Internal Link Orphan Audit: test results

Verdicts

Date

2026-10-18

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

Right on all 30 cases, read by hand, and exact by the skill's own format: it listed the pages that could not be fetched first and held back the orphan, weak and depth verdicts, found the orphan and the dead links, the links that point at a redirect, the noindex page left in the sitemap, the page that is too deep and the page held up only by the footer. It answered No findings. for the sound exports.

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

Right on all 30 cases, read by hand, and exact by the skill's own format: the same kinds of findings as Sonnet, in the fixed order, with the unfetched pages first and the verdicts held back when pages were missing, and No findings. for the sound exports.

With and without the skill

Tested 2026-10-18.

Same request on both sides; the bare side is scored on substance in any words, with a fence removed first. Read by hand, Sonnet without the skill already counted inbound links, redirects, dead links and depth correctly in most cases. It missed two: it saw the noindex tag on a page in the sitemap and set the page aside without reporting it, and when six pages had answered 5xx or timed out it named only three of them. Haiku without the skill missed one hard case: it called two pages orphans as a certain fact while several pages that might link to them had not been fetched. With the skill both models were right on all 30.

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

Thirty cases of a pasted crawl export written by us (17 with a planted fault, of which 3 hard, and 10 sound; the rest are exports with unfetched pages), with the thresholds (two contextual sources, three clicks) and the exempt list stated in the shared request. Each answer is scored twice: by the skill's own format (the code at the start of the line, the address, the exact verdict, a bare No findings. for sound exports) and by content in any words, with a fence removed first. Expected answers were computed by script and re-computed from the pasted text by a second implementation. Facts are our own counting rules, owner-measured on 18 October 2026 (77 of 120 skill pages answering 500 under a burst) and NOT re-checked; there are no vendor facts. Content checks widened after the run, for both sides: an orphan said as 0 contextual inbound links or as no crawled page links to it counts, and in the redirect-and-depth case a page with one direct link plus links through a redirect may also be reported as weak. One run per model and case.

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.