From Perfect Scores to Downloadables: What We Learned Building ZenMasterWorks
ZenMasterWorks launched on June 18, 2026, with one question: could one person, using technology, the mind, the body, and a small cash investment, figure out Google's own website standard and turn it into a living? Here's what happened, including the parts we'd do differently.
The goal: Google's standard, all four scores
The target was a verified 100/100/100/100 on mobile Google PageSpeed Insights: Performance, Accessibility, Best Practices, and SEO. Not once, but repeatably, across hundreds of pages, sample sites, and platform rebuilds.
The method never changed: study, learn, try, apply, fix, deploy, and repeat until perfection is achieved. Most of the code and writing came from working with AI, with a human deciding what to build, setting the standards, and checking every claim against Google's live tools.
What reaching 100 actually took
Less cleverness than discipline. Inline CSS, system fonts, no frameworks, images sized for phones, scripts delayed until after the page loads, and space reserved for anything that arrives late. The hardest point is usually the last one: in October our homepage sat at 99 because one hero image painted 190 milliseconds too slowly. Shrinking it from 77.8 KB to 23.9 KB brought it back to 100, and we published the incident.
Every mistake goes in the public incident log, numbered and never edited afterward, each one ending with the rule it produced. Some of the most useful rules came from the most embarrassing mistakes.
Why we stopped selling services
Working almost nonstop, day and night, solved the standard, and it also caused real burnout. Services meant calls, revisions, and waiting on other people. So ZenMasterWorks stopped offering services to the public and turned its work into downloadables: a book explaining the whole method, and website templates built to the same standard. A buyer can test any live sample first, pay, and download, without waiting on anyone.
What we got wrong along the way
- Price claims. We briefly wanted to say "the lowest price anywhere on the internet." Free templates exist, so that claim would have been false. We published a sourced comparison instead.
- Prices themselves. We dropped templates to $8 to be the cheapest, then raised them to researched prices: $19 for the book, $29 per template, $99 for the bundle. Cheapest isn't automatically best for buyers or for the work.
- Packaging. One upload bundled private files with public ones and briefly exposed them. The new rule: anything placed in the website folder is public.
- Counts that go stale. Pages that said "371 entries" became wrong the moment entry 372 arrived. Now counts are computed live or dated.
- Mixing selling with teaching. Our guide to superintelligence briefly carried a sales banner and a partnership offer. We removed them; information pages should serve readers, not us.
What building with AI taught us
AI made it possible for one person to build and maintain a site this size. It also made confident mistakes, and it was only as good as the checking behind it. The lesson that held up best: let AI do the heavy lifting, but keep a human responsible for every claim, every price, and every decision about what to publish.
What's next
The products stay up, the method stays public, and the unpaid audits continue, free, for organizations whose websites serve their communities. If you only take one thing from this post, take the method: test your own site, fix the single biggest item the report names, and test again.