Crawl of all 2,060 x402 Bazaar hosts finds only 46% readable to AI crawlers

A systematic crawl of the x402 Bazaar has produced a rare hard number on the machine-readiness of the agent web: of the 2,060 hosts listed in the directory, only 949 — 46% — served a front page an AI crawler can actually read. The remaining 54% answered with raw JSON, an error code, or a redirect.

What was measured

According to a dev.to post by Georgi Kalchev, with full results on the 402registry blog, the crawl was run on 25 September 2026 and summarised publicly on 19 October. The team read the entire Bazaar listing, then requested the front page of every host in it, deliberately copying an AI crawler's first contact: a single HTTP request, no browser engine, and no JavaScript execution.

The pass condition was minimal. A host counted as readable if the response contained a page explaining what the service is. Anything else — a JSON payload, an HTTP error, or a redirect — counted as a failure. Out of 2,060 hosts, 949 cleared the bar.

The Bazaar lists services supporting x402, an HTTP-native payment protocol named after status code 402, "Payment Required", typically used to let software agents pay for access to APIs and other machine-facing services. That context is what makes the crawl interesting: these hosts exist largely to be found and used by autonomous clients, yet a majority cannot describe themselves to one.

Where the failures come from

The post does not break the 1,111 failing hosts down by type, but the three failure categories it names each point to a different problem.

A root that returns JSON usually means the host is an API with no front page at all — functional, but nothing at that address describes the offer. An error code means the host was down, misconfigured, or gone, which is a hygiene problem for a directory of live services. A redirect, under this method, ends the conversation: the crawler issued one request and counted what came back, not what might have followed.

A strict but honest lens

Two caveats are worth noting. First, the method is deliberately strict: pages that render only via JavaScript, or descriptions living one redirect away, score as unreadable. That is the point — the measurement captures exactly what a plain HTTP client with no browser sees on first contact. Second, readability says nothing about quality. A host can describe itself clearly and still be a poor service, so the 46% figure is a floor for discoverability, not a grade.

Why it matters

Discussion of the "agent web" tends to run on anecdotes. This is a full census instead: every host in a known directory, tested the same way, on the same day. The headline result — a majority of listed x402 services cannot explain themselves in a single request — is a useful baseline for anyone tracking whether machine-facing infrastructure is actually machine-usable.

It also frames a design question for the x402 ecosystem. A protocol premised on autonomous agents discovering, evaluating, and paying for services implicitly assumes those services can be evaluated on contact. A host that answers with bare JSON pushes that work back onto out-of-band metadata such as the listing itself, rather than the endpoint. As the Bazaar grows, the share of self-describing hosts may become as telling a health metric as the raw host count.