# The GEO Field Manual: An Operating Framework, From Crawl to Citation > Sixteen prior essays distilled into one operating framework: the four-layer funnel, the audit points that survive evidence review, and the sequence that works. - Canonical: https://www.aixindar.com/news/the-geo-field-manual-an-operating-framework-from-crawl-to-citation - Markdown: https://www.aixindar.com/news/the-geo-field-manual-an-operating-framework-from-crawl-to-citation.md - Author: Daoyu Guan — https://www.aixindar.com/experts/daoyu-guan - Published: 2026-09-07T06:53:25.217Z - Last updated: 2026-09-07T06:53:25.285Z - Evidence checked: Not separately recorded in CMS - Editorial status: Published - Corrections: No correction record supplied by CMS. Nineteen essays produce a lot of conclusions. The risk of a series this deep is that the reader leaves with fragments — a volatility figure here, a citation-share number there — and no operating order. This manual is the consolidation. It is deliberately opinionated about sequence, because in the GEO field the order of operations is where most programs fail: teams invest in content before access, measurement before baselines, and mentions before entities. One framing sentence governs everything below, earned through the series' evidence review: **the answer layer rewards infrastructure and penalizes improvisation.** Work flows from that. ## The Funnel: Four Layers, Each a Gate Everything an AI system does with your brand passes through four gates. Each gate is a pass/fail filter, and failure at any layer is invisible from the layers above — which is why diagnosis must run bottom-up. **Layer 1 — Discovery (Can the machines reach you at all?)** The crawl layer, where programs die quietly. The foundational technical fact: most AI crawlers do not execute JavaScript. Client-side-rendered content — increasingly common in modern web stacks — is effectively invisible to the answer layer regardless of quality. The Layer-1 checklist: server-side rendering for AI-relevant content; per-bot robots.txt posture (search, training, and agent bots warrant different policies); no accidental blocking of the fetchers you actually want; and crawl-log verification that AI user agents are arriving and fetching *current* versions of key pages. Layer 1 is the least glamorous work in this manual and the only one whose failure zeroes out everything above it. **Layer 2 — Retrieval (When relevant questions trigger search, do systems pick you up?)** The evidence review established that the retrieval stage is where most optimization battles are decided — and where content rewriting never reaches. Layer-2 work is structural: semantic HTML (headings that mean something, lists and tables where content is list-like or tabular), JSON-LD structured data (Article, FAQPage, Product, Organization — the pillar cluster GEO-16 found most correlated with citations), passage-level extractability (one concept per section, self-contained paragraphs, facts that survive compression), and metadata freshness. The operational bar from the GEO-16 research: pages scoring **G ≥ 0.70 with at least 12 of 16 quality pillars** earned significantly higher citation rates — a concrete audit target, not a vibe. **Layer 3 — Citation (When answers are built, are you in the source pool?)** The earned-media layer, where the citation-pool research applies: earned media accounts for the overwhelming majority of AI citations, comparison/listicle formats alone draw roughly a third, and the platforms diverge structurally (Reddit dominant for some engines, Wikipedia for others, YouTube the strongest single mention signal in the largest brand study). Layer-3 work is the mention portfolio — review platforms, community participation, trade media, original data that others cite — executed per-platform, not blasted uniformly. **Layer 4 — Recommendation (When answers describe your category, are you the suggestion?)** The distinction between being cited and being recommended is the difference between a footnote and a shortlist. Layer 4 runs on entity strength: consistent naming and facts across every surface (the verification loop from our trust-recession piece tests exactly this), knowledge-graph completeness, review depth and recency, and comparison-query coverage — the fan-out types carrying 97% of brand mentions. Entity weakness caps everything above it: a brand the model cannot confidently describe is a brand the model hedges on, and hedged brands do not get recommended. ## The Sequence: A Twelve-Month Operating Order Priority order matters more than completeness. The sequence that the evidence supports: **Months 1–2: Diagnose bottom-up.** Crawl logs for AI bots (Layer 1), an extractability audit of your ten most decision-critical pages, a pre-registered prompt set built around entity and comparison clusters (per the measurement discipline), and a baseline across each platform you care about — separately, never blended. **Months 2–4: Fix the floor.** SSR for AI-relevant content, semantic HTML and schema on priority pages, the extraction repairs (dates on facts, ungated specifications, clean section structure). Unglamorous; decisive; mostly one-time engineering. **Months 4–8: Build the mention portfolio.** Review-platform presence with authentic review velocity; community participation as resident expert; trade-press and original-data placement; per-platform content aligned to each engine's citation profile. This is the longest-horizon work in the manual — start it early precisely because it compounds slowly. **Months 6–12: Close the entity and freshness loops.