AI Website Design Service: What Is Actually Delivered [2026]

AI website production — what it means

AI website production means building a website through a process in which AI handles construction and humans handle finishing and judgment, and which ships as standard with a structure you can keep operating alongside AI after launch.

Hi, I’m Masato — a Webflow specialist who keeps a constant eye on the leading edge of AI. I run a web production company called Webharu in Shiogama, Miyagi, Japan.

When you hear “website production with AI,” what do you picture?

An auto-generator that just asks you questions in a chat window. A service where AI pours text into a template. The more seriously you are considering commissioning an AI-built website, the stronger the worry: “I get that it’s cheap and fast, but is the quality really going to hold up?”

Honestly, what we actually do looks quite different from that picture.

Let me start with the conclusion.

Our AI website production runs on a process where the AI does the work and a human makes the decisions. What gets delivered is a complete website checked by a human against the same quality bar as conventional production, plus — as standard — a structure you can keep operating alongside AI after launch.

Here’s the whole thing.

How is this different from an “auto-generator”?

The single biggest difference between our “AI website production” and a tool that finishes the job from a questionnaire, or a service that pours AI text into a template, is whether a human checks it before it goes live.

A small or mid-sized company’s website is a dense pile of information you cannot get wrong: pricing, past work, company details, contact information. Chase only the speed of auto-generation and exactly this kind of fact-checking is what falls out.

In our process, the AI does most of the building. But deciding the concept and the structure, judging the design aesthetically, verifying the facts in the copy and giving the green light at each stage are always human work. The speed comes from “the AI does the hands-on work,” not from “nobody checks.”

Concretely, that means situations like these.

• Leave the generation of past work and figures entirely to automation and you can end up publishing projects that never happened or prices that are out of date. We verify the facts before launch.
• Defer the basic SEO setup to “later” and the site stays hard to find, in search and in AI search, for a while after launch. We fit it as standard from day one.
• Lock the design from a single option and the result tends to look mass-produced. We generate multiple variations and a human narrows them down.

Most auto-generators don’t go as far as basic SEO configuration, analytics, or structured data for AI search. The assumption is that you add those after launch. Our AI website production inverts that order. Fit it as standard from the start, so that after launch you only have to think about growth. That is why we can say the deliverable has been through the same level of checking as conventional production.

And human judgment doesn’t stop at the moment of launch. How do you read the numbers in Search Console and GA4, and what do you fix next?
Deciding that is, for now, still human work. The AI produces options fast; the human chooses. That relationship doesn’t change after launch either.

What technology are the delivered sites built on?

Technology first. We select the fastest configuration per project from three options: Astro, Next.js and Webflow. We don’t fix on one technology because the right answer changes with the project’s scale and update frequency. A production company that says “we only use this one technology” is, put another way, fitting projects to its technology. We do the opposite: we fit technology to the project.

For the record, our own site is Astro. We chose it for its compatibility with Cloudflare, its loading speed and its SEO performance.

The rule of thumb is simple.

• Low update frequency, display speed above all → Astro
• Existing React assets, or complex interactive screens likely → Next.js
• The client wants to edit it themselves without code, or values a no-code track record → Webflow

We choose backwards from how the site will be used, not forwards from the name of a technology.

What we adopt on most projects is a setup called “static generation plus CDN edge delivery.” It sounds technical; the meaning is simple. Instead of assembling the page on a server at every visit, we prepare finished pages in advance and serve them from whichever server is closest to the visitor. That is why display is fast, server maintenance never arises in the first place, and the whole thing is hard to break.

There is no program running continuously on a server. Which means there is no software to update and no vulnerability to breach, either. The fewer parts there are to break, the less breaks. I think that’s intuitive even if you aren’t an engineer.

Measured figures from our own site

From here it’s my own experience. Our company site is built the same way, and measures at 27 requests, roughly 1MB transferred, with mobile Lighthouse scores of 98 on the top page and 97 on article pages.

Honestly, those scores were not there from the start. The first measurement was 73. The cause was a 1.7MB web font used for headings; subsetting it down to only the characters actually used lifted the score to 98 without changing a pixel of the design. It took an hour. Whether or not you can be bothered to “measure, then fix” is what produces that gap.

Note that the figures in this article are, as a rule, our own measurements and results. Technology and process differ from one production company to another, so they don’t apply to everyone’s work.

