AI Product Index / The state of the agent web
Three things this site measures directly: whether AI agents visit, whether the machine-payable web actually answers, and whether any of it converts into a payment. Every figure below is dated and links to the file it came from.
| reading | requests/30d | agent share | AI crawlers | free scores | paid-endpoint hits |
|---|---|---|---|---|---|
| 2026-07-29 | 1,855 | 3.20% | 60 | 75 | 7 |
| 2026-08-01 | 4,672 | 7.19% | 336 | 108 | 45 |
| 2026-08-02 | 7,345 | 12.36% | 908 | 178 | 49 |
Agent share went from 3.20% on 2026-07-29 to 12.36% on 2026-08-02. The site had no inbound links for most of that period, so this is agents finding a domain nobody pointed them at.
49 requests have hit the paid endpoint and met its HTTP 402 challenge. Zero of them came back with a payment. Every settlement this rail has ever taken was our own test.
This is the finding worth more than the rest of the page, because it is the one nobody could have guessed from the outside. The audience is real and growing; the conversion is not a small number, it is zero. A five-cent price, published in machine-readable form at /api/x402/info, advertised in the 402 itself, payable by the reference client — and no agent in the wild has ever chosen to pay it.
Two readings are consistent with that, and this page cannot tell them apart: almost nothing in the wild carries a funded wallet with authority to spend it, or the free tier already answers the question the paid tier charges for. Both are more interesting than "the price is wrong".
14,661 endpoints sounds like an ecosystem. 27.1% of them belong to ten operators, and the 1,527 distinct hosts are the honest denominator. Prices cluster low — median $0.01, 90th percentile $0.15 — which is what a market of machine-to-machine calls should look like, and it means a five-cent audit sits above the median rather than below it.
| chain | endpoints | share |
|---|---|---|
| base | 14,356 | 97.9% |
| solana | 161 | 1.1% |
| base-sepolia | 118 | 0.8% |
| x-layer | 18 | 0.1% |
| algorand | 5 | 0.0% |
| arbitrum | 1 | 0.0% |
Neither upstream registry checks whether its own entries still answer. The Bazaar keeps a resource for 30 days after its last settlement; the MCP registry lists whatever a publisher declared. So "how much of this is alive" was a question nobody could answer, and the expectation going in was that a lot of it would be dead.
It mostly is not: 97.2% of a rotating sample of the x402 catalog answers, and 91.9% of the MCP catalog. A 402 or a 401 counts as answering here, deliberately — a 402 is the correct reply from a paid endpoint, and scoring it as a failure would mark the entire point of an x402 catalog as dead.
40 sites in this registry are re-graded weekly against the 20-check agent-readability set. The distribution:
| grade | sites | reading |
|---|---|---|
| A | 7 | agent-ready |
| B | 8 | minor gaps |
| C | 8 | partially legible |
| D | 5 | weak |
| E | 6 | barely legible |
| F | 6 | effectively invisible to agents |
Full ranking, with what each site is missing: the leaderboard. What each check means and how to pass it: the checklist.
Client type is inferred from the self-reported user-agent. That is a traffic-shape signal, not an identity claim: a crawler that lies is counted as whatever it claims, and the "script" bucket is honestly named because a generic HTTP client could be an agent or a shell loop. No IP addresses are collected.
Catalog liveness is one probe from one network path, on a rotating sample. It is evidence, not proof — an endpoint is only called unreachable after two consecutive misses, and a recovery forgives the record outright.
Everything here describes one site's vantage point on a young ecosystem. It is a reading, not a census.