Testing landing pages with AI means generating 20 or more variants from one prompt, rotating them between visitors, and letting per-variant conversion data pick the winner. Alec Saluga's lead-gen site Carship Pros runs this live: one variant converts at 30%, another at 0%, traffic shifts to the winners, and the top 5 get new versions every cycle.
Key takeaways
One prompt builds the whole system: "Make me 20 landing pages and rotate them, and give me an admin panel so I can view the analytics" (Alec Saluga, Session 008).
On Carship Pros, variant V3 converts at 30% while V1 converts at 0%. Without rotation data, either one could have been the only page.
The loop is continuous: direct traffic to the winners, take the top 5, build new versions of those, and keep splitting.
Rising CPMs on Meta and Google are the reason to do this: the clicks are paid for either way, so conversion is where the leverage is.
Polish is a variable, not a virtue. Ian Kilpatrick once had a client ask him to make trailer-park websites look cheaper, because too nice read as too expensive.
Why rotate 20 pages instead of A/B testing 2?
An A/B test answers one question slowly. Rotating 20 to 30 variants answers the real question, which of many possible headlines, hooks, and layouts converts, at the speed the market answers it. Alec Saluga calls his version an A-B-C-D-E-F-G test: every visitor to Carship Pros gets a different variant, and the admin panel scores each one.
The numbers on screen during Session 008 made the case by themselves. Variant V3 was converting at 30%. Variant V1 was converting at 0%. Any business running a single landing page is unknowingly running one of those two, and paying the same for the clicks either way while CPMs on Meta and Google keep climbing.
What is the prompt that builds it?
One prompt builds the entire system, and Alec gave it away on the call: "Make me 20 landing pages and rotate them, and give me an admin panel so I can view the analytics. That's it." Any of the current builders takes it: Lovable is what Carship Pros runs on, and Claude Code, Codex, and Replit all handle the same job.
Before AI builders, this system was not expensive, it was impossible in practice. Twenty landing pages in a drag-and-drop world was a project nobody green-lit. Now the constraint is having enough traffic to read the results.
How do you read the results?
Treat the panel as the only opinion that counts. Everyone has a theory about which page will convert; Alec's framing on the call was to ask "where's the data?" and let it answer. The loop is mechanical: push traffic toward the variants that convert, cut the ones that do not, take the top 5, generate new versions of those, and keep splitting.
The read gets sharper every cycle because the losers are gone and the winners have children. That compounding is the entire reason to run rotation rather than a one-off redesign.
Do the pages need to look polished?
Polish is an audience decision. The Carship Pros variants are deliberately plain, simple form-fill pages for a lead-gen business, and that plainness is on purpose. Ian Kilpatrick's story from his agency years makes the point better than any principle: a family of trailer parks asked him to make their sites look cheaper, because too nice was signaling too expensive to the renters they wanted.
The rotation system makes this an empirical question anyway. If your audience converts on polish, the polished variants will win. If they convert on plain, you will find that out for the cost of the clicks you were already buying.
The repeatable order of operations
If you take one thing, take the sequence. It works because each phase feeds the next.
Prompt a builder for the whole system
The prompt, verbatim from Session 008: make me 20 landing pages and rotate them, and give me an admin panel so I can view the analytics. Lovable, Claude Code, Codex, and Replit all take it. Give each variant its own headline, hook, and layout, with the same lead-capture form.
Each visitor gets one variant, assigned evenly and remembered, so two people sharing a link can see two different pages while each person's experience stays stable.
Read the per-variant panel
The admin panel scores visits, submissions, and conversion rate per variant. On Carship Pros that spread ran from 30% (V3) to 0% (V1), a gap no amount of opinion would have predicted.
Shift traffic to winners, cut losers
Direct paid and organic traffic at the converting variants and turn the dead ones off. The clicks cost the same either way; conversion is the only number that moves your cost per lead.
Rebuild the top 5 and repeat
Generate new versions of your best 5 and keep splitting. The system gets smarter every cycle because each generation starts from proven winners.
Frequently asked questions
How do you A/B test landing pages with AI?
Generate 20 or more variants and a rotation system from one builder prompt, send traffic, and read per-variant conversion in the admin panel. Shift traffic to winners, cut losers, rebuild the top 5, repeat. Carship Pros runs exactly this, with variants ranging from 30% to 0% conversion.
What is the exact prompt?
Verbatim from Alec Saluga on Session 008: "Make me 20 landing pages and rotate them, and give me an admin panel so I can view the analytics." Add your business context and require every variant to share the same lead-capture form so conversions stay comparable.
Do AI-built landing pages need to look designed?
Only if your audience converts on design. The Carship Pros pages are deliberately plain form-fills for lead gen, and Ian Kilpatrick has had clients ask for cheaper-looking pages on purpose. Rotation turns the question empirical: run polished and plain variants and let the panel decide.
Tools used in this post
Every tool here has its own page with pricing, who used it live, and honest alternatives.
16:00 Alecwhat you'll see here is, like, an ABCDEFG test. And basically what that means is we're rotating different variants of the website.
16:00 Alecmeta ads, Google Ads, all... CPMs are going up and up and up, and that's going to continue to be the case
17:00 Alecyou can see V3 here, this is getting 30% conversion rate. You can see... you know, V1 is getting 0%.
17:00 AlecWe're gonna take our top 5, we're gonna make new versions of those, and continue on split testing.
18:00 Alecthis is one prompt, literally. Make me 20 landing pages and rotate them, and give me an admin panel so I can view the analytics. That's it.
19:00 Alecwe're getting 30 different landing pages in front of people, and we're getting that feedback in real time on what's converting and what's not
20:00 IanThere's actually a tactic to making it look cheaper, so people don't think that you're too expensive.
Who wrote this
Every AI4NTP post is written by an operator who was in the room when the work happened.
A former B2B salesman with no technical background who self-taught AI. He builds and deploys AI-driven marketing and websites, and has grown a following of over 15,000 teaching AI adoption.
Founder and host of AI4NTP. He sold his first company from his college dorm room, and as a fractional CMO has helped scale multiple businesses past $50M in ARR.
A designer, developer, and serial entrepreneur writing code since age 10. He has worked with Disney, the Golden Globes, and the AMAs, and now runs a fleet of AI agents doing real work every day.
