This glossary defines 69 terms used across generative engine optimization, retrieval systems, and AI search. It is a reference, not a pitch: every entry states what a term means, why it is worth knowing, and a concrete example. Nothing here promises a ranking outcome.
Read it in whatever order is useful. If you arrived from the pillar page or the FAQ chasing a specific concept, jump straight to it with the A–Z index below. If you are building vocabulary from scratch, the eight category sections run roughly from theory to practice: what the field is called, how the underlying models work, how they retrieve information, how they represent entities, how content gets structured, how citation and authority accrue, how any of it gets measured, and finally the infrastructure that gates access to all of it.
One honest note before the definitions start: this field's vocabulary is not settled. GEO, AEO, LLMO, AI SEO and GAIO are used by different vendors to mean roughly the same discipline, and at least one vendor argues publicly that "GEO" is the wrong name. Wikipedia lists several of these as related terms without picking a winner. This glossary uses GEO as its anchor term because it is the term the founding academic paper used, and it defines the alternatives honestly rather than pretending they don't exist. See Foundations below, and the fuller discussion on the pillar page.
A–Z index
AEO · AI Mode · AI Overview · AI Referral Traffic · AI SEO · AI Visibility Score · Answer Engine · Answer Engine Optimization · Answer-First Writing · Assisted Conversion · Brand Mention · Canonical Description · Chunking · Citable Passage · Citation · Citation Rate · Citation Velocity · ClaudeBot · Claude-SearchBot · Content Architecture · Context Window · Crawl vs. Index vs. Retrieve · Cross-Web Consensus · Dead Citation · Digital PR · E-E-A-T · Embedding · Entity · Entity Disambiguation · Entity Salience · Extractability · Freshness · GAIO · Generative Engine · GEO · Google-Extended · GPTBot · Grounding · Hallucination · IndexNow · Internal Link Graph · JSON-LD · Knowledge Cutoff · Knowledge Graph · Knowledge Panel · LLM · LLMO · llms.txt · Mention · Mention Sentiment · Named Author Entity · NAP Consistency · nosnippet / data-nosnippet / max-snippet · OAI-SearchBot · Owned / Verified / Earned Citation Tiers · Parametric Memory · Passage Extraction · PerplexityBot · Pillar Page · Prompt Coverage · Prompt Set · Query Fan-Out · Query Rewriting · RAG · Re-ranking · Retrieval · robots.txt · sameAs · Schema.org · Semantic Cluster · Semantic Similarity · SEO · Server-Side Rendering · Share of Answer · Source Attribution · Structured Data · Temperature / Non-determinism · Tokenization · Topical Authority · Training Data · User Agent · Vector Search · Wikidata / QID · Zero-Click
Foundations
The words for this discipline, and the two Google surfaces most people mean when they say "AI search."
GEO
Definition. Generative Engine Optimization: the practice of influencing whether and how a company appears inside AI-generated answers, and the academic term for the discipline this hub covers.
Why it matters. It is the term with a peer-reviewed origin, which makes it the most defensible anchor when vendors disagree on naming. Everything else in this glossary sits underneath it.
Example. GEO: Generative Engine Optimization (arXiv:2311.09735), presented at KDD 2024 by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, introduced both the name and the first benchmark, GEO-bench, for measuring it.
Related: SEO · AEO, LLMO, AI SEO & GAIO · Share of Answer
Read more: What is GEO?
SEO
Definition. Search Engine Optimization: the discipline of improving a page's visibility and ranking in traditional, link-based search results.
Why it matters. GEO does not replace SEO; it shares SEO's eligibility layer (indexation, crawlability, technical health) and adds an entity and consensus layer on top. A site with no SEO foundation has no GEO foundation either.
Example. A page ranking #3 for "commercial litigation lawyer Toronto" satisfies an SEO goal. Whether that same firm is named when someone asks an AI assistant "who handles construction disputes in Toronto" is a separate, GEO-shaped question.
Related: GEO · Structured Data · Crawl vs. Index vs. Retrieve
Read more: SEO versus GEO
Answer Engine & Generative Engine
Definition. An answer engine returns a direct answer to a question rather than a list of links (Perplexity is the clearest example). A generative engine is any system that produces a synthesized response by generating text from a model, which may or may not involve live retrieval.
Why it matters. The two terms overlap but are not identical: a generative engine can answer purely from parametric memory with no retrieval at all, while an answer engine implies retrieval happened. Knowing which one you are optimizing for changes whether the fix is about training-time presence or live retrievability.
Example. ChatGPT answering "what year was your knowledge cutoff" from memory is acting as a generative engine with no retrieval. ChatGPT answering "who are the top GEO agencies in Canada" by running live searches is acting as an answer engine.
Related: Parametric Memory · Retrieval · GEO
Read more: How AI discovers your company
AI Overview & AI Mode
Definition. AI Overviews are the AI-generated summary blocks that appear above or within standard Google Search results. AI Mode is a separate, more conversational Google Search experience that runs multi-step, multi-query research before answering.
Why it matters. Both draw on the same underlying eligibility gate as ordinary Search results, and both may use query fan-out, but AI Mode goes further, planning and adjusting its search strategy as it works. Confusing them with each other, or with a general "Google AI," leads to the wrong diagnosis when visibility drops.
Example. Google states AI Mode is "issuing multiple related searches concurrently across subtopics and multiple data sources" and that it "makes a plan, conduct[s] searches to find information and adjust[s] the plan based on what it finds." AI Overviews appear in roughly 10% of US search results, per the Reuters Institute's Trends and Predictions 2026.
Related: Query Fan-Out · Structured Data · Crawl vs. Index vs. Retrieve
Read more: Where AI visibility happens
AEO, LLMO, AI SEO & GAIO
Definition. Four alternative names circulating for roughly the same discipline as GEO: AEO (Answer Engine Optimization), LLMO (Large Language Model Optimization), AI SEO, and GAIO (Generative AI Optimization). None has displaced the others.
Why it matters. There is no industry consensus, and pretending otherwise misleads buyers who are comparison-shopping vendors using different vocabulary for the same service. Profound, a GEO measurement vendor, publicly argues "GEO" is a poor label because it collides with geo-targeting, and pushes "AEO" instead. Other vendors use the terms interchangeably. Wikipedia lists them as related without differentiating.
Example. A company searching for "AEO agency" and a company searching for "GEO agency" may end up talking to the same vendor about the same work. This glossary uses GEO throughout because it traces to the founding academic paper, and flags the alternatives so readers are not confused by vendor terminology.
Related: GEO · Answer Engine & Generative Engine · Share of Answer
Read more: What is GEO?
How models work
The mechanics inside a large language model that determine what it knows, what it can look up, and how reliably it says either.
flowchart LR
A["Training Data"] --> B["Tokenization and Embedding"]
B --> C["LLM (trained weights)"]
C --> D["Parametric Memory"]
C --> E["Context Window"]
D --> F{"Answer grounded?"}
E --> F
F -->|"No"| G["Hallucination risk"]
F -->|"Yes"| H["Grounding"]
Training builds the model's parametric memory; the context window is what it can additionally read at answer time. Grounding is what keeps the gap between the two from becoming a hallucination.
LLM
Definition. A large language model: a neural network trained on massive text corpora to predict and generate language, forming the base of ChatGPT, Gemini, Claude and similar systems.
Why it matters. Every downstream GEO mechanic — retrieval, grounding, citation — is a system built around an LLM's core behaviour, which is statistical text prediction, not database lookup. Understanding that distinction explains why the same prompt can produce different answers on different days.
