AI-Built Website Pre-Launch Checklist (7 Fronts)

Lovable, v0, Cursor and Claude Code build the screen you asked for. They do not add an llms.txt, a real sitemap or a working 404, and they will cheerfully ship a secret key in client-side code. Across the first 37 AI-built sites run through Tabkeel's exam, 92% went live with at least one of these holes. The typical site had five.

This checklist has two layers. The table-stakes layer every launch guide already covers, kept short. And the layer almost nobody checks: whether the machines that send you traffic can actually read, index and cite the site your AI just built. The second layer is where the money leaks, so it gets most of the page.

The AI-built website pre-launch checklist

Seven checks before you go live, grouped by what each one protects.

• Serve real HTML, so a crawler that does not run JavaScript can read the page.
• Add an llms.txt and let AI crawlers in.
• Make the site indexable: a sitemap, no stray noindex, a correct canonical.
• Structure content to be cited: FAQ schema, author and date on articles.
• Check what AI says about your product against what the site claims.
• Match your pricing page to your terms.
• Fix dead links, soft 404s and pages that are empty without JavaScript.

The rest of this page is those seven fronts in depth, the data behind them, and the fixes. Before them, clear the table-stakes.

Why AI-built sites ship with the same holes

An AI-built website is a site whose pages and code were generated mostly by an AI tool instead of written by hand. The tool optimizes for the thing you described. A screen that looks right, clicks through, demos well. It does not optimize for the readers you never named: the crawler fetching your HTML, the answer engine deciding whether to quote you, the buyer reading your refund terms before paying.

That blind spot is consistent. In our exams the most common failure was a site the traffic machines could not read. Most AI crawlers, including GPTBot, ClaudeBot and PerplexityBot, do not run JavaScript. They read the HTML your server sends and move on. A site painted entirely in the browser hands them a blank page.

The full surface, and what this checklist goes deep on

Clear the table-stakes first. These sit in every launch guide and every reader expects them. Tabkeel does not check these four, so use the tools built for each. Do not skip the first one.

• Exposed secrets. AI coding tools make it easy to paste a service-role key or API key straight into client code, where anyone can read it from the browser. Grep your bundle for keys, move every privileged call server-side, and rotate anything that already shipped. This is the one that ends in a breach, not a bounce.
• Performance. AI builds love a giant hero image and a heavy JavaScript bundle. Check Core Web Vitals in PageSpeed Insights, compress the images, lazy-load anything below the fold.
• Accessibility. Generated markup often ships images with no alt text and low-contrast type. Run a quick axe or Lighthouse pass so screen readers and keyboard users are not locked out.
• Mobile. Open the site on a real phone. Layouts that look clean on a laptop often overflow or trap scroll at 375px wide.

Now the part the other checklists miss. Everything below decides whether the site your AI built can be found, understood and quoted. Group it into seven fronts, which is how a site actually fails after a fast build.

The 7-Front Pre-Launch Check

Each front protects one thing. Fix by front, worst first. A dead link is a nuisance. A homepage that serves an empty body to every AI crawler closes the channel where your buyers now start.

What we found across 37 AI-built sites

This is an early sample, not a formal benchmark, so read it as direction rather than a census. The pattern was one-sided.

The average site carried 5.4 findings. Only 8% came back clean. Almost everything clustered at the top of the funnel, in readability and indexing, which is exactly the surface an AI builder never prompts you to think about.

The empty-page-without-JavaScript trap

The first time I ran this check on an AI-built homepage, the served response came back so short I assumed the request had failed. It had not. The whole page was one empty div. That case looked finished in a browser and shipped nothing to a crawler, and the served HTML was a single mount point:

<body><div id="root"></div></body>

Everything the founder wrote, the headline, the feature list, the prices, existed only after JavaScript ran. A person saw a full page. GPTBot saw the title tag and an empty body. Nothing was broken on screen, and the most important reader got nothing.

Tabkeel reports this with the evidence and a correction written as a prompt you paste into your own agent, not code it drops into your repo. The prompt reads like this:

My marketing pages render entirely on the client, so a crawler that does not run JavaScript sees an empty body. Convert the homepage, pricing and about pages to server rendering or static generation so the served HTML already contains the headline, the feature copy and the prices. Keep the logged-in app routes client-side. Verify by requesting the page with JavaScript disabled and confirming the product description is in the response.

In Lovable or v0 that is a framework setting and a few minutes of work. In a hand-rolled Next.js build it is moving the page out of a client component. The fix was never the hard part. Knowing the hole existed was.

The five checks AI builders skip most

Ten minutes before launch, in order. Each is cheap to fix and common in the data above.

1. Serve real HTML. Request your homepage with JavaScript off. Empty body means server-render the marketing pages. This one closes the AI channel.
2. Add an llms.txt. More than half the sites had none. It is a plain text file at your root that tells AI readers what you are and where the important pages live.
3. Ship a sitemap and hunt for a stray noindex. A single leftover noindex from a staging build hides the whole site. One in six sites in the sample had one.
4. Follow every link and button. Dead links and buttons that go nowhere hit more than a quarter of sites. Click them, or let a crawler do it.
5. Make your 404 a real 404. A soft 404 returns HTTP 200 for a page that is actually missing, which quietly pollutes your index. Confirm missing pages return a 404 status.

Run the findability half in about two minutes

You can work the seven fronts by hand with curl, a JSON-LD validator and Search Console. Or paste your URL into the Tabkeel exam and get every finding with clickable evidence and the paste-ready fix, free and without an account. To test one front on its own first, the AI readability tool runs with no login. After launch, the Search Console analysis turns your ranking data into the pages worth fixing first. The full method behind the seven fronts is written up in the methodology.

Frequently asked questions