GEO glossary — every AI-visibility term, defined precisely

GEO (Generative Engine Optimization): The discipline of making content likely to be surfaced, quoted, and cited by generative AI systems — ChatGPT, Claude, Gemini, Perplexity, Google's AI answers.
AEO (Answer Engine Optimization): A near-synonym for GEO that emphasizes direct-answer surfaces. In practice the two are used interchangeably.
AI-SEO: An informal umbrella for GEO/AEO plus the classic SEO that underpins it — because live retrieval runs on search indexes, an unindexed page can't be retrieved.
Parametric memory: What a model learned during training — facts that appeared repeatedly and consistently in its training data. Not directly editable, updated only at the next training cycle. The "slow clock."
Grounding / RAG: Retrieval-Augmented Generation: at query time the engine retrieves passages from the web and answers from them, usually with citations. Updatable in hours to days. The "fast clock."
Grounded vs. parametric answer: An answer built from retrieved sources (usually cited) versus one produced from training memory. They measure different channels and must never be averaged together.
Passage / chunk: The unit of retrieval — a self-contained span of text, not a whole page. The thing that earns a citation is a paragraph that makes sense alone.
Chunkability: How well content survives being split into self-contained retrievable passages: one idea per section, no orphan pronouns, the answer stated first.
Entity: A thing a model reasons about — a product, company, concept, or person — with typed relationships to other things (category, maker, alternatives). LLMs reason over entities, not keyword strings.
Knowledge graph: Structured entity-and-relationship data (Wikidata, Google Knowledge Graph) that feeds both model training and live grounding. The base layer of entity visibility.
sameAs: The schema.org property declaring that several URLs refer to the same entity — the literal machine instruction "these are all me." Cheap, durable, frequently omitted.
Product brief (single source of truth): The one approved fact sheet a subject's content is generated from — canonical name, category, definition, specs with units, aliases. Contradictions anywhere downstream of it are the #1 cause of AI hedging and hallucination.
Hub-and-spoke: Site architecture where one canonical page (the hub) is supported by focused pages (spokes), each owning one question, linked both ways with descriptive anchors.
Answer-first: Putting the direct answer in the first sentence under a question-shaped heading, then elaborating. The single strongest "cite me" signal a page can send.
Presence / mention rate: The share of test prompts where the subject is named in the answer. Headline mention rate is measured on category prompts that never name the brand — otherwise you're grading your own exam.
Share of voice: Your mentions as a fraction of all options named on category prompts — visibility relative to the rivals engines actually name.
Owned-citation rate: The share of answers citing one of your own URLs — as opposed to answers that describe you correctly but send authority to someone else's page.
Prompt suite: A fixed, versioned set of test prompts re-run on a schedule to measure visibility. Versioned because changing the suite invalidates every historical comparison.
Judge (LLM-as-judge): A model used to score answers ("was the brand mentioned? recommended?"). A judge has its own error rate, so it requires calibration against human-labeled samples to be trustworthy.
Wilson interval: A confidence interval for proportions that behaves well at extreme rates and small samples — unlike the naive (Wald) interval, which fails exactly where visibility rates live.
Cluster bootstrap: Resampling at the prompt level rather than the response level when computing confidence intervals — because repetitions of the same prompt are correlated, and treating them as independent understates uncertainty.
GEO lint: An automated pre-publish check that scores a page against the GEO ruleset and blocks publishing on hard errors — the checklist as an enforced contract instead of a hope.
AI Overviews / AI Mode: Google's generative answer surfaces. As of mid-2026, AI Mode is the default search experience — one of the specifics that changes fast, so treat any dated claim about it as perishable.
Search agent vs. training crawler: Two different kinds of bot, two different decisions: search/answer agents fetch pages to build the answer a user sees now (blocking them removes you from AI answers); training crawlers collect corpora for future models (blocking them is a legitimate IP choice). Conflating the two is the expensive robots.txt mistake.
llms.txt: A proposed root Markdown file listing your key pages for LLMs. Not broadly adopted as a ranking input — cheap insurance, never a lever. Generate yours →
Cannibalization: When a supporting page out-ranks or out-cites the page that should be canonical — the usual result of fragmenting one subject across many thin pages.
Zero-click influence: The buying influence that happens inside the AI answer, with no visit for analytics to count. Most citations never become a click, which is why presence-in-answers — not referral traffic — is the honest instrument.

The method behind the vocabulary: the GEO Field Guide · why it is its own discipline: why GEO · how we measure it: the methodology · everything else we publish: resources

The GEO glossary