The $1,000-a-Month Service I Couldn’t Take on Faith
I was at a Claude Code meetup in Budapest, talking to someone about answer engine optimisation: getting cited by ChatGPT, Perplexity and Claude, not just ranking on Google. She pointed me to a service called Rozz.site. They build “AI sites” for you: a version of your website rebuilt for AI engines to read and quote. The case studies were convincing, and the hypothesis felt valid. The price was $1,000 a month.
So I did what I always do. I looked under the hood.
And the thing they were charging a grand a month for wasn’t that complicated. Structured content, a question-and-answer format, some machine-readable files pointing the engines at it. Useful, clearly. But not magic, and not $12,000-a-year magic.
But here’s the part that turned a bit of meetup trivia into a weekend’s work. I’ve spent the better part of a year growing StillMind, my meditation app. It sits at around $300 a month in revenue right now: real money, and a long way short of where I’m taking it. Most of my time goes into talking to users so I’m building the right product, then the SEO and App Store work to get it in front of more people. I’m always hunting for the lever that pushes it past its current ceiling. So I don’t hear about a new way to get discovered and file it under “interesting”. I hear it as: is this the edge?
Two thoughts arrived at once. One: this could be that edge for StillMind, and I could build it myself. Two: if it actually works, maybe there’s an affordable version of it for people like me, still paying a grand a month on faith. But both depend on a word I don’t like leaving unanswered. Works. How would I ever know if it worked?
The Test Is the Easy Part. The Harness Is the Point.
Building the AI site was never going to be the hard bit. Of course AI can build the thing. The hard bit, the bit I actually cared about, was being able to trust the answer at the end. Did this work, or did I just convince myself it did?
That’s the difference between running a test and being able to believe the result. And it’s where AI changed what’s possible for me.
A few years ago, measuring something like this properly would have been a project on its own. You’d spend weeks, maybe months, building the monitoring before you could even start the experiment. Or you’d rent a third-party tool that gives you 80% of the data you need, then fill the last 20% with educated guesses and hope.
That’s the bit that’s changed. AI let me build the test and the instrument that measures it, shaped to the exact question I was asking, not the question some SaaS dashboard was built to answer.
If you’re doing anything with AI, this is the unlock. Not the thing you build. The harness that lets you trust whether it worked.
I keep coming back to that word. Harness. The scaffolding that gives you concrete confidence instead of a hunch. It used to be the expensive part. Now it’s a weekend. I’d already learned a version of this lesson building a system for my App Store rankings: the clever idea counts for little until you’ve built the structure that proves it was any good.
What I Actually Built
The AI site lives at getstillmind.com/ai. It mirrors StillMind’s most valuable content (the material on meditation, anxiety, ADHD and sleep) rewritten in the format AI engines actually prefer to read and quote.
I’m learning to optimise for this properly, because I think it’s where discovery is heading. A decade ago, search rewrote how anything got found online, and the people who understood it early had a real advantage. Answer engines feel like the next shift of that kind: more and more, people ask ChatGPT or Perplexity instead of typing a query into Google, and the sites those engines quote will be the ones built for how they read. I want to understand it while it’s still early, rather than scramble once it’s obvious.
And it really is a different discipline. AI engines lean towards a question-and-answer structure. They handle longer, more conversational queries than a typed search. They have a different appetite for content too: they want the direct answer, cleanly stated and easy to lift. So the AI version isn’t a copy of my pages. It’s the same knowledge, restructured for a different kind of reader.
Then there’s the measurement, which is the half that took the real thought. Part of the harness plugs into ChatGPT, Perplexity and Claude and runs a fixed set of 47 queries I care about, the kinds of questions a real person asks long before they’d ever find StillMind. For each one it records whether my site gets cited, where in the answer, and which competitors show up alongside me. The type and the volume of citations, tracked week over week.
That gives me a “before”. You can’t claim a tactic worked if you never measured the state of things before you tried it.
The Trap: I Could Have Destroyed My Own Rankings
This was the part I had to get right, and I knew it going in. Earlier in my career I worked as a developer at an SEO agency, so I’d seen this exact failure before, and I knew the fix before I started.
The AI site mirrors content that already exists on my human pages, pages that already rank on Google. Duplicate that content carelessly and you create a duplicate-content problem: two versions of the same thing, competing with each other, and Google potentially deciding that neither deserves to rank. I could have spent a weekend kneecapping the SEO I’d spent months building.
So every AI page carries a canonical URL pointing back to its human equivalent. It’s a signal to search engines: this isn’t a rival to the original, it’s an alternate read for machines. And because I’d built the harness, I wired ranking and traffic monitoring straight into it, so I’d know the moment the experiment started eating my own SEO.
Which is exactly what it looked like was happening.
The Five Days I Nearly Killed It
Right as I launched the AI site, another experiment was winding down: a “Meditation Statistics 2026” post that had pulled in a wave of impressions. That wave was naturally tailing off.
But I didn’t see two separate things. I saw one Search Console graph going down. Day one, down. Day two, down. By day five or six of watching impressions slide, every instinct was screaming turn it off, you’ve broken your SEO, kill the AI site before it does more damage.
This is the moment the harness earned its entire cost.
Because I wasn’t staring at one panicked number, I could see the rest of the picture. Clicks and impressions across almost every page were healthy and steady. The only thing falling was that one statistics post, settling back to its normal level after an unusual spike. The AI site hadn’t touched my rankings at all.
So I didn’t panic. I rode it out. I’ve got previous form for killing things too early, and this time the data talked me out of it. My organic search is still on its usual trend, up around 20% a month, now with the possible upside of the AI site compounding on top over time.
Without the instrument, I’d have killed a perfectly healthy experiment to fix a problem that didn’t exist.
Why I’m Writing This With Nothing Proven
Here’s the honest part. I don’t know yet whether the AI site works.
It’s an experiment in progress. The citation data is still inconclusive. I can’t tell you “do this and ChatGPT will cite you 30% more,” because I genuinely don’t know that, and I’m not going to pretend I do.
I’m writing it anyway, because the result was never the point. The point is the way of working. See a tactic you can’t take on faith, build your own version, and then build the part most people skip: the thing that tells you the truth about whether it worked. AI has made that second part cheap enough that there’s no longer an excuse to run on vibes.
That’s a growth mindset with a feedback loop bolted on. You miss 100% of the shots you don’t take, and luck has a habit of turning up when you build the opportunity for it. I plan to be very lucky.
If the experiment does pan out, I’ll have something better than a hunch. I’ll have the data: a growth lever I can lean into for StillMind, and maybe the start of that affordable version for everyone still paying $1,000 a month on faith.
Key Takeaways
- AI’s real unlock isn’t building the thing. It’s that you can now cheaply build the harness that gives you concrete confidence in whether the thing worked.
- Answer engine optimisation overlaps with SEO but isn’t identical: AI engines prefer a Q&A structure, longer queries, and clean, liftable answers.
- Mirroring your own content for AI risks a duplicate-content penalty. Canonical URLs pointing back to the human page stop the AI version competing with your own rankings.
- Always measure the “before”. A before-and-after is the only honest way to know a tactic did anything.
- Build the instrument before you trust, or kill, an experiment. It stops you making panic decisions on incomplete data.