Tavily's Penny Search: The Clearest Working Case in the x402 Census, Under Scrutiny
Tavily sells web search to AI agents for one cent a call. In the last 30 days, 302 wallets bought it 36,958 times. That works out to about 122 calls per payer, and that ratio is the whole question.
Start with what Tavily is, because it matters. This is not an anonymous contract deployed last month. Tavily is an established AI web-search company with a real domain, Terms of Use, and a Privacy Policy. Its core business is search-as-API for LLM applications. What it added is an x402 endpoint at x402.tavily.com/search, where any agent can buy a single search in USDC on Base with no API key and no account.
The mechanics are clean and verifiable. An agent POSTs to the endpoint and gets back an HTTP 402 response carrying a full JSON schema for the request, so the machine learns both the price and the exact shape of what to send. It pays, retries with payment attached, and receives up to 50 results in roughly 0.7 seconds. One honest caveat: the 402-then-pay flow is at least two HTTP exchanges, not the single round-trip the marketing framing sometimes implies. Tavily offers two payment rails, direct USDC on Base and a second agent-pay rail that uses a signed quote token, whose underlying pattern we have not verified as anything more than a standard gasless-signature scheme.
| Price / search | $0.01 USDC (Base) |
| Paid calls, 30d | 36,958 |
| Payers, 30d | 302 |
| Calls per payer | 122 |
| Response | 0.7s |
| Payment rails | 2 (USDC + agent-pay) |
Now the numbers, from the Crest x402 census as observed at the facilitator level. Roughly 36,958 paid calls over the 30-day window. 302 distinct payers. That works out to about 122 calls per payer.
That last figure is the one worth staring at. A payer averaging 122 calls in a month is not a curious human clicking once to see if the thing works. It is repeat purchasing. Something, presumably an agent in a loop, keeps coming back and keeps paying.
But 122 is a mean, and a mean is the least honest number in statistics. It can describe a healthy base of steady buyers or a handful of heavy scripts propping up a graveyard of one-time testers. Those are very different markets, and they imply very different futures for this endpoint. The settlement data lets us get part of the way to an answer, and the part it hides is exactly the part that matters. Here is what holds up, what does not, and what would change our mind.
The rest of this piece runs the scrutiny: the $370 arithmetic behind the volume, the distribution we cannot see, why an incumbent bolting on a rail changes the thesis, the API-key counterfactual, and what would change our mind.
Continue reading on tenjin · $0.10 →