How it works
Timbaly builds a website in four steps with a single confirmation. You fill in one short form, it proposes the complete site structure, you approve it once, and then every page is written knowing where all the others live. In a measured run, eleven pages went from form to published site in 166 seconds.
The order matters more than the speed. Most AI site builders interview you page by page in a chat. That produces four good pages and a broken site: an empty menu, no privacy notice, pages that do not link to each other. Confirming the structure first is what fixes all four at once.
The four steps
Trade, business name, city, contact address. That is the whole input. There is no blank canvas to face and no template gallery to scroll, because both of those are decisions you cannot make well before you have seen anything.
While you are still reading the next screen, three photographs are already being generated in parallel — roughly fifteen seconds. They come first on purpose: the page writer is only allowed to reference images that exist in the library, so an empty library produces text-only pages, and filling it afterwards does not repair them.
Timbaly comes back with the whole site: homepage, the service pages your trade needs, about, contact, the privacy and cookie notices, plus a few suggestions you can decline. You tick and confirm. This is the only confirmation in the entire run.
It is not a convenience. Once the structure is approved, the internal-link service hands every page writer both the published pages and the approved plan rows — so each page knows the URL of every other page while it is being written. One page at a time in a chat, each one is blind to the ones that come after it.
Generation runs concurrently, three pages at once. Around a hundred seconds for a small site.
The legal pages are the exception: they are not written by a model. They come from fixed text filled in with your details, because a model that invents a data-retention period does not produce a page, it produces a liability. Each one says so at the top and asks you to have it reviewed once by a lawyer.
The site goes online with sitemap.xml, robots.txt, llms.txt and an IndexNow key served on your own domain, plus canonical tags, hreflang, Open Graph and JSON-LD on every page. An IndexNow push tells Bing, Yandex, Naver, Seznam and Yep immediately.
Google is the exception, and it is not ours to fix: its Indexing API only covers job postings and live events, and sitemap pinging was retired in 2023. There, sitemaps and internal links are the whole story. The full technical checklist lists what runs by itself.
Measured, not estimated
These are numbers from an end-to-end test in August 2026, on a site with eleven pages. They are here instead of adjectives.
| Stage | What happened | Time |
|---|---|---|
| Photographs | Three images generated in parallel, before the first page | ~15 s |
| Structure proposal | Homepage, backbone, legal pages and suggestions | ~50 s |
| Your confirmation | Tick the pages you want. The only manual step | your call |
| Generation | Remaining pages, three at a time | ~100 s |
| Total | 11 pages, navigable, 9 of them illustrated | 166 s |
Generation is not deterministic. In those test runs roughly one page in ten failed on a model loop; it stays in the plan with a Retry button and the rest of the site is unaffected. We would rather publish that number than let you discover it.
After launch
The first version is the easy part. What decides whether you get found is what happens over the following months, and this is the part that usually stops.
The site scans itself and comes back with concrete actions: rewrite this page, merge these two, redirect that one, write the one that is missing. You approve them one at a time. It reads existing measurements rather than inventing new ones.
Search Console tells you which queries already bring impressions but sit between position eight and thirty with few clicks. Those are the pages to rewrite first — the traffic exists and a few positions are missing — and they become plan rows directly.
On the Business plan a nightly job generates and publishes the approved rows on the dates you set, with nobody at the keyboard. On Free and Pro you generate them when you want them.
Two other ways in
Instead of the form, hand Timbaly your existing site. It reads the pages, keeps the content, and takes over the technical layer — head tags, sitemap, structured data, the files at the root. Existing head tags are localised on import, so a Google Fonts stylesheet stops sending your visitors' IP addresses to Google before the cookie banner has asked anything.
If you already pay for Claude or ChatGPT, connect it and write the pages on that subscription instead of the platform's AI budget. You get handed an approved plan, row by row, with the photographs already in the library and the homepage as a style reference — which is a far better handover than pointing an agent at an empty site.
The free plan needs no card and keeps fifteen pages on a platform subdomain. If what comes out is worth keeping, €19 a month puts it on your own domain and makes it indexable.