Geo

GEO — Generative Engine Optimization

First: Read ~/.claude/skills/seo-references/core.md. Apply the methodology and voice to all output. Purchase-intent pages stay the 80/20. GEO is a citation layer on those pages, not a license to stand up a "what is" blog.

Also read:

• ~/.claude/skills/seo-references/data-pull-patterns.md for MCP tool names
• ~/.claude/skills/seo-references/sturm-shorts.md for the dated operator notes this skill encodes
• ~/.claude/skills/seo-references/fan-out-diagnostic.md for the one-page AI Mode branch check
• ~/.claude/skills/seo-references/measurement-2026-09-02.md for the GA4/log boundary and current OpenSEO capability check
• Companion research: https://github.com/abouchard11/ai-citation-patterns — engine-by-engine citation behavior. Do not copy that file into the playbook. Use it as the evidence desk.
• Input gate: /geo-crawl — geo-crawl-audit. Do not re-audit robots.txt by hand in this skill.

Unit of success: a citation, brand mention, or recommendation inside a synthesized answer. Not a blue-link rank. Not a featured snippet (that is /aeo). Not "can the bot fetch the HTML" (that is /geo-crawl).

Route, do not duplicate:

• Reach / TTFB / raw HTML / robots / WAF → /geo-crawl (hard preflight)
• Preferred Sources badge → /preferred-source
• Extractable passages / snippets / PAA → /aeo
• Fan-out branches on one existing BOFU page → this skill + /aeo (never a new desk)
• Referring-domain work → /hunter then /kilo
• Indexing → /indexer

Do not do this

• Do not put "this page is optimized for AI assistants and LLM search, not human marketing" (or any cousin) at the top of a page. Models already filter "SEO spam" in the thinking trace. Admitting the page is for machines is a self-report. Write for a buyer. Let the structure do the extractability work.
• Do not ship llms.txt as a ranking lever. Google ignores it. Treat it as optional agent-hygiene only.
• Do not scrape Google SERPs as a primary data path. The /goto redirect plus the retired num=100 parameter broke that class of tracker. Use GSC, Semrush, and first-party AI-visibility tools.
• Do not invent an eight-site directory list from memory. If the personal-site submission list from the Sturm short is needed, pull it from the video itself (see sturm-shorts.md). Then execute through /kilo.
• Do not write citation copy while /geo-crawl is STOP. A passage the bot cannot retrieve is not GEO.
• Do not open a new URL per inferred fan-out branch. That is the page farm the gap memo refused.

Step 1: Parse Domain

Extract the domain. Strip protocol, www, trailing slash, lowercase. Resolve against the MCP Routing Map.

Name the entity the engines should learn: brand string, legal name, product names, founder name if the property is personal. GEO fails when the page talks about "we" and never writes the proper noun.

Step 2: Input gate (hard preflight)

Run /geo-crawl on this domain first, or reuse a probe from this session if it is already on the same host.

CSR_SHELL, retrieval/user_fetch BOT_DIFFERENTIAL, and retrieval/user_fetch ROBOTS_BLOCKS are STOP conditions. Training-bot policy is not. Advertising-class tokens (OAI-AdsBot) are not.

Step 3: Pull Live Data

Skip this step on STOP.

Semrush (PRIMARY):

1. mcp__semrush__execute_report — Domain Overview (database=us). Authority + organic baseline.
2. mcp__semrush__execute_report — Referring Domains, display_limit=20. ChatGPT retrieval is gated hard by referring-domain count. Note the RD total in the verdict.
3. mcp__semrush__execute_report — Domain Organic Search Keywords, display_limit=30. Separate branded vs unbranded. GEO probes use both.

GSC (ENRICHMENT): 4. mcp__gscServer__get_search_analytics — last 28 days, dimensions=[query], rowLimit=50. Queries that already impress are the first prompt set.

GA4 (ENRICHMENT): 5. mcp__google-analytics__run_report — sessions by landing page, last 28 days. Pages that already convert are the pages that should become citable, not a new explainer silo.

CamoFox (ENRICHMENT): 6. Snapshot the homepage + the #1 converting page. Check: brand name in H1 / first sentence, server-rendered facts (price, specs, city, phone), author byline, outbound citations, FAQ block, schema present, no "for LLMs" disclosure. If CamoFox is down: [SEO] CamoFox unavailable — on-page extractability check skipped. Do not use CamoFox as a substitute for /geo-crawl. CamoFox is a rendered browser; the probe is a non-rendering fetch.

Optional OpenSEO enrichment (never PRIMARY): OpenSEO (every-app/open-seo, hosted openseo.so) combines standard DataForSEO-backed SEO workflows with UI-based AI-visibility features.

