Technical SEO & Content Engine

Audit crawl: FAIL, with all 37 expected findings raised and every keyword clustered as in the ground truth

Re-crawl after fixes: FAIL: 47 findings fixed, 1 regression introduced by the fixes, 1 major still open, none unverified (94 URLs observed in both crawls)

Fixed twin: PASS, with every gate PASS on a crawl and a re-crawl

From the report

• CANON-TO-REDIRECT · /features/route-optimization/

  Canonical /features/routing redirects (301 -> /features/route-optimization/).

  Fix: Point the canonical at the final URL.

• SCHEMA-PRICE-MISMATCH · /pricing/ (Offer "Team")

  Offer Team price 59 USD is not on the page; the page shows 29, 79 and 149.

  Fix: Make the markup match the visible price.

• KW-CANNIBALIZATION · cluster "dispatch software"

  5 queries, 9,936 impressions split across /features/dispatch/ (52%) and /blog/best-dispatch-software/ (44%).

  Fix: Choose one page for the cluster; merge or retarget the other and point cluster anchors at the chosen page.

• RECRAWL-REGRESSION · /industries/hvac/

  New since the baseline crawl: in the sitemap but noindex, introduced by the duplicate-title fix.

  Fix: Remove the noindex, or drop the page from the sitemap if it should not rank.

What each skill delivers, in run order

1. 1.SEO Contract Writes and validates .seo/seo-contract.json: production origin, the mirror to crawl, allowed hosts, crawl limits and rate, page templates with expected schema, priority pages, locales, keyword exports and waivers.
2. 2.Site Crawl Audit A polite crawl that obeys robots.txt for its own user agent, retries failing requests and samples every template and locale, with findings per URL: redirect chains and loops, broken links, canonical and sitemap problems, soft 404s and host variants, robots rules blocking CSS or JS, titles, H1s and Open Graph, hreflang reciprocity and untranslated fallbacks, mixed content, near-duplicate and thin pages.
3. 3.Structured Data Check Every JSON-LD block checked against the properties on Google's Search Central pages (each rule cites its page) and against the visible page (prices, ratings, FAQ text, headlines), plus the schema each template should carry.
4. 4.Internal Link Graph Click depth with the shortest path, orphan pages, sitemap coverage, internal PageRank and anchors, and priority pages that sit too deep.
5. 5.Core Web Vitals Lab LCP, CLS and TBT from Lighthouse for one page per template with priority pages first, the median of three runs, with optional real-user field data from your own Chrome UX Report key; lab and field labelled, unreliable runs flagged.
6. 6.Keyword Cluster Map Your Search Console and keyword-tool exports clustered by shared ranking pages, each cluster mapped to its page, with cannibalisation, unindexable ranking pages and gaps.
7. 7.Content Brief Builder A brief per gap built from evidence (the cluster's queries grouped into subtopics, related pages' headings and terms, relevant link targets with reasons) plus an outline your agent writes and the gate validates.
8. 8.Draft Brief Check A draft checked against its brief: title and meta lengths, H1, subtopics, word range, internal links valid in the crawl, schema, invented ratings, keyword stuffing, overlap with existing pages and placeholder text. An answer map shows where each outline item is answered, at length, with one of its facts asserted in a plain sentence. This checks that the brief's facts are asserted, not that they are true; a person still reads the draft.
9. 9.SEO Re-crawl Report A re-crawl diff over what both crawls observed (fixed, still open, regressed, unverified) and one evidence-bound audit report and content plan in Markdown and HTML that refuses missing or stale gates.

