By Ryan Richardson · Published 8 October 2026
After three rounds of optimising a single page (new hero, new CTA copy, a full design rebuild) produced no change in outcome, the next test changed the question: instead of asking which headline wins on one layout, it asked which layout wins. Five distinct architectures were built in parallel: a 7-question quiz ending in 3 tiered results, a long-form direct-response essay, a call-first page offering a free 15-minute diagnostic, a page styled like a Reddit community post, and a page styled as a personal founder letter. A sixth, multi-section control page carried its own internal test: 3 price points ($19 / $27 / $47), run as separate ad sets so price stayed a clean signal rather than something the bandit sampled.
Each pattern ran as its own Meta ad set at roughly $15/day (about $90/day total for the 6), with the 3 price variants on the control running alongside at about $30/day total. The pages themselves were served from a Cloudflare Workers edge bandit, so page loads stayed fast regardless of how many variants were live.
At the ad level, the patterns separated clearly on click-through rate and on how many of those clicks turned into a landing page view: long-form pulled the highest ad CTR, the quiz pulled the highest post-click retention, and all 6 patterns beat a 5% CTR benchmark. The 3 price points also separated clearly on CTR. None of that mattered for the question that counted: not one combination of pattern and price produced a purchase, so there was nothing for a purchase-based bandit to converge on.
Building 5 distinct architectures and voice-matching every one of them to the same source material took about 10 minutes using parallel AI drafting agents, instead of roughly an hour doing it sequentially. That speed is what makes a 6-pattern parallel test practical in the first place. But speed only helps if there's a conversion signal downstream to read; running more layout variants doesn't create that signal on its own.
| Claim | Value | Source |
|---|---|---|
| Number of parallel page patterns tested | 5 (plus 1 control) | Landing page velocity.md, 'Landing pages live' list |
| Price points tested on the control page | $19 / $27 / $47 | Sixty Steps bandit test logs |
| Pattern test spend | $90/day across 6 ad sets | Landing page velocity.md, 'Live right now' section |
| Per-price test spend | $30/day across 3 ad sets | Landing page velocity.md, 'Live right now' section |
| Long-form DR pattern CTR | 9.09% | Sixty Steps bandit test logs |
| Quiz pattern click-to-LPV retention | 81% | Sixty Steps bandit test logs |
| $27 price point CTR | 10.00% | Sixty Steps bandit test logs |
| Time to generate 5 voice-matched page patterns with parallel AI agents | approximately 10 minutes (vs approximately 60 minutes sequential) | Sixty Steps bandit test logs |