Skip to content

Universal SEO skill for Claude Code. 26 sub-skills + 19 sub-agents covering technical SEO, E-E-A-T, schema, GEO/AEO, agent readiness (Lighthouse Agentic Browsing, WebMCP, llms.txt), backlinks, local SEO, e-commerce, international SEO, Google APIs, and PDF/Excel reporting. 9 optional extensions for live SEO data.

Merge branch 'phase2/content-geo' into release/v2.3.0

Agrici Danielfaffaed01b5a
diff --cc CHANGELOG.md
index 0552e56,0c61312..a1b0468
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@@ -7,68 -7,42 +7,102 @@@ and this project adheres to [Semantic V
## [Unreleased]
+### Security
+
+## [Unreleased]
+
++### Added
++
++- `content_humanize.py` now strips invisible Unicode watermark characters (zero-width
++ codepoints, directional marks/overrides, tag characters) and normalizes exotic spaces
++ before the AI-phrasing pass, which those codepoints otherwise defeat by breaking `\b`
++ word boundaries. Emoji sequences (ZWJ, variation selectors) are preserved. The
++ `seo-content` skill documents the new triggers and the scope limits (statistical
++ watermarks are untouched; intended for the user's own drafts).
++- Templated-metadata detector (`scripts/metadata_template.py`): flags meta descriptions that
++ restate their own title tag verbatim and then close with a stock call to action, the shape bulk
++ metadata jobs produce site-wide. Deterministic string comparison, no model, `method: heuristic`
++ in its output. Exposes a site-level roll-up (`templated_ratio`, `shared_cta_phrases`, `site_risk`)
++ because duplicated/templated metadata is a site-scale signal, not a per-page one. Wired into
++ `seo-page`, the `seo-content` agent, the quality-gates meta description table, and the
++ `seo-programmatic` uniqueness gate, which measures body copy only and therefore cannot see this.
++
+### Changed
+
+- Google Search guidance refreshed through 2026-09-10 from Google-owned sources:
+ site reputation abuse enforcement now differs for EEA searchers (2026-08-28),
+ region-specific aggregator and supplier units are documented for the EEA, South
+ Africa, and Turkiye (2026-09-08), AI Mode adds travel booking and price tracking
+ (2026-08-27), and the CrUX pass-rate figure moves to the August 2026 dataset.
+- `safe_requests_get` and `safe_requests_head` send browser-like default headers
+ instead of `User-Agent: python-requests/x.y.z`, which managed WAFs answer with
+ 403/406 and SSR frameworks answer with an empty client-side shell, both of
+ which callers were analysing as if they were the real document. The values are
+ `fetch_page.py`'s and now live in `url_safety` as the single source of truth,
+ with `fetch_page.py` importing them back. `Accept-Language` is not sent unless
+ the caller passes one: announcing `en-US` makes a multi-locale site serve its
+ English variant, which corrupts hreflang and international audits. (#200)
+
### Fixed
+- The JSON-LD hook now validates every `application/ld+json` block regardless of
+ attribute order, CSP `nonce`, `id` or `data-*` attributes, tag case, or an
+ unquoted type value; such blocks were previously skipped without validation.
+- The JSON-LD hook no longer reports runtime template expressions (JSX, Vue,
+ Svelte, template literals in component files; PHP and EJS everywhere) as
+ invalid JSON, while a malformed literal in plain HTML is still reported.
+- The JSON-LD hook accepts the schema.org `@context` with a trailing slash, in
+ list form, and in `{"@vocab": ...}` object form.
+- `google_report.py` no longer labels a plain-string finding as "Info" in the
+ executive summary's critical-issues box or in the full-audit category
+ findings; the "Info" prefix/badge now appears only when the finding is a
+ dict that carries an explicit `severity`.
+- `parse_html.py` now detects `rel="canonical"` and `rel="alternate"`
+ (hreflang) `<link>` tags case-insensitively, so `rel="Alternate"` or
+ `REL="Canonical"` are no longer silently dropped. Credit to #269 for the
+ report that prompted this investigation.
+- Every live fetch failed behind a configured HTTP proxy: the pinned resolver
+ refused to resolve the proxy's own address, so nothing left the process. The