** Knowledge-graph and reference-layer presence; a maintained fact baseline (brand canon) that every surface agrees with; a refresh production line for proven URLs — the freshness signal that the largest visibility study found predictive when everything else was noise. **Ongoing: Measure the loop, not the citation.** Quarterly rebaselining, per-platform volatility bands, downstream metrics (branded search lift, direct visits, review-profile traffic) — because a citation is a midpoint, and the verification economy pays on confirmation. ## Three Structural Principles Beneath the Tactics **1. GEO is not a channel separate from SEO — it is SEO's retrieval layer made explicit.** The engines fetch from the same web, reward the same fundamentals, and (per Google's own stated position) treat AI-surface eligibility as an extension of search quality. Organizations that split "SEO team" and "GEO team" reproduce the exact fragmentation the field's failed tools died of — one GEO-only startup's shutdown is the cautionary tale for monitoring-without-integration. **2. The retrieval layer has unglamorous depth most of the field never discusses.** Beneath keywords and content sit vector indexes and embedding spaces — the machinery (HNSW-style approximate nearest-neighbor indexes) that decides which passages are even *candidates* for an answer. One documented case study (Pinterest's AI-search program combining vision models, agent-assisted structure, and vector-retrieval-aware internal linking) reported roughly **20% organic traffic growth** — a reminder that the deepest technical layer is where the least-contested advantages live. Most organizations will never need to touch vector indexes directly; all of them benefit from understanding that passages compete in *similarity space*, which is why semantic self-containment (Layer 2) beats keyword density. **3. Organizationally, the work splits into three disciplines.** The community's most durable decomposition: the **Strategist** track (visibility measurement, brand protection, the measurement discipline), the **Builder** track (content structured for retrieval, extraction, and citation), and the **Engineer** track (crawl access, rendering, structured data, agent-facing surfaces). Small teams stack all three hats; large organizations should resist merging them into one "content" function, because the failure modes differ per track and the audit questions differ per layer. ## The Honest Caveats - **The framework is a synthesis, not a law.** Each layer's evidence strength differs: Layer 1's rendering facts are mechanical certainties; Layer 3's platform profiles are observational; the operational thresholds (G ≥ 0.70) come from one observational study on English B2B SaaS pages. Calibrate to your vertical. - **Volatility governs everything.** With monthly citation volatility at 40–59% across engines, any layer's "results" need quarters of longitudinal data before they are interpretable. The manual's timeline is a minimum, not a promise. - **The field moves at benchmark speed.** The agentic turn, protocol standards, and the next benchmark generation will revise parts of this manual within a year. The layers are stable; the thresholds are dated. ## The Bottom Line Nineteen essays of evidence compress into one funnel and one sentence. The funnel: **Discovery → Retrieval → Citation → Recommendation** — four gates, diagnosed bottom-up, each capable of silently zeroing the others. The sentence: **the answer layer rewards infrastructure — access, structure, entities, earned presence — and the organizations that build it in sequence will outperform the ones still buying tactics.** The GEO field's first generation was a land rush — noisy, speculative, heavy on guarantees. The second generation, informed by the benchmark literature and the measurement discipline, looks like this manual: unglamorous, layered, compounding. The window for being early is closing; the window for being rigorous never closes. --- ### Sources - Kumar & Palkhouski, "GEO-16" (2025; arXiv:2509.10762): 16-pillar audit; G ≥ 0.70 + ≥12 pillars → significantly higher citation rates; top pillars — metadata & freshness, semantic HTML, structured data (1,702 citations, three engines; observational, English B2B SaaS) - The GEO Lab GEO Stack framework (2026): retrieval probability / extractability / entity reinforcement / structural authority / system memory - The GEO Community three-track model: Strategist / Builder / Engineer; GEO-and-SEO integration stance (Aleyda Solis); Lorelight shutdown as integration-failure case - C-SEO Bench (NeurIPS 2025) and SAGEO Arena (2026): retrieval-stage primacy; structural signals over content rewriting (the evidence review behind this manual's priorities) - Ahrefs 75,000-brand study (2025–2026): YouTube mentions \~0.737, unlinked web mentions 0.664 — the mention portfolio's empirical base - Moz 50k fan-out dataset (2026): entity + comparison clusters = 97% of brand mentions - Pew Research Center (2025) and Ahrefs (2025): query-structure trigger gradients for the funnel's entry conditions - Growth Memo freshness study (2026): 5.3M results; freshness the only predictive top-3 attribute - Pinterest AI-search case (2026): \~20% organic growth via VLM + agent-assisted structure + vector-retrieval-aware linking - Cross-context: all seventeen prior analytical pieces of this series *Thresholds and layer evidence are as of September 2026; recalibrate quarterly per the measurement discipline.* ## Editorial references - [Editorial policy](https://www.aixindar.com/editorial-policy) - [Research methodology](https://www.aixindar.com/research-methodology) - [Corrections policy](https://www.aixindar.com/corrections)