Geo First SEO

intent

use this skill when you want content to get cited inside ai-generated answers rather than ranked as a clickable link. traditional seo optimizes to rank and get clicked. geo (generative engine optimization) optimizes to be quoted. you run four phases: strategy and query research, content creation or audit against seven principles, technical markup (json-ld, llms.txt, semantic structure), and a scorecard to iterate. this workflow applies to new content or existing pages. skip it if you're doing keyword-rank audits, paid ads, or link-building.

inputs

user-provided:

• topic or page url (whether creating new content or auditing existing)
• target audience (who must trust the answer; shapes vocabulary and depth)
• target ai engines: chatgpt, perplexity, google ai overviews, gemini, copilot (defaults to all)
• query cluster: the real user questions or prompts the content should win citations for (e.g. "what is x", "x vs y", "how to do z")
• for audit mode: existing draft text or live url to analyze

external connections (optional but stronger):

• websearch/webfetch capability: used in phase 1 to scan what engines currently cite for top 2-3 target queries. without it, work from pasted material and ask the user what competing content exists. never invent live search results.

context / knowledge:

• you have access to references/geo-content-tactics.md (before/after rewrites per principle, per-engine notes) and references/technical-geo.md (json-ld snippets, llms.txt template, markup checklist). open them when you need examples or copy-paste blocks.

procedure

phase 1: strategy & intake

step 1.1 (input: user intent) ask the user the missing fields from the intake table below. wait for each answer before moving to the next question. do not invent answers.

field	why it matters
topic / page	subject, and whether creating new content or auditing an existing url or pasted draft
audience	who must trust the answer; sets vocabulary and depth
target engines	chatgpt, perplexity, google ai overviews, gemini, copilot. defaults to all unless user narrows it
query cluster	real user questions/prompts the content should win citations for

output: captured intake data.

step 1.2 (input: web access availability) confirm whether websearch/webfetch is available. if yes, proceed to 1.3. if no, skip to 1.4.

output: availability flag.

step 1.3 (input: live citation research, conditional on web access) for the top 2-3 target queries, search what ai engines currently cite. note which sources win, what claims they make, what is missing or outdated or unsourced.

output: citation-gap summary (current cited sources, gaps, angles).

step 1.4 (input: competitive context, conditional on no web access) ask the user what competing content exists for the target queries. use their answer instead of live research.

output: user-provided competitive overview.

step 1.5 (input: content brief assembly) synthesize the intake, citation gap (or competitive context), and user's angle into a short brief:

• target query cluster (the questions to answer)
• entities and subtopics that must be covered
• the angle / unique substance this content adds (data, first-hand experience, clearer definition, fresh research)

present the brief to the user for confirmation before phase 2.

output: confirmed content brief.

---

phase 2: create or audit

step 2.1 (input: mode selection) confirm whether the user is in create mode (new content from brief) or audit mode (existing draft or url).

output: mode flag.

step 2.2 (input: material for audit mode, conditional) if audit mode: request the existing draft or url. fetch or paste it.

output: audit material.

step 2.3 (input: apply seven geo content principles) apply all seven principles to the content (new or existing). expand examples from references/geo-content-tactics.md as needed:

1. answer-first: put the direct answer in the first 1-2 sentences of the page and of each section (inverted pyramid). lead with conclusion, then support it.
2. self-contained chunks: each section must answer one question and stand alone without surrounding context. no "as mentioned above".
3. entity & semantic coverage: name the entities, define key terms explicitly, cover related concepts and questions a reader would expect. completeness signals authority.
4. citable elements: include statistics, concrete data, named sources, dated facts, direct quotes. these are the units engines lift. attribute every figure.
5. scannable structure: use question-style h2/h3 headings, short paragraphs, ordered/unordered lists, comparison tables, dedicated faq block.
6. authority & freshness signals: show a named author with relevant credentials, link primary sources, include visible "last updated" date.
7. plain, unambiguous language: write so a model can parse and quote a sentence with no surrounding context. avoid vague pronouns, hedging, clever phrasing that obscures the claim.

if audit mode, diagnose specific weaknesses against each principle before rewriting. show the user what was weak.

output: optimized content (full draft or rewrite), and for audit mode a bulleted list of diagnosed weaknesses fixed.