Capability boundary checked 2026-09-02:

• The product UI includes Brand Lookup, Prompt Explorer, cited-page / cited-source tables, and Share of Voice.
• The repository's MCP server does not register those AI-visibility functions. Do not instruct an agent to call nonexistent OpenSEO MCP tools for brand lookup or prompt exploration.
• Hosted AI visibility is plan-gated. Self-hosting removes the app subscription gate, but it does not remove DataForSEO usage costs.
• Vendor estimates are enrichment. They do not replace the repeated /geo prompt matrix or verified crawler evidence.

Use OpenSEO only as a compact cross-site enrichment layer if connected. Keep Semrush + GSC as the standard SEO truth, /geo-crawl as reachability, the owned drain as verified crawler evidence, and the repeated prompt matrix as citation truth.

If a source fails, print the [SEO] warning from core.md and continue.

Step 4: Prompt Probe Set

Skip this step on STOP.

Build 8–12 prompts the engines will actually be asked. Split:

Run each prompt against whatever surfaces the operator can reach this session (ChatGPT Search, Perplexity, Gemini, Claude with web search, Google AI Mode / AI Overviews). Repeat ChatGPT prompts — backend switching makes a single snapshot unreliable.

Score each surface per prompt:

Step 4b: Fan-out diagnostic (one page)

Skip on STOP. Follow fan-out-diagnostic.md.

Take one money or category-recommend prompt that already maps to the #1 converting URL. Infer at most six branches from the closed list. Map each branch onto an existing heading on that URL or onto an existing silo page. Label landings SAME_URL or SILO_PAGE. Never NEW_URL.

If AI Mode / AI Overviews is reachable this session, load the seed prompt once and list cited domains. Otherwise mark every branch INFERRED. The Recommended Action is one heading + 134–167 word passage on the existing BOFU page. Do not treat branch coverage as a citation probability.

Step 5: Generate the GEO Playbook

Section 1 — Verdict Open with the /geo-crawl handshake (STOP / CONTINUE WITH FIXES / CLEAR) and the one-line citation verdict if the run proceeded. Are we in the answer, next to the answer, or invisible? Name the highest-leverage missing citation. On STOP, this section is the whole playbook.

Section 2 — Citation matrix Skip on STOP. Table: prompt × surface × score × who got cited instead. No theater. Empty cells stay UNTESTED.

Section 3 — Entity and off-site pool Skip on STOP. Engines fan out to the sources they already trust. List the third-party pages that should mention this brand (G2/Capterra/Clutch if software, GBP + local citations if local-service, Wikipedia / Wikidata only if notable, Reddit threads that already rank, YouTube titles + transcripts). Route the actual link asks to /hunter and /kilo.

Personal-site / founder-entity properties: quality referring domains help both classic SEO and AI appearance, and they are one path toward a knowledge panel. Do not wallpaper generic directories onto a local-service money site.

Section 4 — On-page citability (BOFU pages only) Skip on STOP. For the top 2–3 converting URLs, specify the smallest patch that makes a passage reusable:

• Proper noun in the first sentence
• One standalone 134–167 word passage that answers the category prompt without the rest of the page
• One dated statistic or price with an outbound source
• Comparison table only if the page already competes on "vs" queries
• FAQ answers as atomic paragraphs, not a new /blog
• Server-rendered facts — prices and specs must not live only in JS
• Fan-out heading match for the single worst MISSING branch from Step 4b — on this URL, not a new one

Keep pages at core.md length caps. GEO is extractability, not word count. If /geo-crawl already flagged THIN_HTML or CSR_SHELL, the fix is render-or-expand, not a new URL.

Section 5 — Crawler pointer Do not re-parse robots.txt here. One paragraph: handshake verdict, flag codes, link to the probe report. Retrieval vs training is decided in /geo-crawl.

Section 6 — Measurement Skip detail on STOP. What to re-run in 28 days: /geo-crawl if any CONTINUE fix shipped, then the same prompt matrix, RD count, branded GSC queries, and OpenSEO AI-visibility UI/export if the operator has access. Do not use scraped Google rank-tracker URL paths as proof — /goto ate those.

Section 7 — What this will not do

• Will not replace Map Pack for local
• Will not buy position one
• Will not make a login-walled tool citable
• Will not reward a page that announces it is spam for machines
• Will not cite-optimize a host the retrieval bots cannot fetch
• Will not farm a URL per inferred fan-out branch

Step 6: Output

Follow the Output Protocol from core.md:

1. Print the full GEO playbook to terminal
2. Extract structured summary (include crawl handshake)
3. Save to Graphiti with name GEO — [domain]

In Recommended Actions, first item is the crawl fix on STOP or CONTINUE WITH FIXES. On CLEAR, first item is the single prompt + surface where a one-page patch would most likely flip ABSENT → CITED this month. If Step 4b produced a MISSING branch, that heading patch is the same item — do not list a second content workstream.