Why running costs are structurally lower

The other advantage of this setup is running cost. As described above, there is no program running continuously on the server side, so maintenance as an activity never arises. Our own site now costs ¥1,500 a year — the domain, and nothing else. The Webflow subscription we used to run cost close to ¥100,000 a year including localization, so changing the technology changes the order of magnitude of the running cost.

What “building with AI” actually looks like

Next, the substance of “building with AI.” Our standard process has four stages and begins with a single 60-minute discovery session, where we go through the goal, the page structure and the assumptions behind the content. Then AI build: the AI constructs every page in one pass and a human finishes it. Then review and revision: you check the site before launch, and we revise up to twice. Finally launch, including DNS configuration and analytics installation.

Here’s a real example. Boiling this process down, the full rebuild of our own site took seven days (four days planning, three days building). That covered building 26 new pages, migrating 62 articles and 24 project entries from the old site, connecting four forms, and setting up analytics, structured data, the sitemap and llms.txt. The whole account is in the seven-day full rebuild article.

That speed wasn’t there from the beginning, though. It took about a month before the design generation was any good — a period of mass-producing garbage with AI. The human role is to pick the good ones out of that mass production, verify the facts, and give the green light at each stage: deciding the concept and structure, judging the design aesthetically, checking the facts and the commercial judgment in the copy, reviewing each stage. AI does the work, humans decide. That division doesn’t change on client projects either.

Won’t everything built with AI end up looking the same?

A common worry. The answer: we don’t simply use whatever the AI produces first. For our own site rebuild we generated about 50 variations and narrowed them down. The AI prototypes in bulk, and a human selects, cuts and refines aesthetically. That is a fundamentally different process from pouring text into a uniform template.

How the process changes on client projects

On our own site we spent four days on planning; on client projects, the 60-minute discovery session takes on that role. In 60 minutes we settle the “foundation for decisions” — concept, page structure, priorities — and everything after it (AI build, review, revision, launch) runs on that foundation.

Sixty minutes may sound short, but it is a length designed around the fact that most of those seven days on our own site were spent deciding. Which is exactly why what you share in the discovery session shapes the result so heavily.

One example. The corporate site for PonTech Lab Inc., an AI startup in Tama, Tokyo. Six pages — home, company, services, careers, contact and so on — plus news and blog update features. We chose Next.js (React) and put it on Cloudflare’s fast delivery. We started in March 2026 and by April it was live with analytics fully set up.

What we put real care into on that project was the page transition: an animation where tiled blocks spread from the lower left to the upper right as the screen changes. The other was the careers page. We handled it from the planning stage, aimed at recruiting former Self-Defense Forces personnel, starting from almost no usable photography. So every visual on the careers page was generated with AI alongside the site build. It’s also a project where AI image generation flipped “we can’t build it, we have no assets” on its head.

The reason it ran longer than the one-to-two-week minimum is that it included that transition work and the planning and asset creation for the careers page. And on this project the two revision rounds went almost entirely unused. Because we take the work to a high level of completion before showing it, revisions basically don’t arise. Updates continue today: between companies, email is the reliable and fast channel, so we take requests by email and apply them.

What “a human finishes it” concretely means

I keep writing “the AI builds and a human finishes,” so here is the substance of the finishing. We test on real devices: a MacBook and an iPhone. Whether links and input forms actually work, the tone and consistency of the copy, a re-check of anything factual such as pricing and company information — we of course go through all the usual items.

On top of that, what I pay particular attention to is the “rhythm” of the design. How the whitespace falls, the intervals between elements, the flow as you scroll. A design that the AI merely produced almost never lands cleanly on the first try. So we judge the design and the animation by their behavior, implementation included, and fix them by hand where needed. It’s unglamorous work, but a human not skipping it is what holds up the quality of AI website production.

What gets asked in the 60-minute discovery session

The discovery session is the starting point of the process, so let me share its contents up front too. There are six things we ask about.

• The purpose of the site (more inquiries? recruiting? something else)
• The readers you are targeting
• Existing sites you’d like to reference, and the design direction
• The pages you want and their priority
• How much copy and photography you already have
• When you’d like to launch

The more concrete this is, the more accurate the AI build becomes and the more likely the two review rounds are enough.

That said, “we’ll leave it to you” is a perfectly fine starting point too — we can begin by organizing the direction together. If something isn’t decided, just say so. In that case we’ll put up a first draft along generally sensible priorities and work from there.

