I asked what it takes to sell websites at volume. The report came back naming the platform I spent August leaving, and calling it mandatory. The uncomfortable part is that it is right.
The question I asked
This week I ran a gap analysis across nine sources: six recordings of operators who sell AI-built websites and microsites at volume, and three weekly research digests covering per-tenant economics, solo-operator monetisation and AI-SEO. Four hundred and seventy kilobytes of other people’s experience. I asked six specific questions. The first was what the offer actually is and what it costs. The last was what none of the sources answer.
Every one of those recordings was something I saved months ago and never watched. There are more than two hundred others sitting in the same queue.
What came back
The report lists the capabilities an operator needs in place before running this at volume. The first line of the infrastructure list is a GoHighLevel Unlimited plan, marked mandatory.
I cancelled that trial on 12 August. I wrote two Build Logs about why: I could not calculate what a single client would actually cost me, so I left.
The figures behind the recommendation are large, and every one of them is a claim rather than a measurement: 1,300 clients managed by three people, $509,000 in monthly recurring revenue, a 70% margin, sites built for ten to fifty dollars each and carried for about $2.50 a month. Churn of eight to twelve per cent. I cannot verify any of it. Neither can the report.
Why the recommendation is correct
The model those operators run rests on one rule: never do bespoke work. The report is explicit that margin is destroyed the moment a manual update or a custom feature enters the process. Everything else follows from that. If nothing is ever bespoke, the platform only has to be cheap per tenant and identical every time.
An all-in-one is the right answer to that question. It was the wrong answer to mine. Two weeks after I left I had a composed stack I could add up, at roughly thirty-three Australian dollars per client per month.
Both things are true at once. I left because I could not price one client. The volume model never prices one client. It needs the marginal client to cost nothing.
The three questions nobody answered
The last section is the only part I have not read somewhere before, and all three of its gaps are Australian.
A thousand-site portfolio carried at $2.50 a month assumes .com domains. A .com.au requires ABN verification, so the arithmetic that turns the portfolio into a moat may not survive the trip here.
Autonomous lead scraping at that volume sits unreconciled with the Privacy Act 1988 and the Spam Act. The recordings treat the scraping as a solved tooling problem.
And across nine sources there is no figure at all for what a self-hosted stack costs in human hours at 1,300 clients. That is the number my own stack would have to survive.
What I changed
Nothing in the build has changed yet. The research did not overturn the decision. It changed which decision I think I am making: not which platform to run, but whether I am selling depth to a few clients or identical units to many. Those two want different infrastructure, and I have been building for one while occasionally pricing for the other.
The most useful research is the kind that recommends what you already rejected. It tells you whether you rejected it for the right reason.
