# Everyone is racing to generate a website in 30 seconds. The work is in the next three months. > Source: > Published: 2026-08-10 21:00:00+00:00 Watch any website builder demo from the last two years and it is the same thirty seconds. Someone types "yoga studio in Lisbon", a progress bar does something reassuring, and a site appears. Hero image, three services, testimonials, a contact form. It looks good. It genuinely is good. Then the demo ends, because the next part does not demo. I work on a [website builder](https://webkio.com), so this is a criticism of my own industry and partly of my own product. Generation was the hard problem for about eighteen months. It is not the hard problem now, and I think a lot of us are still competing on it because it is the part that fits in a GIF. Any team that wants prompt-to-website can have it. The models are good, the tooling is mature, the ideas are public. If your differentiator is that a site appears from a sentence, your differentiator has a shelf life measured in whatever your nearest competitor's next sprint is. That is not a complaint about AI. Blank-page paralysis is real, and watching something appear from nothing is genuinely the moment people decide to try a product. It removes a barrier that used to stop people cold. It just is not where the difficulty went. It is where the difficulty *left*. Take a business that hires things out. Vans, bikes, boats, equipment, holiday lets. The generated site is fine: photos, a fleet grid, a booking button. Now try to actually hire something out. None of that is content. None of it can be inferred from "van hire company in Porto". It is **policy** - a set of decisions only the owner can make, which encode how their business actually runs and where they lose money if it is wrong. And the site with none of it configured looks exactly like the site with all of it configured. That is the whole problem. This is the bit I would want a developer audience to take away, because it generalises well past websites. If the model writes a bad headline, you see it in a second. If it picks a hideous colour, you see it immediately. Generation output is judged by plausibility, and humans are extremely fast at judging plausibility. The feedback loop is instant and the error is cheap. Configuration output has no such signal. A missing turnaround buffer looks like nothing. A per-count availability check looks like nothing right up until the Saturday you hand someone the keys to a van that is already out, in front of them, having taken their money. The error surfaces weeks later, to a customer, as a broken promise rather than a bug report. LLMs are very good at producing artifacts judged by plausibility. They are much weaker at producing artifacts judged by correctness against rules that are not visible in the output. Pricing, scheduling, permissions, tax, availability - the whole category shares that shape, which is why "AI generates your site in 30 seconds" and "AI runs your business logic" are not the same claim wearing different clothes. **Time to site** is a sales metric. It measures how quickly someone can evaluate you. **Time to first transaction** is a product metric. It measures whether they got what they came for. Nobody signs up for a website builder because they want a website. They want a booking, an order, a customer, an enquiry that turns into work. The gap between those two numbers is where the industry actually competes, and it is measured in weeks, not seconds. That is where the three months go: not designing, but working out what a deposit should be, which dates are peak, whether you can afford free cancellation, how long you need between hires, and then finding all of those settings in a product that hid them because they do not demo well. Not writing the hero copy. That is the easiest part of the page and the part the owner will rewrite anyway because they know their business and the model does not. The higher-value use is **generating the setup, not the site**. Arriving with a working configuration for that trade rather than an empty one: a fleet already priced, deposits set, seasonal rates in place, a turnaround gap that reflects how long a handover actually takes, extras attached to the booking flow. The owner then edits a working business instead of assembling one from parts. Deleting and adjusting is enormously easier than building, and far more people finish. It is also a much better use of a model that is good at plausible defaults for a domain and bad at knowing your specific rules. We got this wrong for a long time, for what it is worth. Our own templates looked right and arrived empty, and the feedback was consistent: nobody struggled to change the colours. They struggled with everything underneath the colours. If you are evaluating one, or building one, these are the questions that separate a demo from a product: The thirty seconds is genuinely impressive and it is genuinely solved. I would just rather we started competing on the three months, because that is the part the customer is actually living in.