The Studio
Built by Gabriel Amzallag (@gabrielamzallag) ·
Web Anatomy ↗ webanatomy.ai ✓ Live
€99 First sale, pre-sold
~€150 Monthly stack
551 Screenshots studied
The Receipt Check
Verified: The build is live and the receipts check out.
- ✓webanatomy.ai is live, and the skills pack is public on GitHub under an MIT license (github.com/GabrielAmz/web-anatomy)
- ✓The €99 first payment and the validation-phase numbers are Gabriel's own, self-reported
- ✓The ~€150/month stack (Supabase €20, Vercel €20, Claude Code €90, Firecrawl €20) is itemized by Gabriel, not audited by us
- ✓The 551-screenshot study across 13 homepages is Gabriel's own research
Five years of conversion work inside Qonto, then out alone with Web Anatomy: a library of what actually converts on landing pages, built in public, every line written by an agent. The roadmap has one rule, nothing gets built until someone pays. The first someone paid €99.
Stack: Claude Code · Supabase · Vercel · Firecrawl
The agent can write the code that runs a scoring pass. It doesn't get to decide what counts as a good page. That taste is the product.
Studio is Okane Land’s founder interview series: builders making real things with AI, and the real numbers behind them.
Gabriel Amzallag put up a pricing page for a feature that did not exist. Real price, a “book a demo” button, and a rule: he would only build it if someone paid. Someone did. €99, for the custom benchmark report he used to assemble by hand for consulting clients.
That sale is his whole method in miniature. Amzallag spent about five years running conversion experiments inside Qonto before going out on his own, and he is not a developer. Every line of Web Anatomy is agent-written, on a stack that runs about €150 a month. The one thing the agent never touches is the judgement: what counts as a good page, what goes in a benchmark, what anything costs.
Most founder interviews arrive after the chart goes up and to the right. This one is earlier on purpose. Web Anatomy is in the validation phase, real traffic, real signups, a business model still taking shape in public, and Amzallag answered everything, including the parts most builders skip.
The problem worth building
You spent about five years doing conversion optimization inside Qonto, then went freelance, then built Web Anatomy. What made the recurring problem worth turning into a product?
I was already doing this by hand. Same report for every client: here are the sections worth copying, here’s where your UX breaks, here’s what good looks like. Two things pushed me to build it. The work was identical every time, so it begged to be done at scale. And I kept watching agencies sell beautiful design that didn’t convert. I had the library of what works in my head. Web Anatomy is that library, made accessible.
Why AI pages read generic
Your thesis is that AI ships generic landing pages because it averages its training data. Make the case to someone who thinks AI-generated pages are already good enough.
Two problems, stacked. 1/ The design: emojis in the headings, rounded cards, beige and purple, everything looks the same. 2/ The AI copy on top: em-dashes everywhere, the fake “we don’t do X, we do Y,” the three-step enumeration for everything. That structure feels smart to the model but it exhausts the reader. You end up with a page nobody wants to read.
Pre-selling the feature
You put up a pricing page for a feature that didn’t exist, with a real price and a “book a demo” button, and told yourself you’d only build it if there was real traction. Then you got your first €99. Walk through exactly what happened, and the step by step for a reader who wants to validate demand before building.
I put up a pricing page for a feature that didn’t exist, or only partially. I put a real price, a “book a demo” CTA, and a rule: I only build it if someone pays. Someone did. €99. What I sold was a custom benchmark report, the thing I’d done by hand for clients.
The steps:
- Write the pricing page (any landing page can do the job) to test a value proposition
- Ask for an email to contact users back
- Send real traffic
- Call the users who submit the form to understand their pain
- Deliver exactly what you promised
- Then see how to automate it
The page is the cheapest version of the product, and you can win months by doing that.
Saying the quiet part
You posted, plainly: “I have traffic. I have signups. I have no business model.” Most builders live in that spot and never say it out loud. What did that moment feel like, and what changed first?
I posted it because I wanted to be honest with my community. I like build in public, and it has to include the things you are not proud of.
What I changed first was the question. I stopped asking “how do I get more organic traffic” and started asking “what would one person pay me for today.” Just a different metric.