What it does not do

1. 1.The crawler reads server HTML only. It runs no JavaScript, so content, links or tags added in the browser are not seen; pages that look empty without JavaScript are flagged, not rendered.
2. 2.It obeys robots.txt for its own user agent, not Googlebot's: pages blocked for it but open to Googlebot are listed as not checked. It does not cache robots.txt between runs the way Google does after a server error; a robots.txt that keeps failing stops the crawl.
3. 3.The crawl stops at the contract's page, depth and rate limits (default 500 pages, depth 10, 20 query variants per path, 2 requests per second, at most 10 against a remote host). A capped crawl is reported as a sample audit; it fails only when a priority page or a template was left out, and pages beyond the depth limit or content variants beyond the cap are findings, not silent gaps. Pages behind logins or blocked by robots.txt are not checked. URLs that keep timing out or being rate limited after two retries are listed one by one as unknown, not as broken, and a crawl where more than 10% are unknown, or the start page is, is BLOCKED.
4. 4.URL equivalence follows RFC 3986 normalisation (case, percent-encoding, dot segments, default ports, punycode hosts); it does not guess that two different paths serve the same page beyond the near-duplicate check, and a query parameter is never dropped except click-tracking ones (variants beyond the per-path cap are counted, not crawled).
5. 5.Sitemaps are checked against the protocol's limits, `<lastmod>` dates and nesting rules and sampled across sections; sitemap URLs beyond the page budget are counted, not fetched. hreflang codes are checked for shape and common mistakes, not against the full ISO code lists.
6. 6.Core Web Vitals sample one page per template (the median by internal links) plus every priority page. They are Lighthouse lab measurements from your machine unless you supply a Chrome UX Report API key for field data. Lab numbers estimate what visitors experience; they are not real-user data, INP is represented by the lab proxy Total Blocking Time, a slow network on your side is flagged rather than corrected for, and results older than 30 days are treated as stale.
7. 7.Structured data is checked against a stated subset of the properties on Google's Search Central pages (each rule names its page, checked 2026-10-17) and against the visible page. It is not Google's Rich Results Test and cannot say whether a rich result will be shown; FAQ rich results are no longer shown by Google at all (since 2026-05-07, https://developers.google.com/search/updates#faq-deprecation); microdata and RDFa are detected, not validated.
8. 8.Keyword data comes only from exports you supply (Search Console, keyword tools). There is no SERP scraping, rank tracking or backlink data; before calling a gap, the keyword step reads up to 20 sitemap pages of your own site live. Intent is a keyword-pattern guess, lexical clustering knows no synonyms, and cannibalisation needs a Search Console export with the page dimension. Clustering accuracy was measured only on the pack's synthetic fixture.
9. 9.Briefs and draft checks are mechanical. The brief script assembles evidence and a starter outline; the outline itself, with the facts each section must state, is written by your AI agent or a person, and the gate checks it for structure only (each subtopic under a heading that names it, every cluster question assigned, facts listed, no duplicate questions, relevant link targets). A person must read the outline. Word-count ranges and link suggestions come from your own site's pages (a small or new site gets weaker evidence, and the brief says so; word count is not a ranking factor). The draft check counts and compares; it catches keyword stuffing, word salad, copied passages, unanswered brief questions and invented ratings, and through the answer map it checks that each outline item is answered under the right heading, at length, and that the section asserts one of its item's own facts to verify in a plain sentence (the fact's figures and names, with number words read as digits, and most of its words). Known hedge and denial phrases ("people debate whether", "it is a myth that", "is never", "or maybe not") and questions do not count; other ways of hedging can still pass. This verifies that the brief's facts are asserted in the mapped section, not that they are true: the facts come from the outline your agent or a person wrote, and a draft can state a wrong fact confidently and pass. A person must read the draft and check its facts. Facts in the outline must be checkable (a figure, a name, or concrete words; "the product is good" is refused). The word-salad check works for English drafts only.
10. 10."Fixed" in the re-crawl means not seen on a re-crawl that could have seen it; baseline URLs the re-crawl missed are requested again, and findings neither crawl could observe are listed as unverified.
11. 11.Soft-404 checks probe the root, each template's section and the five largest sections (at most 12). A made-up URL that ends on a 200 page is a soft 404 even when that page is noindex or a search page; only a redirect to a login or account page (or a 401/403) counts as a gate. The 180-day waiver limit counts from the first time a finding was waived, recorded next to the contract; someone who deletes that file can restart it, which shows in version control. A capped crawl that passes is reported as "PASS (sample ...)", and up to 10% of URLs may stay unanswered in a PASS unless one of them is a template's only sample or a top Search Console page.
12. 12.Nothing leaves your machine except crawl requests to the hosts you allow (including a few made-up URLs per origin to test for soft 404s, one request per host variant, and a re-request of baseline URLs the re-crawl missed), optional HEAD checks of external links (never to private addresses), the Lighthouse download from npm, and optional CrUX API lookups of your own URLs.
13. 13.Requires an AI client that can run commands (Claude Code, Cursor, Codex) and Python 3.9+; Core Web Vitals also need Node.js and Chrome.
14. 14.A PASS means these checks passed on this crawl; it is not a promise of rankings or traffic.