The full deliverables list

If we’re going to use the phrase “as standard,” we shouldn’t hide what’s in it. Here is exactly what is included in AI website production at ¥1,000,000 (excl. tax) and up (the formal figure is fixed by the quote after the discovery session).

Item	Details
Discovery	60 minutes × 1
Site build	A complete AI-built site (standard configuration)
Technology	The fastest configuration selected from Astro / Next.js / Webflow
Display speed	Tuning via static generation plus CDN edge delivery
Foundation for AI operation	Structured data, llms.txt, an AI-readable structure
Basic SEO	metadata, sitemap.xml, robots.txt
Analytics setup	Search Console, GA4
Forms	One contact form
Launch work	Going live, including DNS configuration
Post-launch cover	One week of defect support (up to two minor fixes)

Let me unpack the jargon too.
“Structured data” is markup that presents a page’s contents in a form AI and search engines can understand easily. “llms.txt” is a file that tells AI the key facts about a site. Both raise how easily AI understands the site without changing anything a human sees.
“Search Console” records how you appear in search results and “GA4” records what visitors do on your site — both free Google tools. Installing them is included.

Let me be specific about “one week of defect support” as well. It covers the defects that surface right after launch — broken layouts, dead links, form submission errors — with up to two minor fixes. Large changes such as rebuilding the page structure from scratch fall outside those two. Launch is where it really starts, so treat that week as a period for confirming together that everything works.

The three lines I’d especially point at are “Foundation for AI operation,” “Basic SEO” and “Analytics setup.” Structured data, llms.txt and installing GA4 and Search Console are frequently a paid extra in conventional production, or not offered at all. We fit them as standard.

The reason is simple: we build sites on the assumption that from the moment they launch you measure the numbers and improve them with AI. They are not things to add later; they should be there from the start. Not “we’ll answer if asked” but “we write it before you ask.” That, too, is part of the transparency we believe in.

With that “foundation for AI operation” in place, what can you actually do after launch?
Things like these.

• Consult an AI about the site as a whole
• Feed Search Console and GA4 numbers to an AI and have it propose improvements
• Review all your articles in one pass
• Have it build a strategy for the keywords to target next
• Add and revise pages faster and cheaper
• Get found by AI search (ChatGPT and the like) too

All of these are methods we actually use in operating our own site.

Note, though, that bulk article reviews and keyword strategy apply to sites with a blog (a CMS), and a CMS is not included as standard. For adding one, or for article operations, talk to us in combination with article production or AI adoption support.

How quality is assured

“The AI builds it” makes a lot of people nervous about quality. So let me set out the mechanism honestly.

Every page the AI builds is checked by human eyes before launch. Verifying facts — pricing, past work, proper nouns — is human work too. AI is fast at writing copy and at listing past projects, but judging “is that number actually correct?” isn’t the AI’s job, it’s mine. Giving the green light at the end of each stage is also human.

Precisely because the process uses AI, what we watch for is content that doesn’t match reality slipping in. As long as you use AI, the possibility of generating a project that never happened or an inaccurate figure never reaches zero. Which is exactly why we have not removed the step where a human does the final check on pricing, past work and proper nouns.

Including two rounds of review and revision in the price comes from the same reasoning. We don’t assume the AI builds it perfectly on the first pass. Rather than demanding perfection in one go, building the process around checking and fixing produces a faster and higher-quality result — that’s what experience tells us.

There is a reason the count is two, as well. The first round is for structural or big-picture mismatches, the second for wording and detail. Splitting the roles means a limited number of rounds still covers everything. Instead of making it unlimited, we hold the design of “what gets fixed where” in advance.

A word on transparency too. For our own site we publish the production process itself, with the real git history, at /ai-making. How we instructed the AI, what it generated, what we fixed. From the prompts themselves through the generated diffs and the revision exchanges, you can follow the behind-the-scenes of production that is normally invisible. We can show the process without hiding it because we are that confident in it. If you’re curious, take a look after reading this article.

Honestly: what we can’t do, and who it isn’t for

Everything so far has been the good news, so let me be honest about what we can’t do. Four things are not in the standard configuration.

• A CMS (blog update feature) is not included as standard. We are happy to discuss adding one (+¥50,000).
• Multilingual support is not included. It is quoted separately.
• Fine-tuning of original design has limits. This is production that finishes within the range of designs the AI generates, so detailed craftsmanship down to the last texture is out of scope.
• Ongoing operation after launch is not in the price. It is covered by the Site Care Plan (from ¥15,000/month).

