A founder asks an AI assistant to build a company website in minutes. The result goes online the same day, with a hero image, three service pages, and a contact form. Six months later, the site has not generated a single qualified lead. The problem was never the speed. It was that speed was treated as the goal, when the goal should have been a site aligned with what visitors need and what search engines can understand. This distinction separates a disposable AI page from a website that works, and it is the organising principle of any disciplined approach to building websites with AI.

Faster is not the objective

The temptation with any new tool is to ask how it can reduce cost or effort. That question leads into a well documented trap. When every competitor has access to the same tool, the firms that survive are those who use it to achieve significantly better results with the same effort. The firms that use it only to save a little money, or to finish a little earlier, are gradually squeezed out. Applied to web development, this means AI should raise the quality bar, not merely shorten the timeline. A site produced in an afternoon is only valuable if it also converts, ranks, and serves its audience.

Just do not make the mistake of asking how AI can help us reduce costs or effort. Jacobsen, 2027

A second, related error is vendor lock-in. A model-neutral approach treats ChatGPT, Gemini, Claude, and any later model as interchangeable assistants. Each has different strengths and blind spots, and consulting more than one on the same question exposes where the answers are confident and where they are invented. The process should not depend on a single interface, because the requirements of the website do not change when the model does.

Plan before you prompt

The first productive step is not a prompt. It is a plan. Human-centred development begins with two questions: what is the website for, and who is it for. The goals determine which pages and functions are necessary. The target audience determines the language, the structure, and the tone. Only after these are defined does the choice of platform matter. Wix, Squarespace, WordPress, and Webflow each suit different needs, and AI can advise on the trade-offs, but the decision belongs to the project plan, not to the tool that suggests it first.

AI is at its most useful here as a structured thinking partner. A well formed prompt assigns a role, states the constraints, and demands a precise output format. Asked to act as an experienced project manager, a model can produce a realistic estimate of pages, images, and effort in a table. Such estimates carry enormous variation, however, and should be treated as a reference point, never as a fact. Cross checking them against another model, or against real project data, is what turns a plausible number into a usable one.

The seven-step process

A reliable workflow proceeds in seven stages, each one verified by a human before the next begins.

  • Understand what AI can do and where it fails, so expectations match the tool.
  • Define the goals and target groups before any design or platform decision.
  • Select the platform and the models that fit the plan, remaining model-neutral.
  • Determine the site structure, navigation, and page hierarchy.
  • Design and build the interface, using AI for drafts and humans for judgement.
  • Create content efficiently, then refine it through both AI and human feedback.
  • Test, optimise, promote, and measure the site after launch.

Skipping a stage does not save time. It moves the cost later, where a structural flaw discovered after launch costs more to fix than a planning session would have cost to prevent.

SEO is built in, not bolted on

Search visibility is a property of how the site is built, not an afterthought added by a plugin. Performance is a ranking factor, and the relationship is measurable. An improvement of one hundred milliseconds in loading speed has been associated with a one percent increase in sales, and slow sites are penalised in search results. Images must be compressed without visible loss, code must be lean, and the site must be usable on every screen size. An open source audit tool such as Lighthouse checks performance, accessibility, and search readiness together, and its report should be run before launch and after every major change.

The traffic landscape now includes two kinds of engines. Traditional search engines reward structured content, fast pages, and clear authority. AI platforms reward content that answers questions directly and can be cited with confidence. A website optimised for both is built around clear questions and precise answers, with pages that a machine can parse as easily as a person can read. Tracking where visitors come from, and whether they take the desired action, is what separates a visibility strategy from a vanity metric.

Verify everything the model produces

The final discipline is verification. AI drafts copy, suggests structure, estimates effort, and generates images. None of these outputs is trustworthy on its own. Facts are checked against primary sources. Estimates are cross checked against another model or real data. Design decisions are tested with real users, not judged by the team that made them. The model handles volume. The human team handles judgement, and that division of responsibility is what makes an AI built site indistinguishable from the best conventionally built sites, and often better.

Further notes on applied AI engineering are published on the blog, and project enquiries can be sent through the contact page. The full model-neutral process is treated in Creating Websites with AI: A Model Neutral Beginner's Guide by Jens Jacobsen (CRC Press, 2027).