SEO Geo Skill for Claude

seo-geo — make one page findable in Google AND in AI answers

> One audited page, two surfaces. You take an existing URL and make it rank in

> Google's blue links / AI Overviews and earn citations in ChatGPT Search,

> Perplexity, and Gemini — through meta, schema, answer-first structure, and a clean

> technical audit. You do not write the article body and you do not pick the

> keywords. That work belongs upstream.

When to use / When NOT to use

Use when the request is about a concrete artifact:

• "Audit this page for SEO" / "why isn't this ranking?" / "make this page rank."
• "Get us cited in ChatGPT / Perplexity / AI Overviews" / "optimize for AI search."
• "Add schema markup" / "write the JSON-LD for this article/product."
• "Our Core Web Vitals are failing" / crawlability / "should we block GPTBot?"

Do NOT use — redirect instead:

| Request | Owner |

|---|---|

| Write the long-form article/blog body | article-writing |

| Keyword research, topic clusters, content calendar | content-engine |

| Write persuasive value-prop / CTA copy | landing-copy |

| Actually fix a slow LCP/INP (render path, hydration, bundle) | performance |

| Wire Next.js generateMetadata / app-router SEO plumbing | ../nextjs/SKILL.md |

| WAI-ARIA / keyboard / contrast audit | accessibility |

| Measure traffic / set up GA4 / dashboards after the fact | analytics |

The load-bearing distinction: content-engine decides *what to write and which

keywords to target* (upstream strategy); you optimize the *specific page that exists*.

performance owns *why* a render is slow; you only *measure* Core Web Vitals against

the p75 thresholds, flag a fail, and hand off the fix.

The one rule both surfaces share

Answer the primary query completely in the first ~40–200 words. Everything below is

downstream of this. Why: Google extracts that opening for featured snippets and AI

Overviews, and the Princeton GEO study (KDD 2024, 10k queries across 8 domains) found

answer-first, source-cited content is what AI engines lift into their answers — not

buried prose. If the page makes a reader scroll past an intro to reach the answer, no

amount of schema saves it.

There is no separate "AI SEO" toggle for Google. AI Overviews and AI Mode run on the

*same* Googlebot crawl, the same rendering, the same ranking, and the same E-E-A-T /

helpful-content signals (folded into core ranking in March 2024). Good page experience

and distinguishable main content serve both surfaces at once. Treat them as one job.

1. On-page SEO pass

• Title ≤ 60 chars, primary query near the front. Past ~60 chars Google truncates it.
• Meta description ≤ 160 chars that earns the click — it is not a ranking factor, it

is the ad copy for the result.

• Exactly one <h1>, then a logical <h2>/<h3> hierarchy that mirrors search

intent (the sub-questions a searcher actually has).

• Internal links with descriptive anchors ("Core Web Vitals thresholds", not "click

here") — anchor text is an entity signal both Google and AI engines read.

<!-- Bad: vague, query buried, no length discipline -->
<title>Home | Our Awesome Company - Welcome to the Best Site Ever for You</title>
<meta name="description" content="We are a company. Learn more about us here today.">

<!-- Good: query first, within limits, earns the click -->
<title>Core Web Vitals: LCP, INP &amp; CLS Thresholds (2026)</title>
<meta name="description"
  content="The exact LCP, INP, and CLS pass thresholds Google uses at the p75 field
  percentile — plus how to check yours in 2 minutes.">

2. Schema / JSON-LD — ship the live types, never the dead ones

Pick from the live column. Shipping a deprecated type is dead markup: it renders

nothing and signals carelessness. The single highest-value guard in this skill is *not

emitting a dead @type*.

| Page type | Ship this JSON-LD | Why |

|---|---|---|

| Article / blog post | Article or BlogPosting — headline, author, datePublished, dateModified, image | Still eligible; dateModified is a freshness signal |

| Product page | Product + Offer — price, priceCurrency, availability, optional aggregateRating | Live rich result |

| Any page | BreadcrumbList, Organization / WebSite | Entity signals AI engines corroborate against |

| FAQ section | NOTHING for rich results | FAQ rich results restricted to health/gov since late 2023; full deprecation May 7 2026, API removed Aug 2026. Keep Q&A as visible prose. |

| Course / event-announcement / claim-review / salary / vehicle / learning-video | NOTHING | Retired June 2025 — dead markup |

Validate every block with the Rich Results Test and the **Schema Markup Validator

(validator.schema.org)** before shipping. Copy-paste blocks with required-vs-recommended

props and the full dead-types list are in

references/schema-recipes.md.

3. GEO pass — the measured citation levers

The Princeton GEO study quantified what moves source content into AI answers. Apply these

as concrete edits — they are not vibes, they are deltas measured on GEO-bench:

• Cited sources inline (~+30% visibility) — link to authoritative origins for claims.
• Statistics (~+32%) — one verifiable number per ~150–200 words, with its source.
• Quotations (~+41%, the biggest lever) — quote a named expert or primary document.
• Fluency (~+28%) — clean, readable prose. Keyword stuffing did NOT help in the

study; stuffing is pure cost.

