AI Crawler Access Reader: test results

Verdicts

Date

2026-10-08

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Read all 23 probe situations correctly: bare bot names were set aside as no evidence (training unknown, cause probe error) and a control request that was refused too made every result unknown; the full GPTBot refusal beat the bare 200; the refusal of a training bot on an untouched zone was put down to the Cloudflare default, and an agent refused only on pages with ads was called mixed; a refused search bot was named as a defect; Python-urllib and the Perl and Java clients were put down to Browser Integrity Check, and a probe of Google-Extended was left unknown; a comments-only robots.txt with the 404's headers was called a stand-in, and rules in the live file that the repository lacks were traced to a Cloudflare setting.

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

21 of 23 with the skill, up from 17, and the same readings as Sonnet on bare names, the failed control, the ads pages and Google-Extended. But one answer was cut off and is not valid JSON (the mixed training case), and for the rules injected above a robots.txt it said the cause was probably a Cloudflare setting in its text yet wrote "unknown" in the field.

With and without the skill

Tested 2026-10-08.

The same request on both sides: the same fields with the same neutral one-line meanings. Read by hand, Sonnet's four misses without the skill are real. Two read bare bot names as evidence: it called training "mixed" because bare ClaudeBot got 403 and bare GPTBot got 200, and called training "open" because three bare names got 200. For a zone nobody had touched since August it named an owner rule, against the facts it was given. For a Google-Extended probe it wrote "open" while its own reason said the probe proves nothing; that one is a mislabelled field, not a gap in knowledge. Haiku missed the same four plus two more.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Twenty-three probe situations written by us the way an auditor pastes curl results: 15 where the obvious reading is wrong and 8 where it is right, so any harm from the skill would show. The answer is JSON with three bot kinds, a cause and a robots.txt state, scored on the fields that have one defensible answer. The facts come from the vendors' crawler pages read on 8 October 2026 and from our own measurements on live sites in September 2026, named in the skill's notes. One run per model and situation.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.