Is selling AI websites legit? Yes, above a line

Yes, and the qualifier matters more than the answer. Selling websites built with AI is a real business above a quality line and a refund queue below it, and the line has nothing to do with which model you use or whether you disclose it.

The verdict: it is legitimate when the file contains decisions a person made and the client leaves with a document they own. Six checks define that, all six are objective, and a project that fails four of them is being sold as something it is not. The rest of this is the six checks and the money behind them.

Where the distrust comes from#

Videos promising ten thousand dollar websites built in forty minutes are easy to find, and the instinct that something is off is correct. The speed is fine. What is off is that the video ends at the point where the work would start.

Two visible patterns feed the suspicion, and both are real:

• A flood of identical sites. The same hero, three feature cards, the same gradient, the same 150ms transition on everything, sold in quantity to businesses who will not see each other's sites for a year.
• No handover. The client cannot edit a heading, cannot move a section, cannot take the site anywhere, and discovers it the first week the seller stops replying.

Neither is caused by the tool. Both are caused by selling generated output as designed output. A buyer who has been on the wrong side of that once is right to ask the question in the title, and the answer they need is a test they can apply rather than a reassurance.

Often the person searching this is the seller, not the buyer, and the worry inverts: whether charging real money for an afternoon's work will eventually come back on them. It will not, if the afternoon produced decisions. It will, reliably, if it produced a generation with a logo dropped on top.

The quality line, as six checks#

Run these on your own last project before you run them on anybody else's. Five passes is a real business. Two is the thing people mean when they say scam.

Three designs that pass the neutral row on purpose, each with its darks leaning one way:

Note what is absent from that table: nothing about prompting skill, nothing about which builder, nothing about hours. Every row is a value in a file, which is why the test works on somebody else's site as easily as on your own. Open a competitor's page, check the neutral in devtools, and you know which side of the line they sell on. Measuring a site in developer tools shows where each of the six values lives.

Pass rate is also the price. Two rows is a $900 site, four is $4,000, six with the document written properly is $12,000, and the band model behind that is here.

Run them in that order, because they are ranked by how fast a stranger sees them. Type and the neutral are visible in a screenshot. Motion needs three seconds of scrolling. The handover stays invisible until the client needs it, which is exactly why it is the row that decides whether they recommend you to anybody.

The three things that make it not legit#

Being fast is not one of them. Using a model is not one of them. These are:

1. Selling an unedited generation as a custom design. The word custom is a claim about who decided, and if the answer is the framework then the claim is false regardless of how good the result looks.
2. Hosting the client cannot leave. A site on your account, with no export, no repository and no domain access, is a hostage rather than a deliverable.
3. No handover at all. If the client cannot change a phone number without paying you, you did not sell a website.

Fixing all three takes one afternoon and no new skill. Write the specification, hand over the repository or the export, put the domain in their name, record a ninety-second walkthrough. The handover checklist for a local client lists each item, down to the Website field on their Google Business Profile. All three complaints are logistics rather than ethics, which is convenient, because logistics can be scheduled.

Worth saying plainly, because the loud version of this argument has it backwards: not one of the three is about whether a model wrote the components. A studio that hand-codes a site, keeps it on its own account and refuses to export is doing the objectionable thing. A freelancer who generated the markup in an hour and handed over a repository, a domain and a specification is not.

Crowded at the bottom, thin at the top#

The cheap end is genuinely crowded, and it is crowded in the way a commodity is crowded. When every seller ships the same defaults, the only remaining variable is price, and the switching cost for the buyer is close to zero because there is nothing in the file that only you could have made.

The art-directed end is far less crowded, and the reason is supply of judgment rather than supply of labor. Far fewer people selling sites can write down a type scale, name a curve with a duration, or state the hue bias in their neutrals than can generate a page. The tools removed the labor constraint and left the taste constraint entirely untouched.

Then the profitability arithmetic, which is the part nobody in the videos does:

• A $600 site and a $6,000 site generate roughly the same number of support emails, the same number of content chases, and the same one awkward call.
• Ten $600 projects is $6,000 of revenue and ten clients. One $6,000 project is the same revenue and one client.
• The cheap end therefore fails on your hours rather than on demand, which is why it reads as unprofitable from the inside and busy from the outside.

So whether this is profitable is a question about your floor rather than about the market. Set one, refuse work underneath it, and the crowding at the bottom stops being your problem because you are no longer standing in it.

The verdict on selling AI websites#

Legitimate, and profitable above your floor, on one condition: the file has to contain something a person decided, and the client has to leave with it in writing. Fail that and the business is real but miserable. Pass it and you are competing against very few people.

Four lines to act on:

• Run the six checks against your last project and count the passes.
• Fix the handover row first, because it is the cheapest to fix and the most damaging to leave.
• Raise your floor to the pass rate you can actually defend on a call.
• Say you used AI when asked, and put no paragraph of justification after it.

That is the whole ethics of this business, and it takes an afternoon. The people asking whether this is legitimate are usually the ones who would pass five of the six checks with one afternoon of work, which is worth knowing before you talk yourself out of a business that is working, because the reason most generated sites look related is a gap almost nobody is filling.

If writing the direction is the part you cannot do from nothing, Thanor is the library for it: exact values per design, licensed for client work. The prompts themselves stay with you, so the specification the client keeps is your own record of the values you used. For the pricing side, the AI-specific bands are here, and the contract and license side of the business is step 6 of how to start a web design business with AI.

Run the six checks on your last project tonight. Then open Lorikeet and read how its direction settles the neutral row in values, down to where its ground color was measured: that is the standard the table asks for. A free account opens the designs marked free, such as Vesperion, and the plans open the rest of the library.