Glossary

Plain-English definitions of GEO, AI, and LLM terms as we use them in production. Each entry explains what it is and what it isn't.

GEO (Generative Engine Optimization)

What: Optimization discipline for making your content the answer that AI engines cite.

Isn't: Not the same as SEO. SEO fights for top-10 in blue links; GEO fights for citation inside AI-generated answers.

Matters when: Your buyers ask ChatGPT / Perplexity / Gemini before Google.

SEO (Search Engine Optimization)

What: The classic discipline of ranking in Google / Bing search results.

Isn't: Not dead — Google still handles most search volume — but the top-of-funnel is migrating to AI chat.

Relationship: GEO is layered on top of SEO. GEO content usually helps SEO too, but SEO alone doesn't produce GEO wins.

LLM (Large Language Model)

What: A neural network trained on text to produce coherent language output. Examples: GPT-4, Claude 3, Gemini, Llama 3, Qwen 2, DeepSeek V2.

Isn't: Not intelligent, not consistent, not deterministic. It's a probabilistic text generator with strong pattern recognition.

AI Engine / Answer Engine

What: A product that uses an LLM to answer user questions in a conversational or structured way. Examples: ChatGPT, Perplexity, Gemini, Claude, Copilot, Doubao, Kimi.

Isn't: Not just the LLM alone. Includes: retrieval, ranking, prompt orchestration, citation, moderation.

RAG (Retrieval-Augmented Generation)

What: LLM answers grounded in retrieved documents (from your knowledge base, docs, or web index) rather than only from model parameters.

Isn't: Not fine-tuning. RAG changes what the model sees at query time; fine-tuning changes what the model knows internally.

Matters when: Data changes often, or you need to cite sources, or you can't fine-tune (data too sensitive / too little).

Fine-tuning

What: Continuing to train an LLM on your specific data, updating its weights.

Isn't: Not the first tool to reach for. Try RAG first — cheaper, safer, updatable.

Matters when: You need domain-specific style / format the base model can't do, or you need low-latency + tight prompt, or you have a proprietary vocabulary.

Prompt Engineering

What: The craft of writing the instructions given to an LLM to get consistent, useful output.

Isn't: Not a permanent job title — as models improve, prompts get simpler. But prompt versioning is a permanent engineering practice.

Agent (AI Agent)

What: An LLM-driven system that plans, executes tools, and iterates toward a goal — not just single-turn Q&A.

Isn't: Not the same as a chatbot. A chatbot answers; an agent acts (calls APIs, reads files, executes code).

Matters when: Task requires multiple steps, tool use, or long-horizon planning.

Hallucination

What: LLM confidently outputs something false — non-existent facts, fake citations, wrong numbers.

Isn't: Not the same as a general error. Hallucinations look plausible, which makes them dangerous.

Mitigations: RAG (grounding), retrieval-only mode, output verification, human review for high-stakes.

E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness)

What: Google's quality signal for content. AI engines apply similar concepts. Preferred content shows first-hand experience, subject expertise, authority (citations to it), and trust (contact info, real author, credentials).

Isn't: Not just a checkbox. Engines detect voice / specificity / originality.

JSON-LD

What: Structured data embedded in HTML, describing the semantic meaning of a page (WebSite, Organization, Article, FAQPage, HowTo, Person, Product, etc.).

Isn't: Not optional for GEO. AI engines that produce structured answers (Gemini AI Overviews, ChatGPT web search) preferentially cite pages with valid JSON-LD.

FAQPage schema

What: A JSON-LD type that marks a page as a Frequently Asked Questions structure with Question / Answer pairs.

Matters for GEO: Engines can extract Q&A directly and cite. Buyer-question pages in this hub use FAQPage schema — see one.

HowTo schema

What: JSON-LD marking a page as step-by-step instructions.

Matters for: Tutorial content, checklists, procedures — engines cite the steps.

llms.txt

What: A robots.txt-like file for AI crawlers, letting site owners declare policies for GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and others.

Isn't: Not required by all engines yet, but growing adoption. Include one to be explicit.

IndexNow

What: An open protocol (Microsoft + Yandex) for sites to push URL updates to search engines instantly — Bing indexes within minutes.

Matters for GEO: Faster Bing index → faster ChatGPT web-search citation (ChatGPT uses Bing).

Model Routing / Model Gateway

What: An abstraction layer that picks which LLM to use per request (based on cost, quality, latency, region).

Isn't: Not just load balancing. Includes fallback, versioning, cost tracking, per-tenant policies.

Matters when: You use multiple providers (OpenAI + Anthropic + Alibaba + private) and want vendor independence.

Semantic HTML

What: HTML5 tags with meaning — article, nav, aside, section, header, footer — as opposed to div everywhere.

Matters for GEO: Engines use these hints when extracting structure and citations.

What: A URL fragment #section-id that jumps to a specific H1-H4 heading.

Matters for GEO: Engines can cite an exact passage, not just the page. This hub adds copy-link icons to every H1-H4.

Independent Domain

What: Your own domain (ailaopao-geo.pages.dev, yourbrand.com) — not a rented subdomain on someone else's platform (weibo.com/yourbrand, wechat.com/user/xyz).

Matters for GEO: Engines attribute authority to the domain, not to the user account.

Canonical URL

What: The "true" URL of a piece of content, declared via <link rel="canonical">. Prevents duplicate-content issues when the same content exists on multiple URLs.

Matters for GEO: Reverse-migration — content you published on CSDN / Medium / dev.to should link back to the canonical page on your own domain.