovery done vs discovery skipped – what you actually know:

The right column is not a joke. It's a composite of actual briefs I've received. Several of them were accompanied by a Loom video explaining how much thought had gone into them.

Brief quality audit – what a usable discovery document contains:

If the brief has all six, the rest of the process has a fighting chance. If it's missing the first three, the project will revisit them – either intentionally in a structured conversation, or accidentally in the fourth round of homepage revisions.

Positioning is the single most important step and the one most likely to be replaced with a tagline generator. It's the decision about what the site leads with, who it's speaking to, and – crucially – what it's not saying. That last bit is where most people lose their nerve.

The work is writing and rewriting the core message until it's specific enough to exclude people. That sounds counterintuitive. It isn't. Specific positioning feels like leaving money on the table. It's actually how you attract the right clients and stop wasting everyone's time, including your own. A homepage that tries to speak to everyone ends up speaking to the algorithm and nobody else.

Here's a rough diagnostic I use before starting any site project:

The difference isn't tone. It's specificity. Specific positioning is the thing AI cannot generate from a prompt because the founder hasn't told it – and often hasn't told themselves.

What good positioning produces is a single statement – one or two sentences maximum – that passes three tests. First: a direct competitor couldn't use it without lying. Second: the person it's written for reads it and thinks "that's me." Third: it makes someone outside the target audience feel correctly excluded rather than vaguely unwelcome. All three are harder than they sound. The second one usually takes the longest.

Time-wise: half a day to a full day, assuming discovery went well. If discovery produced a ChatGPT brief, add another half day while you work out what the business actually does. Positioning done in under an hour is either genius or avoidance. It's almost never genius.

Positioning in 2026: AI tools have made everyone's positioning sound the same. Not because the tools are bad – they're generating exactly what they've been trained on, which is every SaaS homepage written between 2018 and 2024. The output is fluent, professional, and completely interchangeable. "We help teams move faster." "The smarter way to manage X." "Built for the way modern teams work." All of these are on live sites right now. None of them are wrong. None of them mean anything concrete enough to be useful.

AI generates something that sounds like positioning. It hits all the structural notes – benefit-led, outcome-focused, audience-aware. It doesn't mean anything because meaning requires specificity, specificity requires a choice, and AI doesn't make choices on your behalf. It reflects them back.

This is the step a designer and founder have to do together. A designer alone doesn't know enough about the business. A founder alone is too close to it. The conversation between the two – specifically the part where the designer says "I don't think that's what you actually do" and the founder pauses – is where the positioning lives.

A few things that reliably kill it: positioning by committee, where every stakeholder softens it slightly until nothing remains; positioning for the business you want to be in two years rather than the one you are now; confusing tone with substance ("bold and human" is a tone, not a position); and finalising it before testing it on three real potential clients, which sounds obvious and almost never happens.

What skipping it costs: discovery calls with the wrong people. Conversion rates that plateau regardless of how many times the button gets A/B tested. A vague persistent sense that something's off that never quite locates itself, because the site looks fine. It does look fine. That's the problem.

IA is the structure of the site before any visual decisions are made. Which pages exist, what lives on each one, what order information appears in, what someone is supposed to understand by the end of each page. Not glamorous. Completely load-bearing. The part clients never ask about and always feel the absence of.

The work is mapping every page as a user journey rather than a content inventory. The question isn't "what do we need to include?" – it's "what does someone need to know, in what order, to go from stranger to convinced?" Different questions. Different structures. Most AI-built sites answer the first one. Most failing sites answer only the first one.

A simple audit I run on almost every project that comes to me after an AI build:

None of these are visual problems. All of them are structural ones.

What good IA produces: A sitemap and a page-by-page brief – one paragraph per page describing what the visitor knows when they arrive, what they need to know by the time they leave, and what action the page is trying to produce. Not a wireframe. Not a Figma file. A document. The wireframe comes later and takes a fraction of the time when this exists.

