The GEO SOP

The standard operating procedure for generative engine optimization resource pages.

This document is the substantive spec behind auto-geo’s schema and validation heuristics. Every word-count constraint in core/schema.ts and every soft-warning threshold in core/validation.ts traces back to a section here. Read this document before deviating from any rule the code enforces.

The SOP is calibrated to the public retrieval and citation behavior of major AI search engines (ChatGPT, Perplexity, Google AI Overviews, Google AI Mode, Claude, Gemini, Copilot) as of early 2026. Specific numerical thresholds (e.g., citation lift percentages, density ratios) are heuristic — they reflect the working consensus from internal experimentation and external GEO research. Treat them as starting points, not laws.

§1. Definition

A GEO resource page is a public web page whose architecture, density, and metadata are engineered to be retrieved and quoted by large language models acting as search interfaces.

GEO is the successor to SEO. Where SEO optimized for keyword matching against an inverted index, GEO optimizes for citation by AI engines that answer questions on the user’s behalf. The win condition is not a click — it is being quoted in the AI’s answer, with the user’s downstream click being a bonus.

§2. The three optimization frameworks

A GEO resource page typically serves one or more of:

A page can serve all three. core/schema.ts’s geoMetadata.optimizationFramework field records which.

§3. Question-format H2 headings

H2 headings are written as the questions a user would type into an AI engine, not as topic labels. Ideal length: 6-10 words.

Wrong: “Our approach to onboarding” Right: “How does an effective onboarding process work?”

Rationale: AI engines retrieve against query intent. A page whose H2s mirror prompt phrasing is matched directly. A page whose H2s are internal labels requires the engine to synthesize a connection — synthesis costs the engine; direct matches are cheap.

core/validation.ts flags H2s that don’t contain ? or are outside 4-12 words.

§4. Page type and word-count targets

Five page types, each with calibrated word-count and density targets:

Page type Min words Max words Min tables Min lists Stats/1k words
definitive 3000 — 2 5 10
resource 800 1500 0 3 3
comparison 1000 1500 1 3 3
category 2000 4000 1 5 5
listicle 1500 3000 1 5 3

geoMetadata.pageType selects the expectation set. core/validation.ts surfaces soft warnings when page totals are outside the range.

§5. Structural and quality rules

§5a. Title formulas

Five title patterns map to the five page types. Use the formula that matches the page’s intent:

§5b. Section and paragraph length

Rationale: chunks at this length match the optimal extraction window for current retrieval models. Shorter chunks lack context; longer chunks dilute the answer signal.

Each H2 section should contain 1-2 outbound links to authoritative non-self domains. Zero links flags as under-cited; more than 4 flags as aggregator content. Aggregator-style heavy linking carries a measured citation discount — AI engines prefer to cite primary sources, not link-dense intermediaries.

4-8 entries. Self-links are explicitly forbidden (soft warning).

Related Guides are an entity-graph signal: they tell AI engines this page sits inside a coherent cluster of related content. The cluster signal lifts citation probability for every page in the cluster.

§5e. Key Takeaways

4-6 declarative bullets. Each 10-35 words.

These are the page’s claims, surfaced for extraction. AI engines often pull Key Takeaway bullets verbatim into answer responses.

§5f. FAQ

3-10 Q&A items. Each answer 40-60 words.

The FAQ block drives FAQPage Schema.org markup, which is extracted by retrieval pipelines independently of the page body. FAQ items often appear in AI engine responses as direct quotations of the answer text.

§5g. Entity density

15+ named entities per page (companies, people, products, frameworks, places).

The empirical link: entity-dense pages show ~4.8x higher citation probability than entity-sparse pages of similar length. Entity density is a credibility signal — pages that name specific actors are treated as more authoritative than pages that gesture at abstractions.

core/validation.ts runs a simple heuristic (joined capitalized words after stripping inline markers and sentence-starters) and flags pages below 15.

§5h. Banned promotional superlatives

The following phrases carry a measured citation penalty (~26%) when used without attribution:

core/schema.ts rejects these as hard errors unless:

Rationale: AI engines empirically downweight pages whose voice reads as marketing copy. The ban is a hard constraint, not a guideline.

§5j. Mandatory page architecture

In order:

  1. H1 + last-updated metadata
  2. TL;DR (40-60 words)
  3. Intro blocks
  4. Sections (each: H2 + answer capsule + blocks)
  5. Related Guides (4-8)
  6. Key Takeaways (4-6)
  7. FAQ (3-10)
  8. About the Author (auto-injected)
  9. Disclosure

The order is load-bearing. AI engines extract chunks; chunk position matters. TL;DR at top, FAQ near the bottom — this is the pattern retrieval pipelines are trained against.

§6. The answer capsule

Every H2 section opens with a 40-60 word answer capsule — a fully self-contained answer to the section’s heading, written before any supporting paragraph or block.

Rules:

The answer capsule is the highest-value chunk on the page. It is the most likely thing to be quoted verbatim by an AI engine. Every other block on the page exists to support the capsule.

§7. Multimodal density

§7a. Lists, tables, callouts

Lists and tables are extractable as structured data. The minimum-block-count targets in §4 force a baseline of structure per page.

§7b. Images

Image cadence: ~1 image per 500 words. Each image’s alt text must:

Multimodal pages (text + images) show empirical citation lift of 156-317% over text-only pages of similar length. AI engines that ground their answers in retrieved pages preferentially cite pages with strong visual content.

§8. Information gain

geoMetadata.informationGainStatement is a required field stating what the page contains that is not in current AI engine responses for its target queries.

Rationale: AI engines have an implicit novelty filter. Pages that restate what the engine can already synthesize from existing training data contribute nothing — they are not cited because they are not informative. Pages with proprietary data, original analysis, or first-party research provide information gain and are cited disproportionately.

Before publishing, articulate what makes this page non-redundant. If the answer is “nothing,” the page should not exist.

§13. Schema.org markup

Auto-emitted by core/jsonld.ts from the typed payload. The agent never writes JSON-LD by hand. Five schema types are derived:

JSON-LD is rendered inside <script type="application/ld+json"> tags, with text safely escaped via safeJsonLd to prevent breakout.

§14. GEO metadata

geoMetadata is internal metadata stored on the published payload but not rendered to the page. Fields:

The metadata is what makes a page auditable after publication. A team can run a regular audit asking “is this page still surfacing for its target queries, and if not, why.”

§15. Refresh cadence

GEO resource pages decay. The query landscape shifts as AI engines update their training and retrieval pipelines; the answers AI engines synthesize without your page also evolve over time, eroding your information gain.

Two cadences:

geoMetadata.refreshCadence records the page’s cadence. A separate audit pipeline (out of scope for auto-geo’s core) reads the metadata and surfaces pages due for refresh.

Maintenance

This SOP is versioned with auto-geo. When the empirical thresholds in core/validation.ts change, update the corresponding section here. The intent is that this document and the code stay in lockstep — reading the SOP should always tell you what the code is doing and why.

External GEO research is rapidly evolving. Treat specific numerical claims (e.g., “4.8x citation lift”) as snapshots, not eternal truths.