Example. GPT-4-class and Claude-class models are LLMs; the chat products (ChatGPT, Claude.ai) are applications that wrap an LLM with tools like web search, memory and citation formatting.
Related: Parametric Memory · Context Window · Hallucination
Read more: How AI discovers your company
Parametric Memory
Definition. What a model "knows" purely from its training weights, with no live lookup — knowledge baked in at training time rather than fetched at answer time.
Why it matters. You can only influence parametric memory by existing prominently and consistently in whatever got scraped for training, which happens on a training schedule you do not control and over a horizon measured in years, not weeks. It is why well-established brands get named in generic prompts with zero retrieval, and why a two-year-old company almost never does.
Example. A model naming "IBM" or "Salesforce" unprompted when asked "name an enterprise software company" is drawing on parametric memory. A model naming a five-person regional firm for the same prompt would almost certainly require retrieval.
Related: Training Data · Knowledge Cutoff · Retrieval
Read more: How AI discovers your company
Training Data
Definition. The corpus of text (and increasingly images, audio and code) a model is trained on, which shapes its parametric memory.
Why it matters. Whether your content is included in future training runs is governed by robots.txt tokens like GPTBot and Google-Extended, and that decision is entirely separate from whether you're retrievable in live search. Blocking training crawlers has zero effect on today's search visibility.
Example. A publisher can block GPTBot to keep its articles out of OpenAI's next training run while leaving OAI-SearchBot unblocked, so it still appears in live ChatGPT search results.
Related: GPTBot · ClaudeBot · Google-Extended
Read more: The training decision versus the retrieval decision
Knowledge Cutoff
Definition. The date after which a model's training data stops; the model has no parametric knowledge of events, products, or content published after that point.
Why it matters. Anything that happened after a model's cutoff is invisible to it unless retrieval supplies it live. This is the single biggest argument for making retrieval work: it is the only channel through which post-cutoff facts about your company reach the model at all.
Example. A company founded in 2025 will not exist in the parametric memory of a model with a 2024 training cutoff. It can only be named in an answer if the model retrieves a live page about it.
Related: Parametric Memory · Retrieval · Freshness
Read more: Parametric memory versus retrieval
Tokenization & Embedding
Definition. Tokenization splits text into the sub-word units a model actually processes. Embedding converts tokens, words, or passages into numerical vectors that represent meaning in a way a machine can compare.
Why it matters. Retrieval systems don't search for your exact words; they search for vectors near your meaning. A page can rank for "affordable enterprise CRM" without using that phrase verbatim, because embeddings capture semantic proximity, not string matching.
Example. The phrase "budget-friendly customer relationship platform" and the phrase "affordable enterprise CRM" produce embeddings close enough in vector space that a semantic search for one plausibly retrieves a page written with the other.
Related: Vector Search & Semantic Similarity · Chunking & Passage Extraction · RAG
Read more: What GEO actually optimizes
Vector Search & Semantic Similarity
Definition. Vector search finds documents whose embeddings are mathematically close to a query's embedding, rather than matching literal keywords. Semantic similarity is the property vector search is measuring: how close two pieces of text are in meaning, not spelling.
Why it matters. This is why keyword-stuffing does not work in a generative context and can actively hurt, as the GEO paper found. Writing the concept clearly, in natural language, retrieves better than repeating a target phrase.
Example. A query for "how to reduce churn in SaaS" can retrieve a page titled "Cutting customer attrition in subscription software" even though not one word matches, because the two phrases sit close together in vector space.
Related: Tokenization & Embedding · Retrieval · Re-ranking
Read more: Content tactics that measurably helped or hurt
Hallucination
Definition. A confident, fluent statement generated by a model that is factually false, unsupported by any source, or invented outright.
Why it matters. Hallucination risk rises when a model lacks grounded retrieval and has to fill a gap from parametric memory, which is exactly the situation an obscure or ambiguous entity creates. A company with thin, inconsistent web presence is more likely to be described inaccurately, or invented into, by a model with nothing solid to retrieve.
Example. A model asked about a small company's founding date, with no reliable page stating it clearly, may generate a plausible-sounding but wrong year rather than declining to answer.
Related: Grounding · Entity Disambiguation · Canonical Description
Read more: Entity recognition
Grounding
Definition. The practice of anchoring a generated answer to specific retrieved sources, so the response is traceable to real, citable text rather than pure model recall.
Why it matters. Grounded answers carry citations; ungrounded ones don't, and are more prone to hallucination. Every major engine's citation mechanism — Gemini's url_citation annotations, Claude's web_search_result_location objects — exists specifically to make grounding auditable.
Example. Anthropic documents that Claude's web search always returns citations as web_search_result_location objects containing url, title, and cited_text — a literal span lifted from the source page, which is grounding made explicit.
Related: Hallucination · Citation · Source Attribution
Read more: Where AI visibility happens: Claude
Temperature & Non-determinism
Definition. Temperature is a setting that controls how much randomness a model injects when choosing its next word. Non-determinism is the resulting behaviour: the same prompt can produce different answers across sessions, users, and model versions, even at low temperature.
Why it matters. A single check of "does my brand appear" is an anecdote, not a measurement. This is the reason multi-sample measurement methods exist and why any GEO report claiming precision from one run should be discounted.
Example. Evertune's measurement approach samples each prompt 100 times across 11 models specifically to capture this variance rather than treating one response as representative.
Related: Share of Answer · Prompt Set · AI Visibility Score
Read more: Measuring AI visibility
Context Window
Definition. The maximum amount of text — measured in tokens — a model can consider at once for a given request, including the prompt, any retrieved documents, and the conversation history.
Why it matters. A page that gets retrieved but is too long, cluttered, or buried in navigation may lose the passage that actually answers the question once it competes for space in a finite context window. Front-loading the answer matters partly for this reason.
Example. If five retrieved documents are stuffed into a context window and a company's key fact sits in paragraph fourteen of a 2,000-word article, it competes for attention against four other full documents rather than standing out.
Related: Answer-First Writing · Chunking & Passage Extraction · Extractability & Citable Passage
Read more: Content architecture
Retrieval
How a generative system decides a live lookup is needed, runs it, and turns raw pages into the passages that end up in an answer.
flowchart TD
A["Buyer prompt"] --> B["Query Fan-Out"]
B --> C["Query Rewriting"]
C --> D["Crawl and Index (prerequisite)"]
D --> E["Retrieval"]
E --> F["Chunking and Passage Extraction"]
F --> G["Re-ranking"]
G --> H["RAG: synthesized, cited answer"]
One prompt becomes several searches, several searches become candidate passages, and only the best-ranked passages survive into the final answer.
RAG
Definition. Retrieval-Augmented Generation: an architecture where a model first retrieves relevant documents from an external source, then generates its answer using both the retrieved text and its own parametric knowledge.
Why it matters. RAG is the mechanism behind almost every cited AI answer — AI Overviews, ChatGPT search, Perplexity, Gemini grounding, Claude's web search. If your content isn't retrievable, it structurally cannot enter the "R" half of RAG, no matter how well-written it is.
Example. A user asks Perplexity a question; Perplexity retrieves several web pages in real time, feeds relevant passages into the model alongside the question, and generates an answer with inline citations back to those pages. That is RAG in production.
Related: Retrieval · Grounding · Re-ranking
Read more: How AI discovers your company
Retrieval
Definition. The step where a system runs a live search and pulls candidate documents, as distinct from what the model already knows from training.
