This field has more names than ideas. GEO, LLMO, AEO and AIO describe substantially the same work, while genuinely different things — a citation and a mention, a hallucination and a stale fact — get used interchangeably in ways that cost money. Each entry below states what the term means, what it connects to, and what it is not.
Terms #
Generative Engine Optimization (GEO) #
The practice of making a business or its content more likely to be surfaced, cited and recommended inside AI-generated answers. GEO covers the technical work that makes a site retrievable, the structured data that makes it machine-readable, and the third-party presence that makes its claims corroborated.
LLMO (Large Language Model Optimization) #
Optimization aimed specifically at how large language models represent and recommend a business. In practice LLMO and GEO describe the same work; the term is dominant in Japan, where it is rendered 大規模言語モデル最適化, while GEO is more common in English-language markets.
Answer Engine Optimization (AEO) #
Optimization for being selected as the direct answer to a question rather than one result among many. AEO emphasises question-shaped content, FAQ structure and concise factual statements. It predates the generative-AI wave and originally described featured snippets.
AI Optimization (AIO) #
The broadest of the competing labels, covering any optimization aimed at AI-mediated discovery. Used interchangeably with GEO and LLMO; the choice of term usually signals which market or vendor the speaker comes from rather than a difference in method.
Search Engine Optimization (SEO) #
Optimization for ranked lists of links in a traditional search engine. SEO and GEO overlap substantially — crawlability, structure and authority serve both — but diverge in their goal: SEO wants a click, GEO wants a citation inside an answer that may never produce one.
Retrieval-Augmented Generation (RAG) #
The pattern behind most grounded AI answers: a retrieval step finds candidate documents, then the model generates a response based on what it retrieved. RAG is why current facts about a business can appear in an answer from a model whose training ended long before.
Grounding #
Basing a generated answer on retrieved source documents rather than on the model's internal parameters. A grounded answer usually carries citations and can reflect information published yesterday; an ungrounded one reflects only what the model absorbed in training.
Training data #
The corpus a model absorbed during training, which becomes its baseline knowledge. Information here is fixed until the next model version, which is why a business's outdated price can persist in answers long after the website is corrected.
Knowledge cutoff #
The date after which a model's training data ends. Beyond it, a model knows nothing except what it retrieves at answer time — which is why grounding, not training, is where a business can realistically intervene.
Retrieval #
The search step that selects candidate documents before generation. Retrieval decides which facts and which business names are even available to the answer, making it the highest-leverage stage for visibility work.
Embedding #
A numerical representation of text that places semantically similar passages near each other. Embeddings are why a page phrased the way a customer asks the question retrieves better than one phrased the way a marketing team writes.
Semantic similarity #
Closeness in meaning rather than in wording. Retrieval systems match on semantic similarity, so matching a customer's phrasing matters more than repeating an exact keyword.
Chunking #
Splitting a document into passages that are retrieved and cited independently. Because retrieval returns chunks rather than whole pages, every section of a page should answer its own question without depending on the surrounding context.
Reranking #
A second scoring pass that reorders retrieved candidates before generation. Reranking is why the document a search step ranked first is often not the one an assistant ends up citing.
Citation #
A source an AI answer explicitly attributes and usually links. A citation is stronger than a mention: it names you as the basis for the claim and can send a reader to your site.
Mention #
A business named inside an AI answer without being cited as a source. Mentions build familiarity and can win the recommendation, but do not drive traffic or establish you as the authority.
Presence rate #
The share of tested prompts whose answers name a given business at all. Presence measures whether you are in the conversation; it says nothing about whether you were recommended or merely listed.
Recommendation rate #
The share of answers that actively recommend a business rather than listing it among options. This is the metric closest to revenue, and typically the lowest of the three.
Hallucination #
A confident, fluent statement that is false. In an AI visibility context the costly hallucinations are specific and checkable — a wrong price, a closed location, a service a practice does not offer, a credential it does not hold.
Conflation #
An AI merging two different businesses into one, usually because their names or locations are similar. Conflation is distinct from hallucination: every individual fact may be true, but they belong to two different companies.
Stale fact #
A statement that was true and no longer is — an old price, a former address, a departed partner. Stale facts are the most common category of AI error about a business and the most tractable to fix, because correcting the sources usually corrects the answer.
Entity #
A specific, identifiable thing — a business, a person, a place — as distinct from the words used to refer to it. AI systems reason about entities, so the goal is for every mention of your practice to resolve to one unambiguous entity.
Entity resolution #
Deciding whether two mentions refer to the same real-world thing. When it fails, a practice's reputation gets split across several partial identities, none of which is strong enough to be recommended.
Knowledge graph #
A structured store of entities and their relationships, used by search and AI systems as a stable factual backbone. Wikidata is the most consequential public one for business visibility.
sameAs #
A schema.org property linking a business's canonical page to its official profiles elsewhere. sameAs is the most direct instruction available for telling machines that these accounts are all the same entity.
Structured data #
Machine-readable markup, usually schema.org JSON-LD, stating explicitly what a page is about. Structured data removes the need for a machine to infer facts from prose, and inference is where errors enter.
schema.org #
The shared vocabulary behind most structured data on the web. Relevant types for professional services include Organization, LocalBusiness, MedicalClinic, LegalService, FAQPage and Article.
JSON-LD #
The format in which structured data is normally embedded — a JSON block in the page head. Preferred because it does not entangle markup with layout.
llms.txt #
A plain-text file at a site's root summarising what the site is and linking its most useful pages, intended for AI readers. Adoption is voluntary and unstandardised; it costs almost nothing and gives a model a clean statement of identity.
robots.txt #
The file controlling which crawlers may access a site. For AI visibility the distinction that matters is between training crawlers, search crawlers and user-fetch agents: blocking a search crawler makes a business uncitable in live answers.
AI crawler #
A bot collecting web content for an AI system, doing one of three jobs: gathering training data, building a search index queried at answer time, or fetching a page on demand during a conversation.
Zero-click #
A query answered without the user visiting any source. Zero-click behaviour is why AI visibility cannot be measured by referral traffic alone: the influence happens whether or not a click follows.
Dark traffic #
Visits and enquiries produced by an AI recommendation that arrive with no attributable referrer, because the user searched your name directly afterwards. Dark traffic is the main reason AI visibility looks like it produces nothing in analytics.
Answer engine #
A system returning a synthesised answer rather than a list of links. Perplexity and Google's AI Mode are answer engines; a general assistant like ChatGPT behaves as one when it retrieves.
AI assistant #
A general-purpose conversational AI. Which assistants matter is a per-market question: ChatGPT and Claude are geofenced out of Hong Kong, while Gemini, Perplexity, DeepSeek and Qwen are available there.
Surface #
The set of assistants a given audience can actually reach. Reporting one blended visibility score across surfaces an audience cannot access describes an audience that is not there.
Prompt set #
The fixed collection of buyer-intent questions a business is measured against, held constant so results are comparable over time. A prompt set is the unit of measurement in AI visibility work.
Search engine #
A system returning ranked links to documents. Distinguished from an answer engine by what it returns — links to be evaluated by the user, rather than a synthesised answer.
SimplyAI (2026). The AI Search Glossary. Updated 22 August 2026. https://simplyai.work/learn/ai-search-glossary/ — individual terms are addressable by anchor, e.g. #share-of-voice.