# One Small Website, Six AI Builders, and Dependencies Nobody Ordered > Source: > Published: 2026-11-11 06:55:08+00:00 The errand was small. Pete Millspaugh, who works at Val Town, a company that sells one of these tools itself, needed a website for a conference talk called “Your corner of the internet”. He wanted one page, plus a comments section so the thing would do more than sit there. He typed the same request into six machines: Lovable, Vercel’s v0, Replit, ChatGPT Sites, Claude Artifacts and Val Town’s own Townie. On 27 October 2026 he published [the results on the company blog](https://blog.val.town/how-to-make-a-website-in-2026) under a Dickensian title: it was the best of times to make a website, it was the worst of times. Two weeks later the web developer Jim Nielsen, [writing on his blog](https://blog.jim-nielsen.com/2026/making-websites-in-2026/), picked through the account and asked what any subject eventually asks of a new order. Who decides? Who gets paid? What exactly have I signed? ## A sailboat, an aircraft carrier Millspaugh had kept his prompt vague on purpose. He did not say which database to use, which login service or which framework. He left those choices to the machine and watched what it did with them. Lovable finished in 4 minutes and 55 seconds. Only by clicking through the code did he find that his single page had become a fullstack React application of “dozens of files”, which he thought “a bit heavy”. The comments were stored in Supabase, “which required no intervention on my part to set up.” And then, he wrote, “There’s a whole lot more—analytics, e-commerce and AI integrations, etc.” The sign-in for comments did not work. Publishing to a lovable.app address cost nothing, but a custom domain required the Pro plan. Nielsen wanted to know what “a whole lot more” contained. Even a mostly static site built this way, he noted, arrives carrying a JavaScript framework, a router, a hydration runtime, analytics and component bundles. “You asked for a sailboat, but you got an aircraft carrier,” he wrote. “Do you know? Or more importantly, do you care?” One reflection, offered as nothing more: large establishments are rarely founded by those who wanted them. They grow from small requests that nobody bothered to refuse. ## The man who knows a man With v0 things moved faster. Eleven seconds after he pasted his prompt, Millspaugh wrote, the tool asked whether he would like to install Neon, “because my comment wall needs a real database.” He agreed. Did that mean he needed a separate Neon account? “Vercel creates one for me, and later I’ll find a welcome email from Neon in my inbox.” The comments worked, though with no authentication and no moderation. For Nielsen, this is where it gets interesting. Picking React over Angular is one kind of decision. Picking the company that will hold your users’ logins, whether Auth0, WorkOS or Clerk, is another, and he doubts a non-technical customer would know how to weigh it. Why Supabase and not Firebase? Why Neon? He reaches for an older trade, the real estate agent, who knows a mortgage company, an inspector, a roofer, a title firm. You can choose for yourself. The agent is there to help you choose, and the agent, he observes, very likely does well out of referrals. Nielsen offers no evidence that Supabase or anyone else has paid an AI builder for placement, and he frames the idea as a question. Did the model reason its way to Supabase? Was it simply the statistically likeliest next word? Or had Supabase paid Lovable, in which case he hopes it was disclosed? His bleak observation: “If LLMs are already unpredictable and non-deterministic, nothing will make their actions more predictable and deterministic than some financial incentives.” Then the question underneath all the others: is the customer the person generating the website, or the company willing to pay to be written into the code? ## The currency of the realm Then there was money, or rather the thing that stood in for it. “I’m using up ‘credits’ as the machine works—0.70 so far, now 2.30, now 3.10, oh boy—and I have no clue what that means,” Millspaugh wrote of Lovable. “There are no units. I’d prefer dollars. Are credits just dollars? As it turns out, no.” Their value, he found, depends on the plan you are on. On v0 the build kept stopping. The first stop came after 5 minutes and 13 seconds, for 0.075 credits. Then came two more clicks of “Continue generation”: “5 minutes 36 seconds for 0.070 credits, and 1 minute 29 seconds for 0.024 credits.” Altogether the site took about fifteen minutes. What any of it cost in money, the screen did not say. Nielsen’s complaint is plain. Show the cost in dollars as the prompt runs, not a fractional scrip the customer must convert in his head. “Computers are good at math.” Hosting companies, he writes, pitch credits as a convenience because they can be spent across services, on image transformations or serverless compute. He has an answer to that: another abstraction already works across all those services. “Dollars and cents.” Governments that pay in their own coupons tend to be the ones that would rather you did not compare prices. Companies may be no different. That is a reflection, not a finding. But the customer in this story ends each session not knowing what he has spent. ## Ten minutes, and after Fairness requires the rest of the record. Every one of the six tools produced a working small site in roughly ten minutes. ChatGPT Sites was the fastest at 4 minutes and 17 seconds and gave clear instructions for pointing a domain at it. Claude wrote a few lightweight files and gave Millspaugh the design he liked best. Replit took about six minutes and produced another heavy React app. Val Town’s Townie took about three minutes, with what its own employee called an average design. Results, he warned, vary a great deal from one run to the next, and building a good or large site still takes real work. I am both optimistic and pessimistic about this moment for making websites. I’m optimistic because end-user programming, the idea that everyone should be able to customize their own software, seems more possible than ever. I’m pessimistic because a slopsite, an AI-generated website with little to no human thinking involved, is often worse than no website at all. Nielsen turns the last sentence into a fill-in-the-blank and leaves the reader to supply the noun. His own conclusion is narrower. Generating a website may be easier than ever. Knowing what you generated, whom you now depend on, and why, is not. Somewhere in Millspaugh’s inbox sits the welcome email from a database company he never asked to join.