Why it matters. Retrieval is the mechanism you can actually influence this quarter, unlike parametric memory, which moves on a training schedule you don't control. Access failures — blocked crawlers, JavaScript-only rendering, broken links — kill retrieval before content quality even gets evaluated.
Example. Google documents that AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources" — that entire process is retrieval.
Related: Crawl vs. Index vs. Retrieve · Query Fan-Out · Freshness
Read more: How AI discovers your company
Query Fan-Out
Definition. The technique of decomposing one user question into several related sub-queries run concurrently across subtopics and data sources, then combining the results.
Why it matters. The engine is not searching for the buyer's literal question. It is searching for four or five sub-questions it invented. Optimizing content for the literal prompt is the wrong target; optimizing for the decomposition is the right one.
Example. A prompt like "we're a 40-person medical device company in Ontario, who should we hire to fix our AI search visibility" might fan out into "generative engine optimization agency Canada," "AI visibility consultant medical device industry," and "GEO vs SEO agency Ontario" as separate searches.
Related: Query Rewriting · AI Overview & AI Mode · Retrieval
Read more: Query fan-out, made concrete
Query Rewriting
Definition. The process of transforming a user's original natural-language question into one or more targeted search queries better suited to a search index.
Why it matters. ChatGPT search specifically rewrites conversational input before it ever reaches its search partners. A page written to match the buyer's exact phrasing may miss entirely if the rewritten query uses different terms.
Example. OpenAI documents that ChatGPT search "rewrites your query into one or more targeted queries" sent to search partners including Bing and Shopify before returning an answer.
Related: Query Fan-Out · Vector Search & Semantic Similarity · Retrieval
Read more: ChatGPT
Chunking & Passage Extraction
Definition. Chunking splits a document into smaller sections before it's indexed for retrieval. Passage extraction is the downstream step of pulling out the specific chunk or sentence that answers a given question, which is what actually gets quoted or cited.
Why it matters. A model does not read your whole page and decide to trust it. It surfaces one chunk. If the answer to a real question is spread across three paragraphs with no single self-contained passage, none of those paragraphs is a good extraction candidate.
Example. A 1,500-word article on "employee retention strategies" that never states a single self-contained sentence like "Structured onboarding reduces first-year turnover by giving new hires a defined 90-day plan" gives passage extraction nothing clean to lift.
Related: Answer-First Writing · Extractability & Citable Passage · Context Window
Read more: Common GEO mistakes: burying the answer
Re-ranking
Definition. The step where a system scores and reorders an initial set of retrieved candidates by relevance and quality before deciding what makes it into the final answer.
Why it matters. Being retrieved is not the same as being used. A page can be pulled into a candidate pool of a dozen documents and still lose the re-ranking step to a more clearly written competitor passage.
Example. Ten pages might all get retrieved for "GEO agency Canada," but re-ranking might promote only three into the answer, based on factors like clarity, specificity and corroboration rather than raw keyword match.
Related: Vector Search & Semantic Similarity · RAG · Share of Answer
Read more: What GEO actually optimizes
Crawl vs. Index vs. Retrieve
Definition. Three distinct steps often collapsed into one: crawling is a bot fetching your page; indexing is a search engine storing and organizing that page for future matching; retrieval is a live system pulling the indexed page in response to a specific query.
Why it matters. A page can be crawled but not indexed, indexed but never retrieved for the queries you care about, or retrieved but re-ranked out of the final answer. Diagnosing "we're invisible in AI answers" requires knowing which of the three steps is actually failing.
Example. Google's eligibility gate for AI Overviews and AI Mode requires a page to be "indexed and eligible to be shown in Google Search with a snippet" — that's steps one and two. Whether it then gets retrieved for a specific fan-out sub-query is a separate, later question.
Related: Retrieval · robots.txt · Server-Side Rendering
Read more: Retrieval failures beyond robots.txt
Freshness
Definition. How recently a piece of content was published or updated, and how that recency factors into whether a retrieval system surfaces it for time-sensitive queries.
Why it matters. Freshness is one of the few levers that works on a fast, predictable schedule compared to consensus-building, which takes months. It's also one of this hub's differentiators against most competing GEO resources, which cite pre-2026 guidance as current.
Example. FAQPage rich results stopped appearing in Google Search as of May 7, 2026. A GEO guide published after that date still recommending FAQ schema as a citation tactic is working from stale information, and freshness is exactly the signal that exposes it.
Related: Dead Citation · Knowledge Cutoff · Retrieval
Read more: Common GEO myths: FAQ schema
Entities & knowledge
How a machine identifies who you are, distinguishes you from everyone with a similar name, and encodes what it believes about you.
flowchart TD
A["Entity"] --> B["Entity Disambiguation"]
B --> C["sameAs links"]
B --> D["Wikidata and QID"]
C --> E["Knowledge Graph and Knowledge Panel"]
D --> E
F["NAP Consistency"] --> B
G["Canonical Description"] --> B
E --> H["Entity Salience"]
H --> I["Topical Authority and Semantic Cluster"]
Disambiguation is the hinge: consistent identity signals feed it, and salience and authority are only earned once it succeeds.
Entity
Definition. A distinct, identifiable thing a machine can reason about — a company, a person, a product — as opposed to a mere string of characters that happens to appear on a page.
Why it matters. Before a system can recommend you, it has to know you exist as a distinct thing, separable from every other similarly-named thing. This is the foundational data problem underneath almost everything else in GEO.
Example. "Artlogic" the marketing firm and any other unrelated business that happens to share the name are, to a model doing entity resolution, two candidate entities competing for the same string until disambiguating signals sort them out.
Related: Entity Disambiguation · Knowledge Graph & Knowledge Panel · Canonical Description
Read more: Entity recognition
Entity Disambiguation
Definition. The process by which a system determines which specific real-world entity a name, mention, or reference actually refers to, when multiple candidates share the same or a similar name.
Why it matters. This fails constantly and silently. Three law firms sharing a founder's surname, a software company sharing its name with an unrelated manufacturer, a clinic that rebranded while half the web still uses the old name — in each case the model's internal representation of "you" is smeared across overlapping concepts, and a better-corroborated competitor wins the ambiguity by default.
Example. A structured Organization entry with consistent naming, a stable legal-suffix convention, and sameAs links to every verified profile gives a disambiguation system the signals it needs to resolve "Acme Health" to one specific clinic rather than guessing among three.
Related: sameAs · Canonical Description · NAP Consistency
Read more: Entity recognition
Knowledge Graph & Knowledge Panel
Definition. A knowledge graph is a structured database of entities and the verified relationships between them, maintained by a platform like Google. A knowledge panel is the visible summary box a knowledge graph generates on a search results page for a recognized entity.
Why it matters. Google's AI Mode explicitly draws on the Knowledge Graph alongside real-world information and shopping data when constructing answers. A company with no knowledge graph presence is invisible to that entire data source, independent of how good its web content is.
Example. Google's own AI Mode documentation states the system draws on "fresh, real-time sources like the Knowledge Graph, info about the real world, and shopping data." A knowledge panel appearing for a company's name in a Google search is a visible sign that the entity has been resolved into the graph.
Related: Entity · sameAs · Wikidata & QID
Read more: How AI discovers your company: structured knowledge
sameAs
Definition. A schema.org property that links an entity's structured data record to other verified pages about the same entity elsewhere on the web.