Half a day for a simple site. A full day for anything with more than six pages or a complex service offering. The temptation is to skip straight to wireframes because wireframes feel like progress. They are progress. They're just faster when the thinking happened first.

Information architecture in 2026: Navigation structures are where AI tools consistently underperform. They default to the five-item nav that's been standard since 2015 – Home, About, Services, Work, Contact – without asking whether that structure actually serves this site's argument. Sometimes it does. Often the journey a potential client takes is completely different, and the nav is working against it the entire time.

I recently audited a site where the founder's single biggest conversion moment – a 10-minute recorded demo – was buried three clicks deep because the AI had generated a standard portfolio structure. The demo had a 73% watch-through rate. Three people had found it in six months. The IA was the problem, not the demo.

IA is designer territory, but it requires the founder to sign off on the logic before anything visual happens. The most expensive version of this going wrong is approving wireframes built on a broken structure – at that point, every revision cascades.

Common structural mistakes: building the nav around the company's internal departments rather than the visitor's decision journey; creating pages because content exists rather than because the visitor needs them; treating the About page as a company history lesson; adding a Blog because every site has a Blog and then not updating it for eight months; burying the most convincing proof – demos, case studies, specific results – below the fold or behind an extra click.

What skipping it costs: bounce rates nobody can explain, visitors who reach the services page and leave without contacting, and the specific frustration of having genuinely good work that nobody sees because the structure never gave it a reason to exist.

The words. Not content – copy. There's a difference worth being clear about. Content fills space. Copy does a job.

Every headline is making an argument. Every subheading is either advancing it or wasting the reader's attention. Every CTA is either a natural next step or an interruption. Getting this right takes the most time of any step and is the least visible when it works. Nobody finishes reading a homepage and thinks "the copy was excellent." They just book the call.

In practice: writing, testing, cutting, rewriting. Specifically – writing the homepage headline until it's precise enough to make the right person feel seen and the wrong person feel correctly excluded. On a recent project this took eleven drafts. The final version was eight words. The first version was twenty-three words and said less.

The rest of the page follows the same logic. Each section has one job. The hero establishes who this is for. The next section proves it's real. The section after removes the obvious objection. The CTA asks for one specific thing. Any section that doesn't do a specific job is a section that talks a potential client out of staying.

What good copywriting produces is a complete copy document before anything goes into a design tool. Headlines, subheadings, body copy, CTAs, microcopy – all written, reviewed, and agreed before a single pixel moves. This sounds obvious. It is almost never done. The standard alternative is designing with placeholder text, then writing copy to fit the layout – which produces copy shaped by boxes rather than by argument.

In terms of time: three to five days for a five-page site, assuming positioning is solid. If positioning is still being worked out, add the time it takes to finish that first. Copy written on top of vague positioning is copy that will need to be rewritten. It will be rewritten at the worst possible moment, which is after the design is approved.

Copywriting in 2026: AI copywriting tools have raised the floor and lowered the ceiling simultaneously. The floor: nobody has an excuse for genuinely bad copy anymore. The ceiling: everybody's good-enough copy sounds identical, generated from the same training data, optimised for the same conversion patterns, filtered through the same "make it sound professional" instruction. The result is a web full of sites that are competent and forgettable in equal measure.

The tell is the CTA. AI-generated CTAs are almost always one of four things: "Get Started," "Learn More," "Book a Demo," or "Contact Us." All four are correct. None of them are precise enough to mean anything. "See how we fixed Deutsche Telekom's onboarding in six weeks" is a CTA. "Get Started" is a button.

One test I apply to every headline before signing off on it:

Could a direct competitor use this exact line without changing a word?

If yes, it's not copy. It's furniture. Start again.

Copy is a collaboration whether anyone plans for it or not. A designer writes a first draft. The founder corrects it – "that's not quite how we'd say it" – and in doing so reveals what the positioning actually is. This is valuable. It's also the stage most likely to turn into twelve rounds of line edits about comma placement. Agree upfront on who has final call and what "approved" actually means.