The third one especially deserves candour. If you want to build a brand slowly and deliberately, down to the texture of the details, our conventional “website production” (Figma × Webflow) suits you better. Designing in Figma and then implementing in Webflow gives a different level of design freedom.

The price difference reflects how much effort goes into building the design from zero. Where it really shows is how the cost scales up and how the timeline stretches. AI website production is from ¥1,000,000 with a one-to-two-week minimum; website production (full-scratch build) is from ¥1,000,000 and roughly 1.5–2 months. Launch fast and grow it with AI afterwards → AI website production. Build a brand deliberately → website production. Choose by your goal.

Pricing and timelines

AI website production is priced in three tiers (excl. tax, showing the lower bound).

Plan	Price (excl. tax, from)	Typical timeline
AI Website Production (standard configuration)	¥1,000,000〜	From 1–2 weeks
Full-Scratch Website Build	¥1,000,000〜	About 1.5–2 months
Web Application Development	¥2,500,000〜	About 2–4 months

All figures shown are the lower bound. What moves the number is requirements such as page count, multilingual support and volume of migration, and once a quote is fixed we don’t change it. We can give you a rough range for your own conditions on the spot in a 30-minute free consultation. Note that the larger the project, the lower the per-page rate. The full pricing picture is on the pricing page.

Beyond page count, cost changes when you add options. As a guide: adding a CMS (blog update feature) is +¥50,000, multilingual support +¥150,000, migration from an existing site +¥80,000, photography +¥70,000, and API integration with external systems +¥200,000. The CMS and multilingual support mentioned in the “what we can’t do” section can be added in exactly this form.

The price gap between options is simply the gap in work. A CMS costs what it costs because it designs for ease of updating after launch; multilingual because the copy and the page structure are built twice over; migration because the old site’s data is sorted and carried over one item at a time.

If the number of pages you want grows after we start, we re-quote at the figure for the relevant band. The same applies if the scope ends up smaller than first agreed.

In closing: what kind of company is this for?

The conclusion, once more.

Our AI website production (AI website design) is built through a process where the AI does the work and a human decides. What is delivered is conventional-standard quality that has been through human checking, plus a structure you can keep operating alongside AI after launch, as standard.

There is no special magic in the technology, process, deliverables or quality mechanism described above. It is an accumulation of patient checking.

• A good fit: you want to launch first and grow the site with AI while watching the numbers; you want to cut the time and cost of asking an outside party for every update; you don’t want to defer SEO and AI search readiness.
• Not a good fit: you want to build a brand world slowly and deliberately; you are assuming free CMS editing, multilingual support or e-commerce from the outset.

Conversely, if you want to build a brand slowly, we recommend our conventional website production. First-hand information on service contents and pricing is on the AI Website Production page, and our past work — all of it named — is on the work page. For how far the post-launch operation can be automated, see the operations platform (Japanese) page as well.

The process, deliverables and pricing described in this article are as of July 2026. We’ll tell you honestly which option fits in a free consultation, and we can answer on the spot where your particular project lands.

“It’s AI-built, so it must be cheap.” “It’s AI-built, so I’m worried.” Look inside the process and neither assumption turns out to be necessary. Judge it on the substance, not on the unit price.

Even just telling us what’s troubling you right now is fine. Come and say hello through our free consultation.

FAQ

Doesn’t building a website with AI weaken its SEO?

Building with AI is not itself a disadvantage for SEO. Basic SEO configuration (metadata, sitemap.xml, robots.txt), analytics (GA4, Search Console) and structured data and llms.txt for AI search are all fitted as standard from the start, so they are in place at launch rather than added afterwards.

Is the quality lower than something a human builds?

The quality bar hasn’t changed. AI takes on the speed part of production; the bar is held by checking every page with human eyes before launch and by having humans verify facts such as pricing and past work.

How many days until launch?

From one to two weeks for the standard build, stretching toward two to three weeks as the page count grows. The actual number of days also depends on what is settled in the discovery session and how quickly review feedback comes back.

Who handles updates after delivery?

Post-launch operation is offered through the Site Care Plan (from ¥15,000/month), where updates are handled by simply telling us by email. With the operations platform built in, you also get form handling and automatic news updates from Instagram posts (with a preview check before publishing).