murattunalı.
Reading

Building a website with AI: what it does well, where it falls short.

It is now possible to put a website live in one evening with AI tools. This guide explains what these tools do well and where they fall short, using the numbers from our study in which we measured 40 AI-built sites. At the end there is a ten-point list to go through before you publish.

What kinds of tools are there?

There are three families. The first is platforms that build a site or app through a chat: with tools such as Lovable, Bolt, Base44 and Create, you write what you want, the tool generates the code and publishes the site on a subdomain of its own domain. The second is AI coding assistants that work in a developer’s editor; the developer manages the code and the release. The third is the AI mode of ready-made site builders: builders such as Wix start the site with a chat and let you edit it in their own editor. Starting with AI on Wix is free; connecting your own domain requires a Premium plan.

The numbers in this guide come from the first family: in our measurement article we measured 40 sites published on these tools' subdomains on the same day, against the same nine criteria.

What does it do well?

They are fast at producing a draft. They set up the page layout, the copy and the image placement from your description and publish the site in a single step; on Lovable, publishing is available on every plan and the site goes live at the tool’s own address (project-name.lovable.app). For testing an idea, a campaign page or an internal tool, that speed is a real gain.

The basic tags mostly come in place too: all the sites we measured had a page title and a language declaration, and 98% had a meta description. The total time JavaScript blocked the main thread (TBT) was low, at a median of 44 milliseconds; in other words, the problem is not heavy code but the places described below.

Where does it fall short?

  • Content in the first HTML. On 45% of the sites, the page held fewer than 50 words before JavaScript ran. A browser or bot that does not run JavaScript sees these pages as blank. The question to ask is whether the tool pre-renders the page: some of the Lovable sites we measured were pre-rendered, some were not.
  • Accessibility. The automated check found no violations on only 23% of the sites; the most common problem was insufficient color contrast (68%).
  • Speed. The median PageSpeed Insights mobile score was 76; the largest content took a median 5.8 seconds to appear (LCP), and only 25% stayed below Google’s "good" threshold of 2.5 seconds.
  • KVKK. 8 of the 40 sites set analytics or advertising cookies before the visitor touched anything; a privacy link appeared on only 53% of the sites. We explain what a site needs to publish in our guides to the cookie policy and the privacy notice.
  • Language and multiple languages. 2 of the 11 Turkish sites declared the page as English. If the site is published in several languages, each language needs its own address and the languages need to point to each other; we explain this in our hreflang guide.
  • Structured data. JSON-LD appeared on 65% of the sites; on the rest, search engines and AI assistants have to work out from the page itself who the company is.
  • Security headers. HSTS was present on all sites, a content security policy (CSP) on none. On a site published on a subdomain, this is the platform’s setting.
  • Code ownership and portability. Lovable can sync the project two ways with a repository you own on GitHub, and this is available on every plan; downloading the code directly is on paid plans. Bolt lets you download the project as a zip file and run it outside the tool. On Base44, two-way GitHub sync is on the Builder plan and above. Connecting your own domain on Lovable requires a paid plan. Moving the site elsewhere one day depends on what the tool and the plan allow; ask before you start.

What the eight gaps above have in common is that none of them shows on screen. A page that looks fine can have empty first HTML, nameless buttons or cookies set without asking the visitor; you see these through measurement, not by eye. None of the 40 sites we measured met all nine criteria at once: not a single one had a content security policy. So before you publish the site the tool produced, run the ten checks below once in the browser and once with measurement tools. None of the checks requires coding knowledge; most can be done with the browser’s own tools, with PageSpeed Insights or with the free tools on this site.

When is it enough, and when is it not?

For testing an idea, a short-lived campaign page, an internal tool or a prototype to take to a first client meeting, these tools are often enough. For a company site that expects customers from search engines and AI assistants, a form that collects personal data, an export site published in several languages or an organization with accessibility obligations, the tool is a starting point. Each of the gaps above has to be measured and closed separately; nothing should be called "done" without measuring.

Ten checks before you publish

  1. Open the page source (in the browser, "View page source"): is your text there, or only an empty container? Our Machine reading tool asks the same question.
  2. Is the page’s language declaration right: on a Turkish site, lang="tr".
  3. Does every page have its own title, description and canonical address?
  4. Are robots.txt and an XML sitemap live?
  5. Is there structured data (JSON-LD) describing your company?
  6. Is the text and button contrast sufficient, and does every button and link have a name? The Contrast checker answers the first question.
  7. Is the mobile LCP in PageSpeed Insights below 2.5 seconds?
  8. Are analytics or advertising cookies set on first load without consent? Are the privacy notice and cookie policy linked?
  9. Is your own domain connected, and does your plan allow it?
  10. Is the code in a repository you own, or can it be downloaded; what happens to the site if you leave the tool?

Our approach

We develop this site with AI assistance too; the difference is in measurement and testing: every change is measured against the criteria on this list before it goes live.

If you want to see these answers on your own site, corporate web design delivers design, code and technical SEO from one hand, on a measurable budget.

SOURCES