Why it matters. It is one of the clearest, cheapest disambiguation signals a company can ship, directly connecting a homepage's Organization markup to LinkedIn, Crunchbase, Wikidata, and other verified profiles.
Example. Google's documentation describes sameAs as "the URL of a page on another website with additional information about your organization," and confirms multiple URLs can be listed, such as "sameAs": ["https://www.linkedin.com/company/example", "https://www.crunchbase.com/organization/example"].
Related: Entity Disambiguation · Structured Data · Wikidata & QID
Read more: Entity recognition
Wikidata & QID
Definition. Wikidata is a free, structured knowledge base that assigns every entity a unique identifier called a QID (for example, Q312 for Apple Inc.). It's a common backbone data source for knowledge graphs across the web.
Why it matters. A Wikidata item, when genuinely warranted by notability, is a strong disambiguation and corroboration signal that multiple downstream knowledge graphs can consume. Manufacturing one for a company that doesn't meet Wikidata's notability standards backfires: both Wikidata and Wikipedia actively remove promotional entries, and a deleted entry is worse than never having one.
Example. A well-established, independently-covered regional company might legitimately qualify for a Wikidata item with its own QID; a two-year-old startup with no independent press coverage almost certainly does not, and attempting one risks a deletion that becomes its own negative signal.
Related: Knowledge Graph & Knowledge Panel · sameAs · Entity Disambiguation
Read more: Entity recognition
NAP Consistency
Definition. Consistent Name, Address and Phone data for a business across every directory, profile and citation source that lists it.
Why it matters. NAP inconsistency is one of the most common, most fixable causes of entity confusion. A model resolving conflicting address or name data across sources either hedges, picks one source arbitrarily, or drops the entity to avoid asserting something unsupported.
Example. A company listed as "Acme Corp." on its website, "Acme Corporation Inc." on a business directory, and an outdated address on a third profile gives a consensus-seeking system three slightly different facts to reconcile instead of one clear one.
Related: Canonical Description · Entity Disambiguation · Cross-Web Consensus
Read more: Entity recognition
Canonical Description
Definition. A single, unambiguous one-sentence description of what a company does, repeated near-verbatim across the website, LinkedIn, Crunchbase, industry directories and press materials.
Why it matters. Models are consensus machines. If eleven sources describe a company the same way, that description becomes a retrievable fact. If eleven sources describe it eleven different ways, there is no fact to retrieve. This is often the single highest-yield hour in an entire GEO engagement.
Example. Standardizing on one line — "Artlogic is a growth and marketing firm helping companies expand across Canada, the US, Europe and the Middle East" — and pasting it, unedited, into every bio, directory listing and about page removes the ambiguity a model would otherwise have to guess through.
Related: Entity Disambiguation · NAP Consistency · Cross-Web Consensus
Read more: Entity recognition
Entity Salience
Definition. How prominent and central an entity is within a given piece of content, as distinct from merely being mentioned in passing.
Why it matters. A brand named once in a footer credit line has near-zero salience compared to a brand that is the clear subject of a paragraph. Retrieval and re-ranking systems weight salient mentions far more heavily than incidental ones.
Example. An article that opens with "Acme Health's diabetes management program reduced average A1C by two points across its patient cohort" gives Acme Health high salience in that passage. The same fact buried in a sentence three paragraphs deep, listed alongside five competitors, has much lower salience.
Related: Entity · Topical Authority & Semantic Cluster · Answer-First Writing
Read more: Content architecture
Topical Authority & Semantic Cluster
Definition. Topical authority is the degree to which a source is treated as a credible, comprehensive reference on a specific subject. A semantic cluster is the set of interlinked pages covering related subtopics that, together, build that authority for a broader topic.
Why it matters. A single strong article rarely earns topical authority alone; a coherent cluster of pages that collectively answer the sub-questions a topic generates does. This is directly related to query fan-out: an engine decomposing a prompt into five sub-queries is more likely to find a credible answer from a source that has content for all five than from one that has content for only one.
Example. This hub's five pages — the pillar, methodology, glossary, FAQ, and resources — form a semantic cluster around "generative engine optimization," each addressing a different facet the topic's fan-out would plausibly generate.
Related: Content Architecture & Internal Link Graph · Pillar Page · Query Fan-Out
Read more: The GEO stack
Content & structure
How to build pages that are structured for machine extraction as well as human reading, and what schema does and does not do for that.
Structured Data
Definition. Machine-readable markup added to a page, usually via schema.org vocabulary, that explicitly labels what a piece of content is — an organization, a product, an article, a review.
Why it matters. Google's own AI-features documentation is direct on this: "There's also no special schema.org structured data that you need to add" to appear in AI Overviews or AI Mode. That is not a statement that schema is worthless. It still serves disambiguation (linking your entity to other verified records via sameAs), fact confirmation (helping a system verify what your page's visible text already says), and other non-Google consumers that do read it. What it does not do is flip an AI-visibility switch that does not exist.
Why it matters, continued. The one rule that actually protects you: whatever structured data you ship must match the visible text on the page. Google explicitly recommends this, and mismatched schema is a trust signal working against you, not for you.
Example. Organization schema with a sameAs array pointing to LinkedIn, Crunchbase and Wikidata helps disambiguate an entity. It does not, by itself, earn a citation in an AI Overview.
Related: Schema.org · JSON-LD · sameAs
Read more: What GEO is not
Schema.org
Definition. A shared vocabulary of types and properties (Organization, Article, Product, Review, and hundreds more) maintained collaboratively by Google, Microsoft, Yahoo and Yandex, used to write structured data.
Why it matters. Schema.org is the vocabulary; structured data is the practice of using it; JSON-LD is the format it's most commonly written in. Knowing the difference matters when a vendor pitches "AI schema" as if it were a distinct, proprietary thing — it usually isn't.
Example. Organization, Article, Product and BreadcrumbList are all schema.org types; a company chooses which ones apply to which pages.
Related: Structured Data · JSON-LD · sameAs
Read more: What GEO is not
JSON-LD
Definition. JavaScript Object Notation for Linked Data: the format Google recommends for embedding schema.org structured data inside a web page's HTML.
Why it matters. It's an implementation detail, not a strategy. Getting the syntax right (valid JSON, correct nesting, matching visible text) matters; obsessing over which exact schema types to add without a clear entity or answer-quality problem to solve does not.
Example.
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Example Co.",
"sameAs": ["https://www.linkedin.com/company/example"]
}
Related: Schema.org · Structured Data · sameAs
Read more: What GEO is not
Answer-First Writing
Definition. Structuring a page so the direct answer to its core question appears in the first sentences, with elaboration, nuance and caveats following after.
Why it matters. A 2,000-word article that answers the question in paragraph fourteen is not extractable in any practical sense. Retrieval systems favour self-contained passages, and the earliest sentences of a page are the most likely to be pulled cleanly into a synthesized answer.
Example. Opening a page with "FAQPage rich results stopped appearing in Google Search as of May 7, 2026" instead of three paragraphs of throat-clearing before that fact gives a retrieval system an immediately extractable, quotable answer.
Related: Extractability & Citable Passage · Chunking & Passage Extraction · Context Window
Read more: Common GEO mistakes: burying the answer
Extractability & Citable Passage
Definition. Extractability is how easily a passage can be lifted out of its surrounding context and still make complete sense on its own. A citable passage is a passage that succeeds at this: self-contained, specific, and quotable without surrounding sentences.