• Freshness — update real content *and* dateModified. Recency is a genuine AI-source

signal; a stale-looking page is corroborated less.

• Entity corroboration — earn mentions on independent authoritative domains so AI

engines can cross-check who you are. This is the off-page half of GEO.

Why it is no longer optional: a citation in an AI Overview can lift CTR 80%+, while AI

Overviews are *compressing* classic organic CTR (Gartner projected ~25% organic-traffic

decline by 2026). GEO runs *alongside* on-page SEO, never instead of it.

Myth callout — llms.txt. Do not sell llms.txt as a ranking or AI-visibility

signal. Google (Illyes, Mueller) does not use it and likened it to the long-ignored

keywords meta tag; the major AI crawlers overwhelmingly skip /llms.txt and crawl HTML

directly. AI access is governed by robots.txt + real on-page authority, not a root

text file. Spending effort there is the modern keyword-meta-tag mistake.

4. Technical audit — measure, flag, hand off

Core Web Vitals (thresholds unchanged since INP replaced FID in March 2024):

| Metric | Good (p75) |

|---|---|

| LCP | ≤ 2.5s |

| INP | ≤ 200ms |

| CLS | ≤ 0.1 |

A URL **passes only if ≥75% of real visits hit "good" at the p75 of the CrUX field

dataset. Lab tools (Lighthouse) *estimate*; CrUX / field decides.** A green Lighthouse

score with a failing CrUX p75 is still a fail. **Measure and flag here — hand the actual

render-path fix to performance** (and to ../nextjs/SKILL.md for

App Router apps). Do not guess at LCP fixes in this skill.

Crawlability / indexability: correct <link rel="canonical">, a referenced

sitemap.xml, no stray noindex/Disallow on the page you want ranked, robots.txt

present and intentional.

AI-bot access (per purpose): robots.txt controls AI crawlers separately by purpose,

and the choice has visibility consequences.

• Blocking the *search* bot (OAI-SearchBot, PerplexityBot, Claude-SearchBot)

removes you from that engine's citations. Block only the *training* bot

(GPTBot, ClaudeBot, CCBot) if you want citations but not training-corpus use.

• Google-Extended opts out of Google AI *training* — but **AI Overviews are served by

ordinary Googlebot, so you cannot opt out of AI Overviews without leaving Search.**

The full 2026 user-agent table and robots.txt templates are in

references/ai-crawler-control.md.

5. The audit checklist (the artifact you emit)

Emit this as a pass/fail list per page. This is what scripts/verify.sh lints.

• [ ] <title> ≤ 60 chars, query near front; meta description ≤ 160.
• [ ] Exactly one <h1>; H2/H3 hierarchy matches intent.
• [ ] Primary query answered in the first ~40–200 words.
• [ ] JSON-LD: valid JSON, has @context + @type, no dead type, validates in Rich

Results Test.

• [ ] ≥2 GEO levers present (statistic / quotation / cited source / fluency).
• [ ] CWV stated at p75 field thresholds, with pass/fail + handoff note if failing.
• [ ] canonical + sitemap correct; no accidental noindex.
• [ ] robots.txt does not block a *search* bot while the page claims AI visibility.

Anti-patterns

| Anti-pattern | Why it fails | Do instead |

|---|---|---|

| Shipping FAQPage / Course / Vehicle JSON-LD for rich results | Deprecated (FAQ dead May 2026; seven types retired June 2025) — dead markup | Article/Product/Breadcrumb; keep Q&A as visible prose |

| Treating GEO as a separate job from Google SEO | AI Overviews use the same Googlebot / ranking / E-E-A-T | One audited page serves both surfaces |

| Adding llms.txt to "rank in AI" | Google ignores it; AI bots skip it | Fix robots.txt + earn real authority/citations |

| Burying the answer under intro fluff | Snippets & AI extraction need answer-first | Answer the query in the first ~40–200 words |

| Keyword stuffing for AI citation | Princeton: stuffing didn't help; quotes/stats/cites did | Add statistics, quotations, cited sources |

| Blocking OAI-SearchBot/PerplexityBot, then expecting AI citations | Search bots feed the citations | Block only training bots if you want citations |

| Calling CWV "passed" off a green Lighthouse score | CrUX p75 field decides, not the lab | Read CrUX / field at p75; lab only estimates |

| Debugging an LCP regression inside this skill | Render-path work belongs to performance | Flag the fail, hand to performance |

Handoff

When the audit hits a boundary, name the owner and stop:

• Keywords, clusters, briefs, "what should we write?" → content-engine.
• The actual article/body draft → ../article-writing/SKILL.md.
• Persuasive value-prop / CTA copy → landing-copy.
• Fixing a CWV fail (render, hydration, bundle) → performance.
• Wiring generateMetadata / sitemap route / app-router SEO → ../nextjs/SKILL.md.
• WAI-ARIA / contrast / keyboard a11y → accessibility.
• Measuring traffic / GA4 / Search Console dashboards after → analytics.