What to give away
The hardest part of your model: the skills pack and the library are free, but the MCP server and copy-to-site features are moving behind €9 and €19 a month. How do you decide what to give away when the thing you’re known for is the free part?
That’s the hardest part, and I don’t have a clean answer. I don’t decide up front. I talk to users, try things, and watch whether anyone actually subscribes. When you iterate with AI you can just try fast and let people show you. The rough rule underneath it: free is what earns trust, paid is what saves time. But I found that by shipping.
The stack, and the line the agent can’t cross
You build Web Anatomy largely with AI agents. Give me the real setup: the tools, what it costs a month, and where you refuse to let the agent decide.
I’m not a developer. All of it is agent-written. My stack is cheap: Supabase €20, Vercel €20, Claude Code €90, Firecrawl €20.
Where I refuse to let the agent decide: anything a user pays for or reads as my judgement. Pricing, what goes in the benchmark, the copy on the site, and any published scoring.
The agent can write the code that runs a scoring pass. It doesn’t get to decide what counts as a good page. That taste is the product.
The agent can write the code that runs a scoring pass. It doesn’t get to decide what counts as a good page. That taste is the product.
The human in the loop
Your methodology keeps a human in the loop: you review every published analysis rather than fully automating the scoring. Feature or bottleneck?
I’ve automated most of it. A script pulls 200 new sections a week from top websites. But I review every one before it imports. That’s the step I keep. The quality lives in the database, and that’s the whole wedge. The day I stop reviewing and trust it 100%, the product has nothing underneath it that a well-prompted model doesn’t.
What the data showed
Give me a finding from your data that surprised you.
AI pages are worse on copy than normal ones. What happens is companies reach for “AI” in the hero as a shortcut for copywriting. It feels like a selling point, so they lead with it instead of doing the harder work of saying what the customer actually gets. And it makes the copy weaker every time.
I looked at 551 screenshots across 13 big, famous homepages over two years, and most of them tried putting “AI” in the hero. Most of them rolled it back. Grammarly, Webflow, Semrush all led with AI and quietly returned to plain benefits. The ones that never blinked, Canva, Ramp, HubSpot, kept the same promise the whole time (you get more done, you win more customers) and let AI make that promise truer underneath.
So the lesson: “AI” is not a value prop. “You get more clients” is. Sell the outcome in the hero and let AI be the reason you can deliver faster and better.
Gabriel’s full write-up of the study is on his Substack.
Whether the moat holds
You’re building a business where the product is partly taste, the curation an AI can’t replicate. What do you say to someone who argues a well-prompted model closes that gap in a year?
You said it in the question: a well-prompted model can do this. That’s the catch, “well-prompted.” Most people can’t give the agent the right context. And left alone, the agent is lazy. It takes the path of least effort, won’t go dig for the best answer, and hands you the cheapest generic thing it can produce. The gap isn’t the model getting smarter. It’s the work of getting the right context, and that’s the work I’ve already done.
The wrong take
One popular belief about AI and web building that you think is flat wrong?
That AI makes SEO an easy game.
I worked on SEO for years. Now AI makes content trivial to produce, which is exactly why it stopped working. Spinning up hundreds of AI-generated pages a month, especially on a new domain, is the fastest way to get penalized.
Publish a few strong pages a week instead, and earn authority first. Taking it slower at first has a greater impact in the medium term.
Gabriel Amzallag builds Web Anatomy, a data-backed library of what converts on landing pages. He was in-house on conversion and SEO at Qonto for five years, posts on X as @gabrielamzallag, and writes the Web Anatomy newsletter on Substack.
[ Go see more from Gabriel Amzallag →
](https://gabrielamzallag815268.substack.com/) The Studio
Think you belong here? #
The Studio is invite-only, but some of the best interviews started with a builder reaching out. If you have shipped something real with AI and can show what it cost or what it earns, we would like to hear from you.
One email, when there's something worth sending
Get the receipts in your inbox. #
No fixed schedule, no filler. You get an email when we've tested something, run the numbers, or found a tool worth your time.
Free. Double opt-in, unsubscribe in one click.