Why it matters. The GEO paper found that adding quotations from expert sources produced the largest single gain in visibility of any tactic tested, around 40%. Adding statistics produced roughly a 30% gain. Both work for the same underlying reason: a sentence with a number or a named attribution in it reads cleanly in isolation, which is exactly what a passage needs to survive extraction.
Example. "Otterly.ai found that 19.3% of over 20 million cited URLs across seven AI engines were dead in a one-month snapshot" is a citable passage — it stands alone with its source, its number, and its claim intact. "This can be a real problem for a lot of companies" is not.
Related: Answer-First Writing · Chunking & Passage Extraction · Source Attribution
Read more: Content tactics that measurably helped or hurt
Content Architecture & Internal Link Graph
Definition. Content architecture is how a site's pages are organized and interrelated to cover a topic comprehensively. The internal link graph is the specific network of links between those pages that makes the relationships machine-traceable.
Why it matters. A single strong page rarely wins a fan-out-decomposed query on its own. A coherent architecture, with a clear internal link graph connecting a pillar page to its supporting pages, signals topical depth to both crawlers and readers, and gives a retrieval system multiple entry points into the same subject.
Example. This hub's structure — one pillar page linking out to methodology, glossary, FAQ, resources and case studies, each linking back — is a deliberate content architecture built around one internal link graph.
Related: Pillar Page · Topical Authority & Semantic Cluster · Crawl vs. Index vs. Retrieve
Read more: The GEO stack
Pillar Page
Definition. A comprehensive, central page on a topic that anchors a cluster of more specific supporting pages, all interlinked.
Why it matters. Pillar pages give both humans and retrieval systems a single authoritative starting point for a broad topic, while the supporting pages absorb the long-tail, fan-out-shaped sub-questions the pillar can't cover in depth without becoming unreadable.
Example. This hub's own pillar page covers GEO broadly, while this glossary, the methodology, the FAQ, the resource library and the case studies each go deep on one facet.
Related: Content Architecture & Internal Link Graph · Topical Authority & Semantic Cluster · Answer-First Writing
Read more: The Artlogic GEO Knowledge Hub
E-E-A-T
Definition. Experience, Expertise, Authoritativeness and Trustworthiness: a framework from Google's Search Quality Rater Guidelines used to evaluate content and author credibility, now widely applied to how generative systems weight sources too.
Why it matters. E-E-A-T is a human evaluation framework, not a piece of markup you can install, and no schema tag confers it. It's earned through demonstrable first-hand experience, verifiable credentials, and a track record other sources corroborate.
Example. A legal article credited to a named lawyer with a verifiable bar admission and a linked professional history demonstrates E-E-A-T in a way an anonymous "Staff Writer" byline on the identical text does not, regardless of how well either page is optimized technically.
Related: Named Author Entity · Cross-Web Consensus · Digital PR
Read more: Authority
Named Author Entity
Definition. Content attributed to a specific, real, verifiable person with a professional footprint, rather than to a generic byline or no byline at all.
Why it matters. An author page that links to a LinkedIn profile, a bar admission, a licence number, or a publication record does entity work and trust work simultaneously: it gives the model a verifiable person to corroborate, not just a company.
Example. A construction-law article byline reading "Written by Jane Moreau, LL.B., admitted to the Law Society of Ontario, with a linked professional profile" is a named author entity. "Written by the Editorial Team" is not.
Related: E-E-A-T · Entity Disambiguation · Digital PR
Read more: Authority
Citations & authority
What counts as an engine acknowledging you, how those acknowledgements are tiered, and why they decay if left unmaintained.
Citation
Definition. An instance of an AI engine linking directly to a source in a generated answer.
Why it matters. A citation is a fundamentally different, more valuable asset than being merely named without a link, because it carries a clickable path back to your domain and is directly measurable as a rate. Different engines cite at very different rates and in different formats, so citation behaviour has to be tracked per engine, not as one aggregate number.
Example. Gemini's grounding returns inline url_citation annotations with character offsets pointing back into the generated text; Claude returns web_search_result_location objects with url, title and cited_text. Both are citations; neither format is interchangeable with the other.
Related: Mention & Brand Mention · Citation Rate · Source Attribution
Read more: Citation
Mention & Brand Mention
Definition. A mention is when an engine names an entity in a generated answer without a link. A brand mention is specifically that entity's company name (as opposed to, say, a person's name or a product name) appearing that way.
Why it matters. Citation and mention are not the same asset and should be tracked separately: a mention with no link still shapes a buyer's shortlist, but it produces no clickable referral and is harder to measure without dedicated tooling.
Example. An AI Overview stating "firms known for construction litigation in Toronto include Acme Law and Beta Legal" without linking either is a brand mention for both, with zero citations attached.
Related: Citation · Mention Sentiment · Share of Answer
Read more: Citation
Cross-Web Consensus
Definition. The degree to which multiple independent sources across the web agree on the same facts about an entity, which generative systems treat as evidence those facts are true.
Why it matters. This is the layer most GEO programmes never reach, and it's where durable advantage lives. Generative systems are, functionally, consensus estimators: when independent sources align, a claim becomes retrievable as fact; when they conflict, the model hedges, defers to the most authoritative source, or omits the entity entirely rather than assert something it cannot support.
Example. A company's own site, its LinkedIn page, Crunchbase, a Wikidata entry, and independent trade-press coverage all describing it the same way builds consensus that a single well-written owned-domain page, however polished, cannot build alone.
Related: Canonical Description · Digital PR · Owned, Verified & Earned Citation Tiers
Read more: Cross-web consensus
Citation Velocity
Definition. The rate at which a company earns new independent citations and mentions across AI engines over time.
Why it matters. A defensible programme builds citation velocity at a sustainable, credible pace rather than in an artificial burst, because a sudden unnatural spike in third-party mentions is a pattern that both search engines and generative systems have reason to treat with suspicion.
Example. A steady cadence of earned trade-press coverage and comparison-page inclusion over a year builds citation velocity that a single paid press-release blast, timed to spike all at once, does not replicate credibly.
Related: Digital PR · Owned, Verified & Earned Citation Tiers · Citation Rate
Read more: Citation and Consensus Building
Owned, Verified & Earned Citation Tiers
Definition. A three-tier framework for citation sources. Owned: your own site and documentation, total control, fast to produce. Verified: profiles like LinkedIn, Crunchbase, Wikidata and business registries, high control, medium speed. Earned: trade press, comparison pages, community discussion and analyst mentions, low control, slow, but the most durable and trusted.
Why it matters. Profound's citation data across 11.84 billion citations found roughly 57% of AI citations globally point to brand-owned domains, ranging from about 47% for ChatGPT to 69% for Gemini. Owned is where most citations land, but owned citations are only trusted because verified and earned sources corroborate the entity behind them. A company with excellent owned content and no third-party footprint reads to a retrieval system as a well-produced claim with nothing behind it.
Example. A knowledge hub like this one is an owned-tier asset; a LinkedIn company page is verified-tier; a mention in an independent industry comparison article is earned-tier. A mature programme deliberately builds across all three rather than over-investing in the easiest one.
Related: Cross-Web Consensus · Digital PR · Citation
Read more: Citation architecture
Digital PR
Definition. The practice of securing independent editorial coverage, expert commentary placements, and third-party mentions across trade press and industry publications.
Why it matters. Digital PR is the primary engine for building the earned tier of citations, which is the slowest and hardest to fake, and therefore the most durable. It also supplies the named-author, third-party corroboration that feeds E-E-A-T and cross-web consensus at the same time.
