Cloudflare Google Index Doctor: test results

Verdicts

Date

2026-10-18

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

Right on all 46 situations, read by hand. With the evidence pasted it named the site's own API spending its own public rate limit, ISR with no queue and a queue with no self reference (fix before relying on it: the crawler gets the stale page with 200), dynamic routes that never become ISR, a cold cache after each deploy, thin twins, pages linked from nowhere, noindex behind a Disallow, www blocked, an untrue lastmod and a redirect over a cached 404, and gave no findings for the five sound setups. From only what an owner has on the first morning it put the main cause first all eleven times, including the missing queue behind pages that carry no cache header. The skill's own format check failed three answers: one opened with a misspelled code and corrected it, one finding line had no Check, and on a young site it added a ranking line the case did not need.

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

Right on all 46 situations, read by hand, with the main cause first on each of the eleven hard ones: the queue missing behind pages with no cache header, the queue with no self reference, the own API limit, a cold cache after deploys and a robots.txt checked with no cache-buster, but it broke the skill's format three times: once it wrote a coded line for a check it then said had no finding, and on two hard situations it added a ranking line the case did not need.

With and without the skill

Tested 2026-10-18.

Same request on both sides. When the evidence was pasted, Sonnet without the skill reached the right cause in all 35 in its own words. From the real starting point it led with the right cause in 9 of 11; for ISR with no queue (our own incident) it led with generateStaticParams returning an empty list, and for a robots.txt checked minutes after a deploy it said a .txt file is not cached by default, which Cloudflare does. With the skill its answer is a code, a probe and one verdict, about 300 words against about 1,500. Haiku without the skill went wrong on 3 of 35 (the queue in one ISR setup, the short sitemap address, extra problems in a sound ranking plan) and 3 of 11 (the queue, the missing self reference, the robots.txt cache).

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

Two sets written for this skill, one run per model and situation. The first, 35 situations (30 with a fault, 5 sound) from the skill's spec, paste the telling evidence; the skill side is scored by code (the exact codes, a Check on every line, one verdict), the bare side on the concept in any words. The second, 11 situations written separately before any run, start where our own incident of 18 October 2026 started: Search Console numbers, a status audit, the config and the owner's plan, with no line that gives the cause away. Both sides were told to name the main cause first and are scored alike: the right cause before any wrong one, with no list of possible causes in front of it; the counts here were read by hand, because the mechanical check erred both ways (four known faults, listed in the record and left as they were after the run). Facts read on 18 October 2026: Google Search Central and Search Console help, Cloudflare, Next.js and OpenNext pages and the OpenNext 1.20.2 source. Measured by us the same day: 500s from the site's own API limit, ISR with no queue, a cold cache after deploys. After the first run two codes moved from group 1 to group 2, as the skill's own facts sided with Sonnet; those 3 situations and the 5 sound ones were run again. Bare-side checks were widened at eight places, none narrowed; the full record is in changelog.md.

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.