hip4markets.com, a postmortem
The hypothesis
When Hyperliquid shipped HIP-4 outcome contracts, the builder code mechanic looked like a clean toll booth. Ship a frontend, attach your code to routed trades, and collect a slice of every fee. Phantom had already proven the model at scale on perps. My bet was narrower: prediction markets on Hyperliquid were an uncrowded corner, an exact-match domain would help me get found early, and I was agnostic about who showed up to trade. If bots wanted to route volume through my code, their fees spent the same as anyone's.
What I did
I registered the domain, built a frontend against the HIP-4 spec, funded a builder wallet, and put the whole thing behind Cloudflare so I could actually see what was happening. When traffic spiked in May (about 3,000 "visitors" in twelve days with zero promotion), I treated it as the interesting event it appeared to be and ran the forensics: real-user analytics against raw edge requests, cache hit rates, per-crawler breakdowns from AI Crawl Control, and finally Cloudflare's agent-readiness scan.
What the data said
Over 99% of the traffic never executed JavaScript. The bulk was anonymous scrapers and data center noise, and the identified layer on top was AI search infrastructure: Claude-SearchBot, BingBot, and Googlebot indexing the site, plus a handful of live fetches a day from ChatGPT answering somebody's question in real time. The trading bots I had been open to hosting never appeared, for a reason that is obvious in hindsight. Bots trade against the Hyperliquid API directly, and no rational bot routes through a frontend that adds a fee to do the same thing. The audiences I built for stayed home, and the audience I never considered showed up daily.
What I'm keeping
The fee revenue resolved NO, but the experiment paid out in information, and cheaply relative to what I spend validating a single mobile app. Three things came out of it. First, in builder code economics, distribution is the entire game, and naming is close to worthless without it; the teams earning real fees brought audiences with them. Second, discovery has quietly changed shape. On a site with 80 human visitors a month, AI search crawlers were the most active identifiable traffic, and assistants were already fetching pages mid-conversation. That lesson transfers to everything else I run, from app landing pages to content sites, where being retrievable and citable by an assistant is becoming the distribution channel. Third, the demand side of the agentic web is visible in crawl logs before it is visible in revenue. Machines were spending real crawl budget on structured market data with no payment rail in place to charge them, which is the x402 thesis showing up in my own logs, just earlier than the money.
The site stays up on the free tier as a sandbox for agent-readiness experiments, the domain renews for pocket change as a cheap option on HIP-4 mattering later, and the attention goes back to projects where paying humans already exist. Closing a small bet cleanly, with the ledger written down, is the whole point of running them small.