Example. Securing a named executive as a quoted expert source in a trade publication's article on industry trends produces an earned citation, builds E-E-A-T for that named individual, and reinforces cross-web consensus about the company, all from one placement.
Related: Owned, Verified & Earned Citation Tiers · Cross-Web Consensus · E-E-A-T
Read more: Citation and Consensus Building
Dead Citation
Definition. A citation an AI engine has generated that points to a URL which is now missing, moved, or unreachable.
Why it matters. A citation earned and then broken by a site migration is a citation no longer functioning, silently. This is a real, ongoing maintenance obligation that most GEO programmes never budget for, because link-health auditing has historically belonged to SEO teams, not content teams.
Example. Otterly.ai examined more than 20 million cited URLs across seven AI engines over a one-month window and found 19.3% were dead, with the rate highest on ChatGPT at 25.1% and lowest on Google AI Overviews at 12.6%. This is descriptive vendor data from a single snapshot, not a controlled study, but the practical lesson holds: redirect discipline during any migration is GEO work, not an afterthought.
Related: Freshness · Source Attribution · Owned, Verified & Earned Citation Tiers
Read more: Citation
Source Attribution
Definition. The mechanism by which a generative system credits the specific origin of a fact or passage it used in an answer, typically as a link, a title, and sometimes a literal quoted span.
Why it matters. Attribution formats differ by engine, and the format matters because it determines what's actually clickable and verifiable to the end user. Claude's web_search_result_location objects carry a literal cited_text span; Gemini's url_citation annotations carry character offsets into the generated text; not every engine's attribution is equally granular or equally likely to appear.
Example. Anthropic's terms require that citations be displayed when API outputs are shown directly to end users, which makes source attribution a contractual requirement in Claude's ecosystem, not just a design choice.
Related: Citation · Grounding · Dead Citation
Read more: Claude
Measurement
There is no Search Console for generative answers. These are the metrics that substitute for one, and their honest limitations.
Share of Answer
Definition. Across a fixed set of prompts, the percentage of generated responses that name a given brand.
Why it matters. It's the closest GEO analogue to a rank tracker, but it must be tracked per engine, because variance between engines is large, and it must be sampled multiple times per prompt, because responses are non-deterministic. A single check on a single engine is not a measurement.
Example. Testing 40 buyer-relevant prompts across ChatGPT, Gemini, Claude and Perplexity, each sampled several times, and finding a brand named in, say, 22% of ChatGPT responses but 4% of Perplexity responses for the same prompt set, would be a real share-of-answer baseline. Those figures are hypothetical; the point is the shape of the gap between engines, which in practice is often this wide.
Related: Prompt Set · Citation Rate · Temperature & Non-determinism
Read more: Measuring AI visibility
Prompt Set
Definition. A fixed, repeatable list of prompts representing real buyer questions, used as the consistent basis for measuring share of answer, citation rate, and other GEO metrics over time.
Why it matters. Without a fixed prompt set established before work begins, there is no baseline to compare against, and improvement becomes indistinguishable from noise. This is the most commonly skipped step in the entire measurement discipline.
Example. A 50-prompt set built from mapped buyer questions like "best AI visibility agency for legal firms in Ontario" and its query-fan-out sub-questions, run monthly across four engines, is a usable prompt set. A single ad-hoc ChatGPT check the week before a client call is not.
Related: Share of Answer · Prompt Coverage · Query Fan-Out
Read more: AI Visibility Audit
Prompt Coverage
Definition. The proportion of a company's mapped buyer prompts that it appears for at all, regardless of position or citation status.
Why it matters. A company might have strong share of answer on the five prompts it already tracks while having zero presence on thirty other real buyer questions it never mapped. Prompt coverage exposes that blind spot.
Example. A firm tracking only its core "GEO agency" prompt might discover, once its prompt set expands to cover the fan-out sub-questions buyers actually generate, that its coverage across the full mapped set is far lower than it assumed. Coverage below a third of the mapped set is common on a first baseline, though the figure is entirely company-specific.
Related: Prompt Set · Share of Answer · Query Fan-Out
Read more: AI Visibility Audit
Citation Rate
Definition. The percentage of generated responses across a prompt set that link a company, as distinct from merely naming it.
Why it matters. Citation rate and share of answer answer different questions and can diverge sharply: a brand might be frequently mentioned but rarely linked, which matters enormously for referral traffic even if it matters less for shortlist visibility.
Example. A brand appearing in 30% of responses across a prompt set but only linked in 8% of them has a share of answer of 30% and a citation rate of 8% — two very different numbers describing the same underlying dataset.
Related: Share of Answer · Citation · Dead Citation
Read more: Measuring AI visibility
Mention Sentiment
Definition. The framing and tone of how a brand is described when it is named or cited, distinct from whether it's named at all.
Why it matters. Being named as "a good option for X" and being named as "an alternative to Y" are different commercial outcomes even though both count as a mention. Aggregate mention volume without sentiment analysis can hide a genuinely damaging framing pattern.
Example. A company consistently framed across engines as "a budget option" rather than "a premium provider" has healthy mention volume but a sentiment problem that share of answer alone would never surface.
Related: Mention & Brand Mention · Share of Answer · AI Visibility Score
Read more: Measuring AI visibility
AI Referral Traffic & Zero-Click
Definition. AI referral traffic is site visits attributed to clicks from AI-generated answers. Zero-click is the broader pattern where a query gets fully answered on the results or answer surface itself, producing no click to any source at all.
Why it matters. These two concepts explain the traffic paradox at the heart of this shift: Google organic referrals to news sites fell 33% globally and 38% in the US between November 2024 and 2025, even though search volume did not fall by anywhere near that much. The queries are still happening; fewer of them produce a click. Google still delivers roughly 500 times more referral traffic than ChatGPT from search alone, so the absolute scale gap remains enormous even as the composition shifts.
Example. A buyer asking an AI Overview "what's the average cost of a kitchen renovation in Toronto" and getting a complete numeric answer inline, with no click to any of the sources it drew from, is a zero-click outcome. The same buyer clicking through a citation link in a Perplexity answer to read a full case study is AI referral traffic.
Related: Citation Rate · Assisted Conversion · Share of Answer
Read more: How search actually changed
Assisted Conversion
Definition. A conversion — a lead, a sale, a signed engagement — that followed exposure to an AI-generated answer naming or citing a brand, even when the final conversion happened through a different, unattributed channel like a direct visit or branded search.
Why it matters. Because zero-click answers and hedged AI attribution mean many AI-influenced buying decisions never register as a trackable referral, assisted conversion is often the only honest way to credit the channel at all. Attribution here is genuinely imperfect, and that imperfection should be stated plainly to clients rather than papered over with manufactured precision.
Example. A buyer who sees a company named favourably in a ChatGPT answer, then two days later types the company's name directly into a browser and requests a quote, produces a branded-search conversion that a naive attribution model would credit to "direct," when the AI exposure plausibly did the real work.
Related: AI Referral Traffic & Zero-Click · Share of Answer · AI Visibility Score
Read more: Measuring AI visibility
AI Visibility Score
Definition. A composite, vendor-defined metric intended to summarize a brand's overall presence across AI answer engines into a single number.
Why it matters. No independent third-party audit of any AI-visibility tool's accuracy has been verified publicly, and every vendor scores its own coverage using its own undisclosed methodology. Treat any single composite score as a directional signal from one vendor's dataset, never as ground truth, and always ask what's inside it.