---

phase 3: technical geo

step 3.1 (input: content from phase 2) take the optimized content.

output: markup ready to paste.

step 3.2 (input: schema.org json-ld generation) add structured-data types that fit the content: Article/BlogPosting, FAQPage, HowTo, Organization, author (Person). mark only content that actually appears on the page. use references/technical-geo.md for snippets.

if you do not know a real value (author name, date, url), insert a clearly labeled placeholder like [PLACEHOLDER: author name] or [PLACEHOLDER: publication date]. never fabricate.

output: copy-paste json-ld blocks.

step 3.3 (input: llms.txt generation, conditional on multi-page context) if the content is part of a larger site, generate an llms.txt file (and optionally llms-full.txt) listing key pages and concise descriptions to guide ai crawlers. use the template from references/technical-geo.md.

output: llms.txt block.

step 3.4 (input: semantic html structure) confirm one <h1>, logical heading hierarchy, descriptive <title> and meta description, real faq/q&a markup matching on-page faq.

output: semantic structure checklist or corrections.

---

phase 4: geo scorecard & iterate

step 4.1 (input: content + markup from phases 2-3) score the result against the seven criteria in the table below.

criterion	pass condition
answer-first	page and each section open with the direct answer
chunk self-containment	every section stands alone when read in isolation
citable elements	real stats / quotes / named sources present and attributed
entity coverage	key entities defined; expected subtopics and questions covered
structure & markup	question headings, lists/tables, faq, valid json-ld present
authority & freshness	named author, primary sources, last-updated date
query coverage	target query cluster is each answered explicitly somewhere on page

output: scorecard with pass/fail per criterion.

step 4.2 (input: user review or iteration) present the scorecard. for any failing criterion, name the fix and revise. repeat until user is satisfied or all criteria pass.

output: final revised content and scorecard.

---

decision points

• if user skips phase 1 (quick audit only): confirm before skipping. do not skip in create mode; a content brief is required.
• if web access is unavailable: work from pasted material. ask the user what competing content exists. never invent live-search results or citation data.
• if audit mode and weaknesses exist: always diagnose and show them to the user before delivering the rewrite. do not hide problems.
• if a statistic, quote, source, date, author credential, or search-citation claim would strengthen the content but you lack it: mark it [verify] or ask the user. do not invent authority.
• if the content is multipage and has an existing site structure: generate llms.txt. if single-page or unknown context, skip llms.txt.
• if json-ld values are unknown (author, date, url): insert labeled placeholders like [PLACEHOLDER: author name]. never fabricate.
• if the user expresses a need outside the four phases: append "this skill may not fully cover your situation. suggestions for improvement are welcome open an issue or pr."

---

output contract

phase 1 deliverable:

• confirmed content brief: target query cluster, entities/subtopics, unique angle

phase 2 deliverable:

• optimized content: full draft (create) or rewritten page (audit)
• in audit mode: bulleted list of diagnosed weaknesses fixed

phase 3 deliverable:

• json-ld blocks (copy-paste ready), with placeholders for unknown values
• llms.txt block (if multipage context)
• semantic structure checklist or corrections

phase 4 deliverable:

• geo scorecard: pass/fail per criterion
• revised content addressing any failed criteria
• final signed-off scorecard when all or most criteria pass

all content delivered must be publishable, not instructional. hand over usable text and markup, not a lecture about geo. never fabricate data, sources, or credentials. if uncertain, mark [verify] or ask the user.

---

outcome signal

you know the skill worked when:

• the content brief captures the user's intent and target queries with no ambiguity
• the optimized content passes at least 6 of 7 scorecard criteria on first iteration
• each section can be read standalone and quoted without surrounding context
• json-ld validates (using google's structured data tester or similar)
• the user confirms the content is ready to publish

optionally: after publishing, the user reports that the content appears in ai answer citations (visible in chatgpt, perplexity, or google ai overviews results) within 2-4 weeks of indexing.

---

key rules

• geo is "be quoted", not "be ranked". optimize passages for extraction and standalone sense, not keyword density.
• the content brief is required before drafting in create mode.
• in audit mode, always diagnose weaknesses before delivering the rewrite.
• mark json-ld and metadata only for content that actually appears on the page.
• never invent statistics, quotes, sources, study results, dates, author credentials, or search-citation data. fabricated authority damages credibility and carries legal risk.
• web access is optional and read-only. only fetch urls the user provides. never publish, push, deploy, or recommend cloaking, hidden text, or manipulative tactics.
• ask one question at a time when required information is missing, then wait for the answer.

---

references

expanded examples and copy-paste templates live in:

• references/geo-content-tactics.md (before/after rewrites per principle, per-engine notes)
• references/technical-geo.md (json-ld snippets, llms.txt template, markup checklist)

open them when you need specific examples or ready-made blocks.

---

original author: archlab-space