A few reliable mistakes: writing copy to fit the design rather than designing to express the copy; using the founder's internal language instead of the client's language, which are almost never the same; CTAs that describe an action rather than a benefit ("Submit" vs "Get the audit"); body copy explaining the process in detail nobody asked for on a page trying to convert a first-time visitor; and letting the about page become a timeline. Nobody reads the timeline.

What skipping it costs: trust, built or lost in four seconds. Generic copy signals a generic offer regardless of how specific the service actually is. A site with sharp positioning and weak copy is a gift with bad wrapping. A site with sharp copy and weak positioning doesn't exist – good copy forces the positioning question whether you're ready for it or not.

Colour, typography, layout, imagery, spacing – the UX and UI design decisions that communicate tone, trust, and category before anyone reads a word. It's also the step most people think of when they think "web design," which is why the previous four steps keep getting skipped.

Here's the part that tends to annoy people when I say it out loud: this step causes the least damage when it goes wrong. A site with impeccable visual design and broken positioning will underperform a site with average visual design and sharp positioning. Every time. I've seen it enough times to have stopped being surprised by it. The corollary – good visual design on top of solid positioning – is extremely effective. It's just not where the work happens, which is why it always gets done first.

The visual decisions that actually matter are the ones that support the argument rather than decorate it. Typography that matches the tone – a fintech company and a creative studio should not have the same typographic hierarchy, even if both brief says "trustworthy and modern." Colour that communicates something rather than just avoiding anything wrong. Imagery that's either genuinely relevant or deliberately absent, because stock photography of people in glass-walled offices has been doing active damage to trust for about four years now and nobody's stopped using it.

What this step produces, when done properly: a design system light enough to be maintained by one person. Colour palette, type scale, spacing rules, component library. For a small site this doesn't need to be a 200-page Figma document – it needs to be tight enough that nobody's making decisions by eye and getting something different every time. The 200-page version happens when nobody agreed on scope in discovery. Which brings us back to discovery.

How long it takes: One to two weeks for a five-page site, assuming positioning is done and copy is approved. That includes exploration, feedback, refinement. It does not include the round where someone wants to "see it in blue." That adds a day. It also doesn't include the conversation about whether the logo needs updating, which should have happened in Step 1 and almost never did.

Visual design in 2026: Neutral sans-serif, lots of white space, subtle gradients, one accent colour, sections alternating between white and very light grey. This is on roughly sixty percent of SaaS sites launched in the last three years. It works. It also communicates "professional" without communicating anything else, which is the aesthetic equivalent of describing yourself as "a person who does things."

The question isn't "does this look good?" It almost certainly does. The question is whether it looks like this business specifically, or a well-designed version of the category. Different question. Requires different decisions. Most AI tools answer the second one.

AI is genuinely useful here – more so than any other step. Layout iteration, component generation, first-pass visual direction. For a founder with clear positioning and approved copy, an AI tool gets you eighty percent of the way there faster than any human can. The remaining twenty percent is where the specificity lives. Where the difference between "this could be any company" and "this is unmistakably us" gets decided. That part still requires a human with an opinion and the positioning document open.

A few things that reliably go wrong at this step:

Starting before copy is finalised. Design built around placeholder text gets redesigned when real copy arrives – and real copy always arrives longer than expected, in a font size nobody accounted for, on a mobile screen that's smaller than the Figma frame.

Typography chosen by aesthetic preference rather than legibility at the sizes that actually matter. Body copy on mobile. Not headlines on a 27-inch monitor. That's where most people are reading.

Everything animating because the tool made it easy to turn on. Motion should direct attention. When everything moves, nothing is emphasised and the whole site feels like it's trying to distract you from something.

And finally: approving the desktop design and assuming mobile will sort itself out. It won't. It never does.

What it is: The part where the designed thing becomes the real thing, and approximately forty percent of decisions get revisited because they worked in Figma and don't work on a phone.

Build and QA – including how AI is changing what's possible here, what breaks in AI-built sites, and how to actually test before shipping – is getting its own guide. It deserves more than a section. That one's coming soon.

