AI Website Renewal in 7 Days: Our Full Webflow→Astro Rebuild Log

I’m Masato Kurita, Representative Director of Webharu Inc.

We rebuilt the Webharu website from top to bottom with AI. It took 7 days in total — 4 days of planning and 3 days of building. The one doing the hands-on work was AI (Claude Code); the only human involved was me.

The conclusion first: an AI can fully handle a website renewal. The reason this one took only 7 days, though, is that “what to build” was already completely settled. What information to put on the site, and what design would earn the reader’s trust — that part is still a human’s job.

Key takeaways

Webharu renewed its website with AI (Claude Code): 4 days of planning plus 3 days of building, 7 in total, with one human — the founder
Built 26 pages from scratch, migrated 62 articles and 24 portfolio pieces, connected 4 forms, and set up GA4, structured data and llms.txt
Moving from Webflow to Astro cut running costs from almost ¥100,000 a year to ¥1,500. Lighthouse went from 73 to 98, and 24 of the old 30 articles stayed indexed the day after launch
Japanese-page search went from 641 clicks and 17,920 impressions in June 2026 to 746 clicks and 32,100 impressions in September, after launch (including new articles added in between)
The design was narrowed down through about 50 patterns in a back-and-forth with AI. Scrutinizing the text and checking every page by eye stayed the human’s job

AI website renewal — what it means: AI website renewal is an approach where humans keep the planning, the design judgment and the fact-checking, and hand the page building, migrating existing articles and SEO setup to AI. Webharu rebuilt its own 26-page site this way in 7 days (4 days of planning, 3 of building).

Contents

What the AI website renewal actually involved

We did four things.

Built 26 static pages from scratch (including the cherry-blossom animation on the top page)
Migrated 62 articles and 24 portfolio pieces from the old site (Webflow)
Connected 4 forms, including contact and brochure requests
Set up GA4, Search Console, structured data, the sitemap and llms.txt

I narrowed the target based on my own track record, and rebuilt the site to match it. Most of the 7 days went into planning. The article that turns this into a procedure anyone can follow is the full process of building a website with Claude Code.

The 7-day process of renewing a website with AI

Here is how each day went.

Day 1: Planning and setting the goal

I started from the planning strategy — where the traffic comes from, who the target is, what kind of clients we have had so far. I took stock of all of that and decided what the renewal was meant to achieve.

Day 2: Migrating the old site and setting the design direction

I scraped the old site’s articles and portfolio pieces out and moved them over, checking whether what was currently published still matched the target, and brought the material down locally. At the same time, I settled the foundations of the design here too — color, button placement, and the direction I wanted the site to express.

Day 3: Making around 50 design patterns

I built the new design in HTML — roughly 50 patterns in all. I put the weight on the top page, telling the AI exactly what I disliked about the parts I disliked, and went back and forth making three patterns at a time.

Day 4: The sitemap and the lower-level pages

Once the top page was settled, I made the sitemap. I didn’t make it first because I didn’t want to put odd constraints on the AI. Normally the sitemap comes first, so this was a slightly unusual order. Once the top page’s design is decided, the lower-level pages just follow it, so the design work shrinks a lot.

Day 5: SEO and AI-search optimization, building in Astro

I optimized every page for SEO and AI search, and first built just one page of the article and portfolio detail pages in HTML. From there I moved into building in Astro. I chose Astro over Next.js because I had confirmed that, for compatibility with Cloudflare and load speed, it delivers the best performance as an SEO platform.

Day 6: Connecting everything and scrutinizing the text

I connected Google tags, the forms and Cloudflare, and got it to just short of going live. From here, though, the work of checking the whole site’s text against the plan began — the pages the AI had built were full of wording and numbers I hadn’t intended. I also redefined the direction using the numbers from GA4, and fixed everything in one pass.

Day 7: Checking every page by eye, then launching

Last, I look at every page with my own eyes. The people who come to this site are staff at Japanese companies, or the heads of companies who want to commission a build. We can’t afford to get the company’s own information wrong. I checked everything, then launched.

