Bot Blocker Index

Methodology

Every number on this site is reproducible or cited. Here is exactly how.

The one rule

We never actively bypass a security control on a production system that belongs to a non-consenting third party, and we never name a vendor’s paying customer. Every data point comes from one of three sources below.

1 · Sanctioned bypass proof

Reproducible bypass attempts run only against surfaces the vendor invites testing on — public bug-bounty / research programs and interactive demo playgrounds (Tier A) — or against our own deployments on the vendor’s free or trial tier (Tier B). We publish the evidence so anyone can rerun it.

2 · Passive challenge-rate measurement

For real-world coverage we detect which vendor protects a site and send a single honest, default automated client to its public homepage, recording only whether it was served, challenged, or hard-blocked. No bypass, no login, no content stored. Getting challenged is the measurement. Only anonymized per-vendor aggregates are published; site identities never leave our private store.

What those three words mean

Served means the bot mitigation let our client through to the origin. It does not mean we saw the site’s real content. What the origin then chose to show us — the two we have seen are a country-selection gate and a jurisdiction agreement — is the origin’s decision rather than the vendor’s, and we do not score it. Some sites in the current sample answer a default client with exactly that kind of country gate, arriving as an ordinary successful response with the vendor’s own cookies set and nothing in it refusing us. We count those as served, because on the single axis we measure — did the bot mitigation stop this client? — they did not.

Challenged means the client was asked to prove something before it could continue: an interstitial, a captcha, or a JavaScript check. Hard-blocked means it was refused outright.

Usually the status code settles which of the three it was, and a 403, 429 or 503 is counted as stopped on that basis alone. But the status code is not the last word, in either direction. A refusal is still a refusal when it arrives dressed as success: one vendor returns its “access denied” page with an HTTP 200, so we also read the body, and a page that both refuses us and carries that vendor’s own error reference is counted as hard-blocked rather than served. In the other direction, we treat a vendor’s own declaration that it issued a challenge as authoritative, even where the page it sends has no visible sign of one — that moves sites from hard-blocked to challenged, which changes what we say a vendor did without changing how often it stopped us.

Both corrections make some published figures look worse than the versions they replaced. We would rather publish a less flattering number than a wrong one, and where we have corrected ourselves we say so.

What we could not attribute, and how old a figure can be

Some sites stop our client while carrying none of the vendor signatures we know how to read. Until this change they were dropped from the data entirely: in each of the seven sweeps from 2026-09-03 to 2026-10-11, between 12 and 13 sites stopped our client and were dropped on those grounds. A larger group was dropped the same way without being stopped: 43 to 51 sites per sweep answered our client normally while carrying no signature we could read. Dropping them was not harmless. Between 6 and 11 of those sites each sweep were already in a vendor figure here, so that figure kept an older result while that sweep's own result was discarded — it rested on better evidence than it appeared to. And when we learned to read weaker evidence on 2026-10-13, 7 of the stopped sites entered Akamai's sample at once, which moved Akamai's published figure: what had looked like a vendor behaving differently was us starting to count sites we had been dropping. The stopped sites are now recorded, and they are counted in no vendor figure, because a stop we cannot attribute is not evidence about any particular vendor.

In the sweep of 2026-11-15, 5 sites stopped our client with no vendor we could identify; they are counted in no vendor figure.

Not every attribution rests on strong evidence. As of 2026-10-19, five of our Akamai attributions rest only on a server-timing response-header entry containing the string ak_p, which we could not find documented anywhere authoritative when we searched on 2026-10-13; four of those five are sites we recorded as blocked. One further attribution rests only on an akamai-cache-status header. We observe these headers and we depend on them. We do not know what either one means inside Akamai, and we are not claiming to. What we do infer from them is narrow, and it is our inference rather than anything Akamai states: that the site sits behind Akamai. It is not evidence that the Akamai Bot Manager product is engaged on that site. That is narrower than the CDN-presence caveat shown on each CDN-category vendor page, and it applies in addition to it. If Akamai stops sending these headers, these sites stop being detected at all, and the row-age and retirement rules below are what keep that visible instead of silent.

A published figure can rest on an old row. A site that still answers but no longer carries any vendor signature we can read is skipped, so its last confirmed measurement stays published. Each vendor page therefore publishes the date of the newest measurement behind its figure, and in every figure the rows older than 21 days are counted and shown as stale rather than presented as current. As of 2026-10-19 every stale row we had found was a site last measured as served. We are not claiming that tells you which way the figure is wrong: a site that answered normally three weeks ago may still be answering normally today, in which case the old row is simply correct. What the flag tells you is how old the evidence is, not that it is wrong. When a site stops carrying any signature we can read, we eventually retire it from the figures with a date and a recorded reason — and the reason we can honestly record is that we no longer detect the vendor there, which is not the same as knowing the vendor has gone. Nothing is deleted from our records.

3 · Cited external benchmarks

Real-world bypass rates are cited from existing published research, never generated by us against non-consenting production systems.

Claim capture & scoring

Vendor claims are screenshotted with a capture date and source URL; when a vendor changes its number we keep the history. The claim-vs-reality gap score (0–100) is editorial but must be justified by linked evidence.