I checked 183 vibe-coded apps. 92% are almost blank to AI crawlers.
This week I got curious about a simple question: when ChatGPT, Claude or Perplexity "look at" a website, what do they actually see?
Short version: a lot of their crawlers read your page the way curl does. They download the HTML and that's it. They don't run your JavaScript.
That's fine for a WordPress blog. It's a problem for an app built with Lovable, Bolt or Replit, because those are usually single-page React apps. The HTML that ships is basically a title and an empty <div id="root"></div>. Everything a human sees is drawn by JavaScript afterwards.
So I went and measured it.
What I did
I took the list of public app links we collected for an earlier study (apps people shared on Hacker News) and kept the ones hosted on AI builders: lovable.app, replit.app, base44.app, bolt.host, emergent.host and trickle.host. 283 links, 183 still alive (most of the dead ones return 404, apps get deleted a lot).
For each one I fetched the home page, robots.txt and llms.txt. No JavaScript, no logins, nothing clever. Just what a crawler gets. I only kept totals, so no app is named here.
For comparison I ran the same thing on 144 random apps from vercel.app and netlify.app.
The numbers
| AI-builder apps (183) | Vercel / Netlify sample (144) | |
|---|---|---|
| Median words of text in the raw HTML | 7 | 142 |
| Fewer than 50 words | 168 (92%) | 45 (31%) |
Empty app shell (<div id="root"></div>) | 147 (80%) | 23 (16%) |
| Has a meta description | 141 (77%) | 99 (69%) |
| Has structured data (JSON-LD) | 28 (15%) | 20 (14%) |
Has llms.txt | 17 (9%) | 8 (6%) |
robots.txt blocks an AI bot | 4 (2%) | 1 (1%) |
Seven words. In most cases that's the app name and whatever is in the title tag.
The good news is hiding in the last row: almost nobody blocks AI bots on purpose. People aren't hiding. Their sites just don't say anything until JavaScript runs.
Does this actually matter?
Depends who you ask, so here's what I'd point at:
- Vercel measured this at the end of 2024: the crawlers from OpenAI and Anthropic fetched JavaScript files but didn't execute them. Google's crawler does render JS, which is why "my site shows up on Google" doesn't mean much here.
- Pew Research (July 2025) found that when Google shows an AI summary, people click a regular result 8% of the time, vs 15% without one.
I'm not going to tell you Google is dead. It isn't. But more people now ask an assistant first, and the assistant can only recommend what it can read.
Check your own site in 30 seconds
Open a terminal:
curl -s https://your-app.lovable.app | sed 's/<[^>]*>/ /g' | wc -w If the number is tiny, that's roughly what a non-JS crawler sees. You can also just right-click → View Page Source. If you can't find your headline or your pricing in there, neither can the bot.
How I'd fix it (in order of effort)
- Put the basics in
index.html. A real<title>, a<meta name="description">that says what you do and for whom, and a short<noscript>block with your headline, what you sell and how to contact you. Ten minutes. - Add
/llms.txt. A plain-text page that describes your product, pricing and key links. Some assistants read it, and it costs nothing. - Add JSON-LD. A small
<script type="application/ld+json">with your organization, product and price. Agents parse this much more reliably than prose. - Pre-render the important pages. This is the real fix: static HTML for your landing and pricing pages, so a crawler gets actual content. On Vite/React you can ask the builder for a prerender step.
If you use Lovable, a prompt like this gets you most of the way:
Add SEO for AI crawlers: put a full title, meta description and Open Graph tags in index.html; add a noscript section in index.html with our headline, a 3-sentence product description, pricing and contact email; create public/llms.txt describing the product and main pages; add JSON-LD (Organization + Product with price) to index.html. Don't change the visible design.
I did this on our own site today. Our checker gave it 76/100 before (no llms.txt, no structured data) and 100/100 after. It took one small commit. Our site already rendered on the server, so we were luckier than most Lovable apps.
What this doesn't do
It doesn't guarantee ChatGPT recommends you. Nothing does. It just makes sure that when an assistant looks at your site, it finds something to work with instead of seven words and a loading spinner.
If you'd rather have someone check it for you: I run these checks and write the fixes as a small service on Fiverr. Questions in the comments are fine too, happy to look at what you're seeing.