Example. Rankscale claims an "AI readiness score" built from 200-plus factors on a credit-based, largely undisclosed methodology; Semrush's AI Visibility product reports a different composite from its own prompt-level dataset. Neither number is comparable to the other, and neither has been independently audited.
Related: Share of Answer · Mention Sentiment · Temperature & Non-determinism
Read more: Tools, and what they actually cover
Infrastructure & control
The technical layer that gates whether any of the above ever gets the chance to matter. Getting the crawlers right here is the single most consequential factual distinction in this glossary.
flowchart TD
A["New crawler request hits robots.txt"] --> B{"Which agent?"}
B -->|"GPTBot"| C{"Allow AI training on this content?"}
C -->|"Yes"| D["Allow — no effect on ChatGPT search"]
C -->|"No"| E["Disallow — no effect on ChatGPT search"]
B -->|"OAI-SearchBot"| F{"Want to appear in ChatGPT search?"}
F -->|"Yes"| G["Allow"]
F -->|"No"| H["Disallow — removed from ChatGPT search"]
B -->|"ClaudeBot"| I{"Allow AI training?"}
I -->|"Yes"| J["Allow — no effect on Claude search"]
I -->|"No"| K["Disallow — no effect on Claude search"]
B -->|"Claude-SearchBot"| L{"Want to appear in Claude search?"}
L -->|"Yes"| M["Allow"]
L -->|"No"| N["Disallow — reduced Claude visibility"]
B -->|"PerplexityBot"| O{"Want to appear in Perplexity?"}
O -->|"Yes"| P["Allow"]
O -->|"No"| Q["Disallow — removed from Perplexity"]
B -->|"Google-Extended"| R{"Allow AI training and grounding in some Google systems?"}
R -->|"Yes"| S["Allow — no effect on Search indexing"]
R -->|"No"| T["Disallow — no effect on Search indexing"]
Every one of these agents has a different job. The most expensive robots.txt mistake we see is blocking a retrieval agent while intending only to block a training agent.
robots.txt
Definition. A plain-text file at a site's root that tells compliant crawlers which agents may or may not fetch which paths.
Why it matters. It is the single point of control for the most consequential GEO mistake in the field: conflating the training decision with the retrieval decision. Blocking GPTBot is a legitimate content-licensing choice with zero effect on ChatGPT search visibility; blocking OAI-SearchBot is self-removal from ChatGPT search entirely. Reading this one file resolves a meaningful share of the "why are we invisible" audits we run.
Example.
User-agent: GPTBot
Disallow: /
User-agent: OAI-SearchBot
Allow: /
This configuration opts out of AI training while remaining fully visible in ChatGPT search.
Related: User Agent · GPTBot · OAI-SearchBot
Read more: Retrieval
User Agent
Definition. A string a crawler or client sends with each request identifying what software is making it, used by robots.txt rules to target specific bots.
Why it matters. Exact string matching is what makes a robots.txt rule apply to the right bot and no other. Guessing at a bot's name instead of using its documented exact string is a common, silent misconfiguration.
Example. OpenAI's documented OAI-SearchBot string is Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36; compatible; OAI-SearchBot/1.4; +https://openai.com/searchbot. A robots.txt rule must use the token OAI-SearchBot, exactly, to target it.
Related: robots.txt · GPTBot · OAI-SearchBot
Read more: Retrieval
GPTBot
Definition. OpenAI's crawler used to collect training data for its models. Documented UA: Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; GPTBot/1.4; +https://openai.com/gptbot. Its IP ranges are published at openai.com/gptbot.json.
Why it matters. Blocking GPTBot opts a site out of being used for OpenAI model training. It has no effect on whether the same site appears in ChatGPT search results, which is governed by an entirely different agent.
Example. A publisher choosing to license its content rather than let it train future models can Disallow: GPTBot in robots.txt without any consequence for its visibility inside ChatGPT search.
Related: OAI-SearchBot · Training Data · robots.txt
Read more: Crawlers and robots.txt control
OAI-SearchBot
Definition. OpenAI's crawler that surfaces sites in ChatGPT search. Documented UA: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36; compatible; OAI-SearchBot/1.4; +https://openai.com/searchbot. IPs published at openai.com/searchbot.json.
Why it matters. This is the single most consequential crawler distinction in the whole glossary. Blocking OAI-SearchBot removes a site from ChatGPT search entirely. It is a completely separate agent from GPTBot, and teams that block "the OpenAI bot" without checking which one frequently block this one by mistake.
Example. A company that intended only to opt out of AI training, but accidentally disallowed OAI-SearchBot along with GPTBot in a blanket rule, would be invisible to ChatGPT search while believing it had only made a training-licensing decision.
Related: GPTBot · ChatGPT · robots.txt
Read more: Crawlers and robots.txt control
ClaudeBot
Definition. Anthropic's crawler for training data collection. It supports the non-standard Crawl-delay directive. IP ranges published at claude.com/crawling/bots.json.
Why it matters. Like GPTBot, blocking ClaudeBot is a training-data-licensing decision. It has no effect on whether the site appears in Claude's search results, which a different agent handles.
Example. A site can Disallow: ClaudeBot while leaving Claude-SearchBot allowed, opting out of training while remaining visible to Claude's live web search.
Related: Claude-SearchBot · Training Data · robots.txt
Read more: Crawlers and robots.txt control
Claude-SearchBot
Definition. Anthropic's crawler that indexes content specifically to improve Claude's search quality, distinct from ClaudeBot's training role and from Claude-User, the agent used for user-directed page fetches.
Why it matters. Blocking Claude-SearchBot reduces a site's visibility in Claude's search results. This is a retrieval decision, not a training decision, and conflating it with ClaudeBot produces the same class of mistake as conflating GPTBot with OAI-SearchBot.
Example. Anthropic documents three distinct agents — ClaudeBot for training, Claude-SearchBot for search indexing, and Claude-User for fetching a page a person has explicitly pointed Claude at — each governed independently in robots.txt.
Related: ClaudeBot · Claude · robots.txt
Read more: Crawlers and robots.txt control
PerplexityBot
Definition. Perplexity's crawler that surfaces and links sites within Perplexity's answers. Documented UA: Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot). Perplexity states it does not use this crawler for model training.
Why it matters. Blocking PerplexityBot removes a site from Perplexity's index entirely. There is no separate training-vs-retrieval split for this agent the way there is for OpenAI and Anthropic; it is purely a retrieval crawler.
Example. A site disallowing PerplexityBot under a blanket "block all AI bots" policy loses Perplexity visibility with no corresponding training-data protection benefit, since Perplexity states this bot was never used for training in the first place.
Related: Perplexity · robots.txt · User Agent
Read more: Crawlers and robots.txt control
Google-Extended
Definition. A robots.txt token Google documents specifically to let sites opt out of having their content used for AI training and grounding in some of Google's other systems.
Why it matters. It does not affect Search indexing, and therefore does not affect eligibility for AI Overviews or AI Mode, which run on the standard Googlebot eligibility gate. It is purely a training and grounding opt-out for other Google AI products.
Example. A site can disallow Google-Extended while remaining fully indexed by Googlebot and fully eligible for AI Overviews, since the two decisions are governed independently.