I’ll be honest, though. It took about a month before I could generate designs well, and along the way AI produced a huge amount of junk. There were times I threw away a half-built site and started again from zero. It took that long to arrive at a setup that runs for free, server costs included.

What changed by moving from Webflow to Astro

We swapped out the entire build system, so the numbers changed clearly too.

Running costs: down from almost ¥100,000 a year on the old Webflow subscription to ¥1,500 a year — just the domain fee
Load speed (Lighthouse, measured on mobile): up from 73 at first to 98 on the top page and 97 on articles, after trimming fonts
Indexing: 24 of the old 30 articles were still indexed the day after launch
Ahrefs health score: up from 6 on the old site to 98 on the new one

The thing I was most careful about in the migration was not changing the articles’ URLs. Keep the URL the same, and Google carries over the ranking signal it has already built up. The detail behind the health score is in how to use an Ahrefs site audit, and how we went from 6 to 98.

What happened to the search numbers after the AI renewal

We launched on July 16, 2026. Looking only at the Japanese pages in Search Console, June 2026, before the renewal, had 641 clicks and 17,920 impressions; September 2026, two and a half months after launch, had 746 clicks and 32,100 impressions.

We also published new articles in that window, though, so not all of the increase can be credited to the renewal. What I can say is that swapping out the entire system end to end didn’t make the search numbers drop. The detailed ranking check is in the article testing whether an AI-built site is weak at SEO.

We did run into two incidents along the way, too: the Korean pages disappeared from search results, and GA4 stopped for two weeks. In both cases the site itself kept displaying normally, so we didn’t notice right away. The full story and how to prevent it is in how a site renewal affects SEO.

What I — the human — was doing

Every call that required thinking stayed with the human. AI only does what it’s told.

Deciding the concept and structure (what the site is there to say)
Aesthetic judgment on the design
Fact-checking the copy, and the business calls (pricing, how to present it)
Reviewing every stage and giving the go-ahead

In other words: AI does the hands-on work, the human makes the calls. That division of labor is what decides the quality.

In a way, I think a company president with strong opinions should build the site themselves. The beautiful image in your head and whether it’s actually beautiful once it takes shape are two completely different things. Cram in everything while it’s still glittering in your head, and you end up with a site that looks pretty but is a mess. My job was to tidy that up.

AI will offer to tidy it up for you too. But humans are inefficient — we reject the AI’s suggestions and keep making things messier. I was the same at first. That’s why deciding the direction up front matters more than anything.

For instance, I originally planned to rebuild this site for large-company projects. But looking back at our work since 2020, large companies made up only about 10% of it, and that work didn’t repeat well. The other 90% were startups, small and mid-sized companies, people about to start a business, and people wanting a renewal. By region, it breaks down to roughly 20% overseas, 60–70% domestic outside Miyagi, and about 10% Sendai and Miyagi. Weighing that against pageviews and traffic sources too, I narrowed the target to startups and small and mid-sized companies nationwide.

What doing it with AI taught me

Three conclusions first.

Fine-grained tweaks are slower than the old days of brute-forcing it with no-code
The wall of an unfamiliar language or technology has almost disappeared (I even shipped an App Store app)
Leave it 100% to AI and every site ends up with the same face

A small tweak spreads into a whole-site tweak

Honestly, I still can’t say for certain whether it made things faster. You can’t just fix the one spot you’re looking at. Say I want to widen the gap between the About and Blog sections a little — in the past I’d simply click that spot and widen the margin. Tell AI to widen only that, and it either doesn’t go well or the design starts to come apart elsewhere. A small adjustment spreads into an adjustment of the whole.

Of course, reviewing the whole thing does produce a cleaner design. But honestly, I suspect the brute-force days were faster. Widening a bit of whitespace is nothing more than my own small preference, and that kind of task piles up and is hard to whittle down. That, I think, is the trap of building with AI.

