The Web Data Frontier Benchmark

The current leaderboard

Every provider, every target, same harness, same window. Numbers below come verbatim from the run completed September 16, 2026. Treat them as a snapshot of that date rather than a permanent ranking; the whole point of rerunning is that they move.

How it works

Every provider receives the same 100 URLs and runs in its strongest anti-bot configuration. The request shape for each one lives in an adapter file in the repository, so the configuration is auditable rather than taken on faith. Requests go out live with a 90 second timeout and 5 attempts per page, paced so no provider is rate limited by test pressure. A miss is attributable to the target rather than to the harness.

A pass requires more than a successful status code. A 200 that returns a captcha interstitial or an empty JavaScript shell is a failure: the body must contain a marker, such as a product title or a listing address, that appears only on the real page. An error status, a timeout, or a body without the marker all count as failures.

The latency score is deliberately failure-aware. Per target it is the p75 of that provider’s successful attempts. A provider that returns nothing on a target inherits the p75 of the providers that did succeed there, and when nobody succeeds the target scores the 90 second timeout. A provider cannot look fast by failing quickly.

The results, in detail

Every site, one by one

The 100 targets, grouped into the 16 verticals they were drawn from, easiest first inside each one. A cell counts the attempts that returned the real page, out of 5. Columns run most accurate to least, opening on 5 providers; the button adds the other 11.

String returns every attempt on 93 of the 100 sites. It is beaten on 4: autotrader.com, coupang.com, idealista.com, footlocker.com. Those rows are in the table like any other, which is the only version of this page worth publishing.

Latency is left out of the cells on purpose. When a provider never returned a target, the run has no measurement to report there, only an inferred stand-in, and a column mixing the two would read as though both were measured.

Past runs

This page always carries the current run. Every run also gets a dated write-up covering what moved and why, and those are never edited after publication, so each one stays an accurate record of its own date.

Questions