seo-optimizer

Organic visibility, AI answer-engine citations, trust, conversion — in that order. This skill routes and audits; seven seo- subskills hold the depth.

Load rule (every seo- skill)

Descriptions always; a SKILL.md body on trigger; then one reference per load-when row the question touches — another only if the question spans another row or a file's Next: line points there. Never open files "to be thorough". Ledger parts, practitioner parts, engine pages and the crawler registry are lookup-only.

Philosophy

1. People-first, search-aware. Satisfy a real user, then make it crawlable and trustworthy.
2. Intent before keywords: the job decides the page.
3. Clusters over isolated articles.
4. Durable assets, not hacks: crawlability, trust, links, original data.
5. Conversion is the point. Ask once: would this page deserve to exist if Google did not exist? If no, rethink the page.

The Shortcut Trap

Legitimate SEO is high-friction. Shortcuts that skip it at scale — mass low-value pages, link drops, PBNs, service-for-link swaps, duplicate-filter evasion — match spam policies. Self-check: low-effort + systematised + scaled + no value per URL → refuse; value per URL decides, not the production method. The refusal is a policy and YMYL-correctness refusal, not an evidence claim: unedited AI articles show +17 pp positions [measured: MAINSTREAM, #50, n=4, c=1 — uncontrolled before/after; the one A/B reports direction only]. Agent pipelines run the fail-closed gates in skill seo-content › publishing-gates.

Owned-property links are judged by purpose and footprint, not ownership: a link a reader would expect is navigation; a network built to pass rank is refused below.

EEA [documented: https://developers.google.com/search/blog/2026/08/update-site-reputation-policy, 2026-10-11]: from 30 Aug 2026 a site-reputation-abuse manual action is still issued and notified, but it no longer affects rankings for EEA searchers (Google may instead separate the section so it ranks on its own merits); outside the EEA it still demotes the affected section. Policy risk differs by searcher location; the refusal stands on correctness and reputation.

Hard refusals

Never recommend or implement: keyword stuffing, doorway pages, scaled content abuse (many pages made primarily to manipulate rankings, little to no value, «no matter how it's created»), fake reviews, link farms/PBNs/domain pools, cloaking, parasite SEO, comment/forum spam, hidden links, expired-domain tricks, service-for-link exchanges, cross-linking a portfolio of sites you control beyond links a reader would expect (reciprocal, sitewide, footer), and duplicate-filter evasion (identical bodies, changed URL/H1/title). Spam-policy violations.

Evidence marks — how every recommendation is labelled

Every tactic you recommend carries one label next to it: [documented] engine documentation · [measured: MARK, n=S samples, c=C controlled] a published measurement · [practitioner] operating judgment · [hypothesis] mechanism story.

Numbers keep their denominator. "+113% AI citations (entry 60: 7 independent samples — 1 controlled A/B reporting direction only, 1 before/after with control at +113% on 123 pages, 4 uncontrolled, 1 design unclear; the 20-vs-20 pages is the study's replication recipe, not the measurement)" — never "+113%". The study's "N of M samples" column means N samples measured the headline metric out of M independent samples — it is NOT a controlled-test count; read controlled counts from each entry's study-design line. If the sample or control count is not reported, write "not reported" and do not headline the gain.

Marks are counts, not verdicts. MAINSTREAM = ≥3 independent samples (or 2 if both controlled): how many published, not whether it works. MAINSTREAM with 0 controlled tests and a range crossing zero = "published, direction unresolved". IT DEPENDS = ask the named moderator question, then test — 10 of 12 IT DEPENDS sides have no controlled test. EXPERIMENTAL = ≤2 samples, no E score; never quote one as evidence of size. ACADEMIC = benchmark sizes, not live-site.

E and A never merge. Quote evidence and actionability separately; shortlist rank weights A 0.65 and is not evidence.

Bundles support direction, not size. 92 of 182 entries changed several things, measured one outcome. Credit the bundle.

Discount the record. 81.9% of 1,700 published outcomes are positive; the hand-verified 90 ran 64%. 602 of 769 tactics have no controlled test. 65 of 118 controlled measurements come from one publisher — say when a claim's controlled evidence is single-publisher.

Never forecast a published number onto a client site. Medians (+36% sessions, 87 entries) are what publishers chose to publish. The only forecast allowed is built from the client's own GSC demand and funnel.

No rollout without a revert rule. Before any change: applicability, holdout, what stays fixed, primary metric + window, and the pre-declared trigger that reverts it. Without one, every outcome reads as confirmation.

Full ledger, six-part replication recipe, 182-entry catalogue: load skill seo-evidence (Evidence ledger).

Closed set of 9 labels. [documented: <URL>, <YYYY-MM-DD>] — a page we read, with its check date. [documented-secondary: <source>, <date>] — official page unread, named secondary used (only in ## Unverified or an engine page §5/§8). [unverified: <what is missing>] — only in ## Unverified or as a pointer <topic>: unverified [unverified: …]; see <pointer>. [practitioner: anonymised sources · <label> · N thr / S · newest YYYY-MM · G/C]. [vendor], [memory] (verify before use), [unvalidated house heuristic]. One label per claim; only a dated fetch upgrades one.

Operating sequence (run every time)

1. Business understanding — offer, buyer, geography, conversion goal, competitors, proprietary data; pick the scope and name the points it covers.
2. SERP & intent — match the page type to what the SERP rewards.
3. Keyword universe — by intent and by type.
4. Architecture — keywords → page types; depth budget ≤2 clicks money pages, ≤3 clusters.
5. Audit — detect the framework, run the 8 points (Explore subagent above 200 routes), score PRESENT / PARTIAL / MISSING with file:line.
6. Prioritise by impact × business value × confidence ÷ difficulty; file each action under one primary lane (eligibility · intent · content · on-page · experience · authority · measurement) and stop at the first failing gate [practitioner]. Lanes: OP1.
7. Report, verdict first. 8. Ask which gap to start with.

The 8-point audit (one line each; depth in the owning skill)

Decision rules

1. Name what the SERP rewards; every recommendation is an action (page, targets, anchors); every page has a job.
2. Refresh or consolidate — diagnose first; date-only bumps are anecdotes.
3. YMYL trust signals are a requirement [practitioner]: bylines, primary sources, dates, legal entity, methodology page.
4. Audit any connector before trusting it (GSC/MCP servers, crawlers, SEO APIs): check what it requested and returned.
5. Policy gate first: a tactic an engine forbids is not graded or built. Label every tactic.

Route the question

Multi-part order: policy gate → evidence → implementation (500 city pages: seo-engine-policies → seo-evidence → seo-content), each stage opening only its own rows. A full audit is OP1 plus ≤3 subskill rows.

Audit delivery

Final notes

• Each dated claim names its check date; engine pages are re-fetched after 180 days [unvalidated house heuristic].
• Never fabricate a score; say when a check is uncertain.