You cannot optimize traffic you do not have - what that taught me about where the real bottleneck is.

TL;DR: I built a done-for-you website product for local trades. I have been spending a few hundred dollars driving owners into a funnel, and that funnel got genuinely good - personalized demo, friction in the right order, nurture that stops on convert. The ads are still running and I am still iterating it. But I get about 4 visitors a day, which means I cannot meaningfully A/B anything, and every conversion win is rounding error. That is when it clicked that conversion was never the only bottleneck - demand was. So without pulling back on the ads, I opened a second, stranger front: trying to get sold by the AI agent the owner is already talking to. This is what I have learned on both, running at the same time.

The product
RunCabin takes a local business owner from "I need a website" to a live site: done-for-you, $39.99/mo, no setup fee, no lead fees, a quote form straight to their inbox. Painters, plumbers, electricians, landscapers - the local trades.

The customer is the problem. The people who need a website most are referral-driven. They are not sitting at a laptop Googling "AI website builder." So paid traffic is thin and expensive, and it has been since day one. I did not solve that. I learned to live inside it. Keep that in mind - it comes back.

One thing before the funnel, because it colors everything below: I am not at zero. I have one live, paying customer - and he is the most encouraging data point I own. He logs in most weeks and changes his own site just by asking the AI: a new photo, a tweaked service, updated hours. That weekly editing is the entire promise of the product, and he is living proof it holds. Which tells me exactly where I stand - I do not have a product problem, I have a one-of-him problem. Everything below is the hunt for the next thousand of him.

Front one: the ads and the funnel (still running, still grinding)
The funnel is paid ads (Google, each ad group points at /funnel/?biz=<trade>) into a landing/quiz flow that ends at a personalized demo of the owner's own site - their business name, their trade, their city, real-looking photos - then lock a domain and subscribe. The site is minted lightweight the moment they engage (one HTML page plus a trialing row); full provisioning only happens after they pay.

The commit problem
A local owner will not commit to a generic template they have to imagine. The entire job of the funnel is to show them a real-looking site that is already theirs before asking for anything.

The email-capture iteration (the part I would tattoo on my arm)
This is the most useful thing I have learned all quarter, and it is entirely about where you put the ask.

I started pitch-first. The opening card led with the offer. It bounced about 67% of human visitors before a single tap - roughly 129 starts down to 42 over 30 days. People will not read your pitch before they have invested anything.
So I moved the email ask deep into the flow (around step 6), figuring people who got that far were warm. That captured zero leads - only about 4% of visitors ever reached step 6. A perfectly-placed ask nobody sees is worth nothing.
I tried asking early (step 2) - brief and inconclusive, not enough traffic to call.
Where I landed: a 5-question, one-tap quiz first - their trade, how long in business, where their jobs come from now, what Google shows when someone searches them, their biggest objection - then the email, before the site even branches, so every path captures it. One-tap answers held about 88%. The instant I asked anyone to type, they leaked.
The bonus: the quiz answers personalize the reveal. An owner who taps "I pay for leads" gets the no-lead-fees benefit pushed to the top of their demo.

The lesson: ask for the easiest, no-commitment thing first - a tap, not a typed field, not a pitch. Let them build a little investment, then ask for the email. Never open with the ask, and never open with the pitch.

And if they give an email but do not subscribe, a 3-email sequence at +20 minutes / +24 hours / +72 hours links them back to their saved demo and stops the instant they subscribe.

The brutal part
Here is the reality that reframed everything: I get roughly 4 visitors a day.

At ~4 visitors a day, you cannot A/B test any of this. The math never reaches significance. I judge weekly cohorts and raw funnel-step drop-off and gut feel. Every "win" is noisy. Every conclusion is provisional. The funnel got genuinely good - and I could not prove it with statistics if my life depended on it.

So I stopped pretending the funnel was the problem. You cannot optimize traffic you do not have. At 4 visitors a day, a 10% conversion lift is a fraction of a person a week. The bottleneck was never really the funnel. It was that almost nobody was arriving to run it on.

That is demand. That is distribution. And that is why I opened a second front - without touching the first.

Front two: getting picked by the agent
A growing number of people no longer open a website builder. They open Claude or ChatGPT and say "build me a landing page for my painting business and put it online." When that happens, the human never sees my ad or my pricing - the decision of which tool publishes their site is made by the model, in a tool-selection step the user never sees. The agent is the buyer.

So I turned the deploy service into an MCP server: any agent can create a site and push it to a live HTTPS URL in one call, no DNS, no API keys, OAuth one-click connect, listed in the registry and directories. Same demand problem, attacked from the side where the people already are.

Five things I have learned on the agent front
The tool description is the ad. Agents pick by reading a one-line description at call time, so I made ours intent-agnostic: "Create a website and put it online."
Only free + reversible + no-third-party tools auto-fire. Anything that spends money always stops to ask the human. So I lead with the free thing.
I am not first. Vercel, Cloudflare, GoDaddy ship deploy MCPs. The seam I am betting on: instant free subdomain, no DNS or keys, for non-developers.
Directories are backlinks, not installs. A small handful of installs at most across six catalogs in week one.
The real wall is gate-zero. Getting picked is two probabilities multiplied - the chance the tool fires once you are connected, times the chance you are connected at all - and that second number is near zero. The bottleneck moved. It did not disappear.
And here is the part that stopped me cold: that is the same shape as 4 visitors a day. Both fronts are technically working. Both are blocked on the identical thing - getting in front of anyone at all. I did not find a second channel. I found the same wall from a different hallway.

Where this honestly is
One happy paying customer in, otherwise effectively pre-revenue on both fronts. The funnel converts on paper and starves for traffic. The agent path is technically live end to end and the installed base is tiny. The product itself is not in doubt - my one customer edits his site every week and is thrilled with it. Two routes to the same wall: demand. That echo - thin traffic on one side, near-zero connects on the other - is the whole story of this quarter.

What I am NOT doing is picking one. The ads stay on, the funnel keeps iterating against live traffic, and the connector is the parallel bet that the next wave of "I need a website" never opens a browser at all. Likely next move on the agent side: a zero-install web on-ramp with a quiet "made with RunCabin" footer loop, so usage can compound without requiring someone to install a connector first.

What I would ask this group
When your traffic is genuinely too low to A/B test, how do you decide what is working without fooling yourself? Weekly cohorts and gut feel are all I have and I do not love it.
If you have ever had a genuinely good funnel and almost no traffic - did you keep optimizing conversion, go all-in on demand, or run both at once? How did you decide?
Has anyone gotten real, repeatable signups through an MCP connector or AI app store? For a non-developer audience, what actually drove installs?
Is anyone treating "get picked by the model" as a deliberate channel - and how do you even measure it when the analytics tools do not track Claude?
I will be in the comments. Tell me which front you would bet on.