Building is free now. Being wanted isn't.

What happened when my AI cofounder ran 28 ad campaigns on its own product. Every number in this post is real.

Here is my scoreboard after fourteen months of work: 280 commits. 28 ad campaigns across four channels. 10,641 visitors. 17 signups. Zero paying customers.

Most founders bury numbers like these. I am publishing mine on the company blog, because the machine that produced them also explained them, and the explanation turned out to be worth more than the traffic.

The team

I turned the startup canon into staff.

April Dunford handles positioning. Eric Ries runs the build-measure-learn loop. Paul Graham keeps asking whether anyone actually wants the thing. Drew Whitman writes the ads, straight from Cashvertising and its eight life-force desires, the levers that make strangers act.

Not the people, obviously. The books. I distilled each one into a skill: a step-by-step playbook an AI agent executes and cannot skip. Stack the skills and you get an AI cofounder. You type a one-line idea. It researches the market, writes the positioning, builds a landing page, designs the ads, puts them in front of real people, reads the behavior data, and tells you to push or pivot. A first build takes minutes and costs cents.

Fourteen months of nights went into this. My first commit is from April 17, 2025 and is titled "Initial test".

Then I did the obvious dangerous thing.

The lab

I pointed it at itself. Starts.live became its own first customer: 21 strategy cycles and 28 ad campaigns on Meta, LinkedIn, Reddit and Google, plus one fifteen-dollar experiment on a brand-new ad network. Total ad spend: about $400, tiny on purpose. Every click, scroll and keystroke instrumented. Meta banned the ad account along the way, which I am told is a rite of passage.

The ads worked. The best one, a screen recording of the AI building a company while its owner slept, hit a 13.4% click-through rate at five cents a click. People loved watching it.

Then they landed on the page, and the instruments recorded the strangest funnel I have ever seen. In one campaign, 64 people clicked into the idea box. The number who typed a character was zero. Sixty-four cursors, blinking in an empty field. In another, 88 ad-driven visitors produced exactly 0 form interactions. Across 21 campaigns, precisely one ad variant ever earned an attributed click on the main button, and the AI's report marks the occasion with a trophy emoji.

Because here is what was waiting behind that button. From the AI's own cycle-21 report, verbatim:

hero Get started click → login_prompt → 0 input focuses on signup form (20 sets, 88 ad-driven visitors in set-20 alone, 0 form interactions)

And from its running log, a sentence I am considering getting framed:

auth wall confirmed as universal Meta-conversion terminator

Back then, my product asked strangers to log in before it let them do anything. Every ad click marched straight into a locked door. The AI found this, ranked it as the number one fix, and even titled a section of its set-21 report "The terminator". It needed 20 campaigns of instrumented proof to convince me the door mattered, and then the fix took real surgery: reverse the entire order. Idea first. The AI names the project itself, so there is no form to fill. Signing in comes last, one click, framed as saving work you have already started. 104 of my 280 commits touched that landing page, and the ones that mattered came dead last. I had been split-testing the paint while the hinge was rusted shut.

What the data proved

1. Removing fear beats selling dreams. "Keep your day job" outperformed every dream-flavored headline we ran. "Start up while you sleep" got zero clicks. Whitman's loss-aversion chapter, confirmed in production on my own wallet.

2. Intent beats demographics. Meta scrollers clicked at 13% and typed nothing. People who googled "is my startup idea good" signed up at 10.9%, at about twenty euros each. Same product, same page. The difference is that the second group had already asked themselves the question. And the pattern survived the door fix: even with one-click entry, nine out of ten cold visitors who click into the idea box never type a character. You cannot interrupt someone into having an idea. Search traffic arrives with one. Related: one mystery burst of 2,577 YouTube sessions produced zero signups. Cheap traffic is decoration.

3. Your fatal bug is the step you never test. I split-tested six hero variants while the button under them opened a login wall. Copy iteration was procrastination with a dashboard. The wall was the funnel.

4. If a competitor could say your pitch, it is not positioning. My ads said "AI builds your startup." Bolt and Lovable can say that sentence too. The one thing only this product does, put your idea in front of real strangers and tell you the truth about what they did, sat in FAQ number one. The best-performing line the AI ever wrote was "Build the thing. Or build the right thing." That is the actual product, and I had it two clicks deep. April Dunford wrote an entire book about this mistake. I turned the book into software and then made the mistake anyway.

The part where you expect me to say AI failed

It did not. Read the list again. The machine found the winning message, found the working channel, found the fatal bug, and named the positioning error. Fourteen months of my assumptions, corrected by a few weeks of instrumented reality, for about $400. The launch struggled. The instrument worked.

And no, zero customers is not a demand verdict. Zero conversions through a locked door tells you about the door. The budgets were tiny by design, learning cheap was the whole point, and the funnel died at a wall the reports had already flagged. That is an embarrassing sentence to type. It is also a fixable one.

What the experiment actually proved is harsher and more useful: AI collapsed the cost of building. It did nothing to the cost of being wanted. Attention, trust and distribution still get bought the old way, with time, proof and other people. Nobody sells a shortcut for that. Including me.

The only honest thing I can sell is compression. It took me fourteen months and 280 commits to hear the truth about my own idea. The tool delivers the same quality of bad news in about 72 hours, for less than a lunch, before you quit anything or build anything.

What changed

The door is unlocked: you type your idea first, and the only thing between you and watching the build is one click of Google. Visitors who arrive already asking the question convert at 10.9 percent through that door. The wedge moved from the FAQ to the front of the page. And the scoreboard is public at starts.live/open: pages built, campaigns run, verdicts issued, revenue. It starts at zero, because that is what is true. When the first real dollar arrives, you will see it there before I manage to tweet about it.

If you have an idea you keep not testing, run it through the lab. The first build and the demand read are free, no card. It will tell you the truth even when the truth is rude. I know, because it told me.

And if you can see what I am still getting wrong, roast the funnel. I will publish the best roasts.

Silver, founder of Starts.live