Leave it entirely to AI, and the brand thins out

Publishing whatever AI makes without touching it is entirely possible. But do that, and the “AI look” comes through strongly, and everyone’s site starts to converge. In the end, it can dilute the brand’s value and leave you thinking, “I’d have been better off building it with no-code.”

The wall of an unfamiliar language has almost disappeared

On the other hand, I can now handle code I don’t know at all. I currently have several apps live on the App Store. An app needs a language called Swift, but as long as I can think through the plan, the strategy and the design, I can now build in pretty much any language.

This time I used Opus and Fable in Claude Code. For building a website specifically, I think AI has already reached something close to the ceiling. Give it a clear instruction for what you want, and it can build almost anything.

If you’re renewing a website with AI, spend your time on the planning

Building a site itself has been something an amateur could do since the no-code era. But what design to use, how to present it, and which target to reach — I think there is still professional work left to do there.

So for a client project, allow at least 2–3 months. The build itself might take 7 days, but that was our own site, where I was the only one deciding and checking. There’s also the path of learning the fundamentals yourself and running the site long-term in-house, but constantly chasing trends and adopting the newest technology can be hard work for some people. I chase trends almost as a hobby, and the sites we run in-house get the newest information applied the moment it appears.

Client sites take at least 2–3 months

Some articles say “it’s fast because we use AI.” In practice, a person reviews the code AI writes, every page is built properly, and we bring all of our knowledge and experience to bear. Even a small site needs time for planning and structure, and a redesign starts by tracking down why the current site isn’t producing results.

What AI changed is not the timeline but the price of large sites: a site that used to cost around ¥8–15 million can now be built for roughly ¥3–5 million. “Are there really sites that expensive?” you might wonder — in deals with large companies, those figures are normal. A large company that handles a lot of data won’t outsource to a firm that takes the job for around ¥1 million.

On top of that, the work starts with meetings. Among the companies I’m working with now, one has been in nothing but meetings for three months. For pre-order meetings that is honestly long, and sometimes it makes me uneasy. Still, I think long meetings are part of how business-to-business deals work, and the stage where our research and answers earn the client’s trust.

AI website production, built the same way, starts at ¥1,000,000 (excl. tax, at least 2–3 months), and the post-launch Site Care Plan starts at ¥15,000/month (excl. tax). The division of labor between AI and humans is also laid out on the AI website production role breakdown.

If you want to renew your own company’s website, reach out through the contact form. The behind-the-scenes details and the follow-up numbers go out first through the newsletter (emails are in Japanese).

Selected work

FAQ

Can a website renewal be done with AI alone?

The build alone, yes. But the planning, the aesthetic judgment on the design, and fact-checking are the human’s job. Even on our own site, the pages AI built contained wording and numbers we hadn’t intended, so we checked everything against the plan on Day 6.

How many days does an AI website renewal take?

Our own site took 4 days of planning plus 3 days of building, 7 in total — but that was our own site, with every premise already settled. For a client project, allow at least 2–3 months.

What changed by moving from Webflow to Astro?

Running costs dropped from almost ¥100,000 a year to ¥1,500 (just the domain fee), and Lighthouse on mobile went to 98 on the top page and 97 on articles. We kept the articles’ URLs unchanged, so 24 of the old 30 articles stayed indexed the day after launch.

Do the search numbers drop when you renew with AI?

They didn’t on our own site. The Japanese pages went from 641 clicks in June 2026 to 746 clicks in September, after launch. That period also included new articles, so not all of the increase can be credited to the renewal.

How much does AI website production cost?

AI website production starts at ¥1,000,000 (excl. tax, at least 2–3 months), and the post-launch Site Care Plan starts at ¥15,000/month (excl. tax), as of October 2026.

Which AI did you use to build it?

Claude Code (Opus and Fable). The division of labor was AI for the hands-on work, and the human for deciding the concept, aesthetic judgment, fact-checking and the go-ahead to publish.