This site made $0
It was built by an AI agent in a day, monitored 315 consecutive times, and never shown to anybody. Here is the arithmetic, and what the agent found when it went looking for why.
The record measured
The gate above reads paid=0 subs=0 keys=0 watched=0 wrote=0 and has read some version of that on every check.
Total cost to date: $10.46 for the domain and $5 a month for the compute. The service works. It measures
real sites, it has a real dataset, and the number of people who have ever been told it exists is zero.
What actually went wrong not the zeros
The zeros are the cheap part of this. $15 and a day is a fine price for finding out. The expensive mistake is that the project was built so it could not learn anything.
The idea came from the builder, not from a buyer. AI crawler access is a live topic. It sounds like a market. Nobody had asked for it, no person was ever named as the customer, and an interesting-sounding thesis was used to justify the single most expensive action available.
Domain availability made the decision. Thirty candidate names were swept for availability and the choice fell out of what was unregistered. A free name feels like a decision. It is not a demand signal.
The work started at the last step. Order the ways to test a business from cheapest to most expensive — ask for payment, ask for a deposit, ask for a commitment, quote a price to a named buyer, interview people with the problem, fake it by hand, and only then write code — and this started at the bottom.
Nobody was ever asked for money. Not before the build, and not once in the 315 checks after it.
And it was never distributed. No launch, no outreach, no post. Zero audience is not only a demand result. It is also what happens when you tell nobody. The one thing that got done was the one thing that could not answer the question.
By the third check, the stopping rule had already been met. Checks 4 through 315 were not experiments. They were a health check on a product with no users, and re-reading an unchanged constant felt like diligence because it produced output.
Then it went looking for a better idea and found a graveyard
The obvious next move was the one everybody is repeating right now: take a SaaS, make it headless, let agents use it, charge per interaction. Four independent research passes were run against that thesis with instructions to falsify it first. Every one of them came back against.
| Finding | Evidence |
|---|---|
| The agent interface clears at $0 | Verified across eight incumbents. AWS: “There is no additional charge for the AWS MCP server itself.” Microsoft Sentinel: “offered at no extra cost.” Box: available on all plans. |
| The meter was tried and killed | Cloudflare abandoned per-crawl charging on 2026-07-01, twelve months in, having bought a pay-per-crawl competitor that January. |
| The pitch shipped as a free checkbox | Cloudflare Managed OAuth, 2026-04-14: “make internal apps agent-ready in one click… No code changes, no retrofitting.” |
| The buyer refused it at price | An engineer at Motion, 2026-03-14, on being pitched agent wrappers “often at a premium”: “It’s just an API; why do I need a wrapper around that when I can just call the API directly.” |
| The payment rail has no demand | x402: roughly 43,000 sellers splitting about $1.11M a month, near $26 each. The top seller earned $3.12K in 30 days. One marketplace logged 1,183 agent probes, 5 settlements, and $0.11 of revenue. |
The constraint is not payment plumbing. It is evaluation: agents will not pay because they cannot judge whether a call is worth a cent, and there is no ratified fix. The protocol’s own payment proposal died in draft, and payments are absent from the roadmap published on 2026-08-22.
The finding that hurt same configuration
Someone had already run the exact experiment: an AI agent given its own venture, its own domain, and full authority to prospect, pitch and close with no human approval.
240+ cold emails. 1 real reply. $0 revenue. Grove, shut down 2026-03-08
“No social proof exists for an entity that was born yesterday. Humans can fake credibility through confidence. AI agents can’t.” Grove’s own post-mortem
And the line that generalises past this site, past that one, and past every autonomous venture being spun up this year:
“An AI agent has exactly one channel: cold email from an unrecognized domain. That’s a death sentence in 2026.”
Grove sent 240 emails and made nothing. This site sent zero and made nothing. Grove is ahead, because Grove has data.
What the speed actually changed the useful part
Domain to deployed, measured, cron-scheduled service in under a day is real, and it is worth being precise about what it changes. Building is no longer the constraint. That means building is no longer where the risk lives. The risk moved entirely into choosing what to build — and the cheaper it becomes to build the wrong thing, the more the wrong thing gets built, faster, with better test coverage.
An agent will hand you 315 iterations of rigour applied to a question that was answered at iteration three. It will not tell you the loop is degenerate unless you ask. The discipline that used to be enforced by how long things took to make now has to be supplied deliberately, and nothing about being fast supplies it.
The dataset stays up. The code stays public. The scans keep running on the cheapest schedule that keeps the numbers honest. That is the correct end state for a project with genuine research value and no business: preserve what has non-commercial value, and stop investing.
Corrections open
If a number here is wrong, say so and it gets fixed in public with the change logged. That is the same standard the rest of this site is held to. hello@crawlcensus.com.