+ proxy host `requests` selects for the URL is now exempt from that check, and
+ only that host. (#280)
+- `unlighthouse_run.py` no longer passes `--max-routes` as `--scanner
+ '{"maxRoutes": N}'`, a CLI flag unlighthouse-ci's parser never reads (the
+ crawl silently ran uncapped). Route count and a new per-page timeout are
+ now set via a generated `unlighthouse.config.mjs` passed with
+ `--config-file`, confirmed against unlighthouse's CLI source and docs.
+ `ci-result.json` is parsed as the array the default `jsonSimple` reporter
+ actually writes, with a tolerant fallback for the `jsonExpanded` object
+ shape, instead of assuming a dict. `extensions/unlighthouse/install.sh`
+ no longer aborts on a marketplace/plugin install: it now also checks
+ `${CLAUDE_PLUGIN_ROOT}` and the plugin cache before requiring the manual
+ `~/.claude/skills/seo` layout. Fixes #189.
+- Raised `maxTurns` on all 16 agents `seo-audit` can spawn (`seo-technical`
+ 20→45, `seo-content` 15→45, and thirteen others that were below 30) so a
+ large-site audit doesn't hit its turn budget before finishing. Every one of
+ those agents now writes a partial findings file after its first analysis
+ pass and overwrites it with the complete findings at the end, so a
+ turn-budget stop never throws away completed work; `seo-audit`'s
+ error-handling table documents the same contract for the orchestrator.
+ Fixes #177, #272.
+ - AI crawler claims are now checked against the crawler that actually governs them.
+ `GPTBot` was documented as "ChatGPT web search" in the `seo-geo` crawler table; it is
+ OpenAI's model-training crawler, while `OAI-SearchBot` is what determines ChatGPT
+ Search citability. `Google-Extended` governs Gemini/Vertex training and grounding only
+ and is no longer treated as a Google Search readiness signal (Google Search, AI
+ Overviews, and AI Mode all follow `Googlebot`). The same GPTBot-shaped conflation was
+ found for Claude: `ClaudeBot` (Anthropic's training crawler) was listed as a
+ search-visibility crawler in `seo-geo`'s table, agent, and recommendation line, while
+ `Claude-SearchBot` (the crawler that actually governs Claude search citability) was
+ missing entirely. `seo-technical` already had `ClaudeBot` correctly labelled
+ training-only, so `seo-geo` and the `seo-geo` agent were brought into agreement with
+ it rather than the other way around. Adds `Claude-SearchBot` and `Applebot-Extended`
+ (Apple's training-opt-out token, distinct from `Applebot` search indexing) to both
+ skills, a claim-to-bot mapping table for all four vendors, citations to each vendor's
+ own crawler documentation (OpenAI, Google, Anthropic, Apple), the missing
+ `OAI-SearchBot` row to the `seo-technical` crawler table, and requires training access
+ and search citability to be reported as separate findings. The `anthropic-ai` row is
+ marked unverified: it does not appear on Anthropic's current crawler support article.
-### Added
-
-- `content_humanize.py` now strips invisible Unicode watermark characters (zero-width
- codepoints, directional marks/overrides, tag characters) and normalizes exotic spaces
- before the AI-phrasing pass, which those codepoints otherwise defeat by breaking `\b`
- word boundaries. Emoji sequences (ZWJ, variation selectors) are preserved. The
- `seo-content` skill documents the new triggers and the scope limits (statistical
- watermarks are untouched; intended for the user's own drafts).
-- Templated-metadata detector (`scripts/metadata_template.py`): flags meta descriptions that
- restate their own title tag verbatim and then close with a stock call to action, the shape bulk
- metadata jobs produce site-wide. Deterministic string comparison, no model, `method: heuristic`
- in its output. Exposes a site-level roll-up (`templated_ratio`, `shared_cta_phrases`, `site_risk`)
- because duplicated/templated metadata is a site-scale signal, not a per-page one. Wired into
- `seo-page`, the `seo-content` agent, the quality-gates meta description table, and the
- `seo-programmatic` uniqueness gate, which measures body copy only and therefore cannot see this.
## [2.2.6] - 2026-09-10