GEO Factory Operations — Prompt Research, Production Pipeline & Citation Measurement

For factory operators. The KB already documents what works in GEO (the llm-search mechanics, the Princeton-validated tactics, the technical plumbing). This document is the missing operating system: the three repeatable loops that turn those tactics into a factory — discover (prompt research) → produce (content pipeline) → measure (citation tracking + experiment) → re-prioritize. It also covers what changes for a local-language market (the running example: a Romanian-language site).

The thesis: GEO tactics are necessary but not a factory. A factory is the loop that decides which prompts to win, produces extractable content against them at cadence, measures whether you actually got cited, and feeds that back. Without the loop you're hand-optimizing pages; with it you compound.

1. Discover — Prompt Research (the "keyword research" of GEO)

Traditional SEO researches keywords; GEO researches prompts — the full natural-language questions buyers ask an LLM. The method:

Source the prompts (where real questions live):
- First-party demand: sales-call transcripts, support tickets, and the site search box — the highest-intent questions your buyers actually ask.
- Community mining: Reddit, niche forums, Facebook groups, Q&A sites — phrased the way real people ask.
- Search-adjacent: People Also Ask / AlsoAsked / autocomplete — still the best proxy for question phrasing.
- Query fan-out expansion: every buyer prompt decomposes into 4–8 sub-queries (the RAG fan-out from llm-search); expand each seed prompt into its likely sub-queries and target those too.
- Competitor citation gaps: prompts where competitors are cited and you aren't — the highest-ROI targets.

Classify by intent: informational ("how do EV chargers work"), commercial ("best home EV charger"), comparison ("X vs Y"), and local ("EV charger installation near me"). Commercial and local prompts convert; informational prompts build the entity authority that earns commercial citations.

Prioritize by buying-intent × citation-gap: a high-intent prompt where you're absent but could be cited is worth more than a high-volume informational prompt you already win. This prioritized prompt set is the factory's backlog — and the fixed panel you measure against (§3).

2. Produce — The GEO Content Pipeline (agent workflow)

The repeatable production loop, an agent-workflow-pattern specialized for GEO and orchestrated via agent-orchestration:

prompt (from backlog)
  → answer-first brief  (the question + the exact answer + required entities/stats/sources)
  → draft               (apply the validated tactics: lead with the answer in <=150 words,
                         add a statistic, add a named quote, cite sources, fluent prose)
  → HUMAN GATE          (factual accuracy — hallucinated specs are a brand/safety risk)
  → schema + structure  (JSON-LD: Organization/Article/FAQPage/HowTo; semantic headings)
  → publish             (dual-format .html + .html.md; sitemap; IndexNow push)
  → entity hooks        (link the brand entity; seed third-party presence — see content-authority)
  → enqueue for measurement (§3)

This reuses the content-machine for production but is GEO-shaped: every output is built as independently-extractable 50–150 word chunks, not a flowing essay. The human-review-gate on factual accuracy is non-negotiable (an LLM inventing a charger's kW rating or a price is a real liability). Off-page authority (content-authority) runs as a parallel program: entity definitions (Wikidata), and presence on the third-party sources LLMs over-cite in your category.

3. Measure — The Citation Loop (and why it's a small-N experiment)

You cannot use Search Console for AI citation. The factory runs its own measurement:

The prompt panel: run the prioritized prompt set (§1) across engines (ChatGPT, Claude, Perplexity, Google AI, plus any locally-dominant engine) on a fixed schedule; for each, record whether you're cited, where in the answer, and which competitors appear.

Metrics: AI Citation Frequency (AICF) / Share of AI Voice (your citations ÷ total category citations), brand coverage rate (% of target prompts where you appear), and citation-gap (prompts competitors own that you don't). (Tooling examples — Otterly, Peec AI's "citation gap analysis", Profound — are 2026 vendor options, cited as secondary; the metrics are the durable part.)

Why measurement is hard — and is a small-N problem. Citation is probabilistic: the same prompt fans out differently and returns different sources run-to-run, so a single check is noise. This is exactly minimum-detectable-effect territory. Treat "did adding statistics lift our citation rate for these prompts?" as an experiment:
- Sample each prompt multiple times to estimate a citation rate, not a yes/no.
- Use the bayesian-decision-rule (posterior over citation rate) and seed it with a marketing-benchmark-prior so thin samples don't whipsaw.
- Don't over-react to a single run; pre-register what lift would justify a content change — this is the experiment-loop applied to GEO.

The loop closes: measured citation gaps and wins feed back into §1's prioritization.

4. Local-Language GEO (e.g. Romanian)

Running GEO in a non-English market changes the factory in specific ways:

  • Thinner citation supply = bigger opportunity. LLMs have far fewer authoritative sources in smaller languages, so the bar to become the cited source is lower — if you deliberately build the entity and publish extractable local-language content, you can own a category's answers faster than in English.
  • Mine local-language demand: prompt research must use local forums, local Facebook groups, and local-language PAA/autocomplete — English prompt sets don't transfer.
  • Entity authority in-language: define the brand entity in the local Wikidata/Wikipedia and seed locally-trusted third-party sources; LLMs lean hard on whatever authoritative local content exists.
  • Local intent dominates: for a geographically-bound business, "near me / in [city]" prompts are the commercial core — pair GEO with local schema (LocalBusiness, address, service area).
  • Verify in-language: run the measurement prompt panel in the local language; citation behavior differs from English.

5. Factory Integration

GEO becomes a closed loop the factory runs continuously: discover (prompt research → prioritized backlog) → produce (content-machine pipeline with validated tactics + schema + dual-format publish) → measure (prompt-panel citation tracking as a small-N experiment-loop) → re-prioritize. It is orchestrated by agent-orchestration, gated by human-review-gate on factual accuracy, and grounded in the llm-search mechanics. Most of it is agent-ownable; the human owns factual sign-off and strategic prompt prioritization.

Provenance Note

  • Builds on (already verified in this KB): the GEO mechanics, the Princeton-validated 9 tactics, citation economics, robots.txt/schema/llms.txt plumbing — see llm-search and its source (Aggarwal et al., KDD 2024, previously fetched & verified).
  • Synthesized from 2026 GEO practitioner/vendor sources (secondary, NOT a single fetched primary): the prompt-research methodology, the "prompt intelligence / prompt set" framing, and the AI-visibility metrics (share of voice, citation rate, brand coverage, citation-gap). Tool names (Otterly, Peec AI, Profound) are illustrative, not endorsements or verified specs.
  • First-principles (this KB): framing citation measurement as a small-N probabilistic experiment and wiring it to the experiment-loop / bayesian-decision-rule / marketing-benchmark-prior; the local-language analysis.

Sources

Concepts

Extracted from this source: geo-prompt-research · geo-content-pipeline · geo-citation-measurement

Related concepts: llm-search · content-authority · schema-markup · content-machine · agent-orchestration · experiment-loop