The short version: test on a real phone before going live. Not a browser simulator. An actual phone, held in an actual hand, tapped with an actual thumb. The number of sites that skip this step and then wonder why mobile conversion is half the desktop rate is, conservatively, most of them.

I use AI tools daily. Cursor for code, Claude for thinking out loud, Midjourney for visual direction, v0 for component exploration. None of them have replaced a single discovery conversation. All of them have made the work faster once the thinking was already done. I've also spent enough time thinking about what AI does to your reasoning when you let it do too much of it that I wrote a book about it.

The pattern I keep seeing – in my own projects and in the sites people bring me to fix – is AI used at the wrong stage. Not because the tools are bad, but because they're genuinely impressive at the beginning, which makes it easy to confuse "this looks like progress" with "this is progress." Generating a homepage in forty seconds feels like skipping ahead. It is skipping ahead. The question is what you've skipped.

Here's where AI actually earns its place in the web design process – and where it consistently doesn't:

The build and code stage is a separate conversation – and a genuinely interesting one. I'm writing a guide specifically on converting human-produced design into production code using AI, because that's where the tools are transforming the process in ways worth understanding properly. That one's coming.

For now: AI is most useful after the thinking is done. Use it to go faster once you know where you're going.

The site looks fine. That's not the problem. Here are the actual signals:

You're explaining the site on every call. The first five minutes of every discovery call are a correction. What the site implies, what the business actually does – there's a gap, and you're filling it manually, on every call, indefinitely.

The wrong people are reaching out. Not bad people. Just wrong ones. Close enough to take the work, wrong enough that every project is harder than it should be. The positioning is too broad, too vague, or accidentally aimed at someone other than your actual client. All fixable. None of it is a design problem.

Traffic exists, conversions don't. People arrive. They look around. They leave. Something is failing in the argument the site is making, and that failure is in the copy and structure – not the font choices, not the hero image, not the CTA button everyone keeps suggesting you make bigger.

You've changed it six times and nothing moved. New hero image. Different headline. Repositioned testimonials. A/B test between two versions of the same vague statement. None of it helped because none of it addressed the actual problem, which is three steps upstream of everything you changed.

You're embarrassed to send people to it. This one doesn't get said out loud often. You mention the site exists but you don't link to it. You describe what you do on calls rather than sending people there first. That instinct is correct. The site is telling a story you've already moved past.

You're about to raise a round, hire senior people, or go upmarket. These are the moments when the site stops being background noise and becomes the first impression for people who will absolutely notice if it's average. This is also, historically, when startup web design stops being a weekend project.

In none of these cases is the answer a new AI-generated layout. The answer is the web design process – starting at the beginning, before anyone opens a tool.

A site built on a skipped process is a problem that compounds. The longer it's live, the more it shapes what clients arrive, how they describe what you do, what rates feel reasonable to ask for. A founder I worked with last year had used an AI tool to build their site two years prior. It looked fine. The positioning was fuzzy. For two years, every client who came in was slightly wrong – close enough to take the work, wrong enough that every project was harder than it should have been.

The site took six weeks of product design work to redo. Not because it looked bad. Because it was describing the wrong business and attracting the wrong clients, consistently, for two years.

The AI didn't do that maliciously. It did it because nobody gave it a clear brief. Nobody gave it a clear brief because the brief requires the process, and the process requires the uncomfortable thinking, and the uncomfortable thinking is the one thing that's genuinely hard to outsource – because it requires knowing what's actually true about your business rather than what sounds best about it.

That's the whole guide in one sentence, if you want it.

The brief – what it should contain, how to write one that actually works, and why most of them don't – is getting its own guide. Because "nobody gave the AI a clear brief" is only half the story. The other half is that most people don't know what a clear brief actually contains, which is not a criticism, it's just a gap nobody teaches. That one's coming soon.

If any of this sounds familiar and you'd like to talk about it, the contact page is right there. I don't bite. I do ask uncomfortable questions, but that's the job.