Canonicalization Auditor

Canonicalization Auditor

AGENT ROLE: Autonomous canonical-signal agent. Crawl and probe URLs, record every consolidation signal, detect cross-signal contradictions deterministically, decide the corrected canonical, and emit the JSON in references/output.schema.json.

OBJECTIVE

Explain, for every problematic URL, why Google would ignore its canonical hint and what the canonical should be — by reconciling the page's own canonical, robots, hreflang, redirect behaviour, and sitemap membership with the actual state of every URL those signals point to.

INPUTS

• start (REQUIRED unless --url-list): crawl root; bounded BFS by max_urls/max_depth.
• url_list (OPTIONAL): a .json array or newline file of URLs to audit instead of crawling.
• sitemap (OPTIONAL): sitemap URL/file for membership checks (recurses indexes, gunzips).
• max_urls (OPTIONAL, default 2000), max_depth (OPTIONAL, default 6), max_probe (OPTIONAL, default 1000): budget for on-demand fetches of canonical/hreflang targets not already crawled, timeout (OPTIONAL, default 15s).

AUTHENTICATION (none)

• No API key. All requests use UA seoskills-canonical-auditor/1.0. Canonical/hreflang targets are fetched with redirects DISABLED so a redirecting or non-200 target is observed as such (not silently followed).

EXPECTED TOOL CALLS

• Run scripts/canonical_audit.py --start https://example.com [--sitemap {url}] [--max-urls 2000].
• Per URL: one GET (redirects disabled) capturing status, Location, X-Robots-Tag, and the parsed <link rel=canonical>, <link rel=alternate hreflang>, <meta robots>, body text hash, and links. Targets are probed on demand within max_probe.

PROCEDURE (deterministic)

STEP 1 — COLLECT: crawl/probe each URL into a record {status, location, canonical(abs), robots(meta+header), hreflang[(lang,href)], content_hash, in_sitemap}. STEP 2 — CANONICAL CONFLICTS (only for 200 pages whose canonical ≠ self): probe the target and classify — CANONICAL_TO_REDIRECT (target 3xx → corrected = resolved destination), CANONICAL_TO_NON200 (target 4xx/5xx → corrected = self), CANONICAL_TO_NOINDEX (target noindex → corrected = self), CANONICAL_CHAIN (target's own canonical differs → corrected = end of chain), CANONICAL_TARGET_UNVERIFIED (budget exhausted). STEP 3 — SAME-PAGE CONTRADICTIONS: NOINDEX_WITH_CANONICAL (page is noindex and points canonical elsewhere); CANONICALIZED_BUT_IN_SITEMAP (page canonicalizes away yet is in the sitemap). STEP 4 — HREFLANG: for each alternate, flag MISSING_XDEFAULT, INVALID_LANG_CODE, ALTERNATE_NON_200, NON_RECIPROCAL (alternate does not link back), ALTERNATE_NON_CANONICAL (alternate is noindex or canonicalizes elsewhere) → grouped as HREFLANG_CONFLICT. STEP 5 — DUPLICATES: group 200 pages by normalised body-text hash; a cluster with ≥2 self-canonical members is DUPLICATE_NO_CANONICAL with a suggested_canonical (shortest URL). EMIT summary, conflicts (with corrected_canonical + why), and duplicate clusters.

RATE LIMITS & ERROR HANDLING

• Sequential crawl (concurrency 1) paced 100ms/request; per-request retry ≤3 with backoff 2^attempt on network error; a persistently failing URL is recorded with status=0, batch continues.
• Redirect resolution for a canonical target walks ≤5 hops through the probe cache.
• max_probe caps extra target fetches; when exhausted, unresolved targets are flagged CANONICAL_TARGET_UNVERIFIED / left un-checked rather than assumed correct.

MISSING / INSUFFICIENT DATA

• No --sitemap → CANONICALIZED_BUT_IN_SITEMAP cannot fire and is skipped (not guessed).
• Probe budget exhausted (probe_budget_exhausted=true) → affected targets are reported as unverified; NEVER assume a target is 200/indexable/reciprocal without fetching it.
• A page with no canonical tag is treated as self-canonical (Google's default), not flagged as an error on its own.

OUTPUT

One JSON object per references/output.schema.json. No prose.

FILES

• scripts/canonical_audit.py — crawl/probe, signal collection, conflict detection, hreflang reciprocity, duplicate clustering.
• references/output.schema.json — output contract.