Related: Training Data · robots.txt · AI Overview & AI Mode
Read more: Crawlers and robots.txt control
nosnippet, data-nosnippet & max-snippet
Definition. Three related controls: nosnippet (a meta tag or HTTP header preventing any snippet), data-nosnippet (an HTML attribute excluding a specific page section from snippets), and max-snippet (a directive limiting snippet length). Google states these apply to AI features exactly as they apply to ordinary text snippets.
Why it matters. These are the actual, documented controls over how a page's content can appear inside AI-generated answers, unlike llms.txt, which no major vendor documents consuming for this purpose. If you want to prevent a specific block of content from being quoted in an AI answer, this is the mechanism that is verified to work.
Example. <meta name="robots" content="nosnippet"> on a page prevents any AI Overview or ordinary snippet from quoting it; wrapping a specific paragraph in <div data-nosnippet> excludes just that paragraph while leaving the rest of the page snippet-eligible.
Related: robots.txt · Structured Data · llms.txt
Read more: Google AI Overviews and AI Mode
IndexNow
Definition. A protocol, co-created by Bing, that lets a site push instant notifications of new or changed content directly to participating search engines rather than waiting for the next crawl cycle.
Why it matters. Participating engines are Microsoft Bing, Naver, Seznam.cz, Yandex and Yep, per IndexNow's own listing. Because Copilot is generally understood to inherit Bing's index, IndexNow adoption is one of the few concretely documented, comparatively uncontested levers available for that surface. Notably, Google does not participate.
Example. A firm publishing a new case study can ping IndexNow's API immediately after publishing, getting it into Bing's index faster than waiting for Bing's own crawl schedule, with the same push reaching Yandex and the other participating engines at once.
Related: Server-Side Rendering · Freshness · robots.txt
Read more: Microsoft Copilot
llms.txt
Definition. A proposed markdown file, published at a site's root, intended to give language models a curated summary of a site's content. It originated as a September 2024 proposal by Jeremy Howard of Answer.AI, published at llmstxt.org.
Why it matters. State this plainly: no major AI vendor — not OpenAI, Google, Anthropic, Microsoft or Perplexity — has published documentation stating that its crawlers or answer engines consume llms.txt. It has been adopted by some documentation-generation tooling, but it is not a verified AI-visibility mechanism for any answer engine. Google's own AI-features documentation states you "don't need to create new machine readable files, AI text files, or markup" to appear in its AI features. Shipping an llms.txt file costs an hour and harms nothing; paying a vendor for it as a visibility service is paying for an unverified hypothesis.
Example. A documentation site publishing /llms.txt with a curated list of its key pages may help some documentation-aware tools navigate the site. It is not documented to change whether that site is retrieved by ChatGPT search, Gemini, Claude, or Google's AI features.
Related: robots.txt · Structured Data · nosnippet, data-nosnippet & max-snippet
Read more: Common GEO myths: llms.txt
Server-Side Rendering
Definition. Generating a page's full HTML on the server before sending it to the browser or crawler, as opposed to sending a near-empty page and building content client-side with JavaScript.
Why it matters. Google's own guidance is explicit: keep important content in textual form. A crawler or retrieval agent that cannot execute the same JavaScript a browser does may see an empty shell instead of the actual content, which fails retrieval before content quality is ever evaluated. This is one of the more mundane, more common causes of AI invisibility, alongside content locked in images or PDFs and infinite-scroll archives with no crawlable pagination.
Example. A pricing page that renders its actual prices only after a client-side JavaScript fetch completes may show a crawler an empty container element instead of the numbers a buyer's AI assistant is trying to retrieve.
Related: Crawl vs. Index vs. Retrieve · robots.txt · Extractability & Citable Passage
Read more: Retrieval failures beyond robots.txt
Continue in the Knowledge Hub
- GEO Pillar Page — the full discipline in one page: what GEO is, how it works, and the methodology behind it.
- GEO Methodology — the five phases in operational detail, including the entity, content and citation work referenced throughout this glossary.
- GEO FAQ — sixty questions answered educationally, many of which define terms in context.
- GEO Resource Library — primary sources, tooling comparisons, and structured learning paths.
- GEO Case Studies — how this vocabulary shows up in actual engagements.
Related services: AI Visibility & GEO · Search Dominance · Authority Architecture
Work with Artlogic
Vocabulary is not a strategy. If you want to know where your company actually stands, in the terms this glossary just defined, across the engines your buyers actually use, the useful first step is a baseline: a fixed prompt set, measured across all six surfaces, with a named competitor for comparison.
{
"@context": "https://schema.org",
"@graph": [
{
"@type": "DefinedTermSet",
"@id": "https://artlogic.ca/geo/glossary#glossary",
"name": "GEO Glossary",
"description": "A sourced glossary of terms used in Generative Engine Optimization, retrieval systems and AI answer engines.",
"url": "https://artlogic.ca/geo/glossary",
"inDefinedTermSet": "https://artlogic.ca/geo/glossary",
"hasDefinedTerm": [
{ "@type": "DefinedTerm", "name": "GEO", "url": "https://artlogic.ca/geo/glossary#geo" },
{ "@type": "DefinedTerm", "name": "SEO", "url": "https://artlogic.ca/geo/glossary#seo" },
{ "@type": "DefinedTerm", "name": "AEO, LLMO, AI SEO & GAIO", "url": "https://artlogic.ca/geo/glossary#aeo-llmo-ai-seo-gaio" },
{ "@type": "DefinedTerm", "name": "LLM", "url": "https://artlogic.ca/geo/glossary#llm" },
{ "@type": "DefinedTerm", "name": "Parametric Memory", "url": "https://artlogic.ca/geo/glossary#parametric-memory" },
{ "@type": "DefinedTerm", "name": "Hallucination", "url": "https://artlogic.ca/geo/glossary#hallucination" },
{ "@type": "DefinedTerm", "name": "Grounding", "url": "https://artlogic.ca/geo/glossary#grounding" },
{ "@type": "DefinedTerm", "name": "RAG", "url": "https://artlogic.ca/geo/glossary#rag" },
{ "@type": "DefinedTerm", "name": "Query Fan-Out", "url": "https://artlogic.ca/geo/glossary#query-fan-out" },
{ "@type": "DefinedTerm", "name": "Entity Disambiguation", "url": "https://artlogic.ca/geo/glossary#entity-disambiguation" },
{ "@type": "DefinedTerm", "name": "sameAs", "url": "https://artlogic.ca/geo/glossary#sameas" },
{ "@type": "DefinedTerm", "name": "Structured Data", "url": "https://artlogic.ca/geo/glossary#structured-data" },
{ "@type": "DefinedTerm", "name": "Answer-First Writing", "url": "https://artlogic.ca/geo/glossary#answer-first-writing" },
{ "@type": "DefinedTerm", "name": "Citation", "url": "https://artlogic.ca/geo/glossary#citation" },
{ "@type": "DefinedTerm", "name": "Dead Citation", "url": "https://artlogic.ca/geo/glossary#dead-citation" },
{ "@type": "DefinedTerm", "name": "Share of Answer", "url": "https://artlogic.ca/geo/glossary#share-of-answer" },
{ "@type": "DefinedTerm", "name": "robots.txt", "url": "https://artlogic.ca/geo/glossary#robots-txt" },
{ "@type": "DefinedTerm", "name": "OAI-SearchBot", "url": "https://artlogic.ca/geo/glossary#oai-searchbot" },
{ "@type": "DefinedTerm", "name": "llms.txt", "url": "https://artlogic.ca/geo/glossary#llms-txt" }
]
},
{
"@type": "BreadcrumbList",
"itemListElement": [
{ "@type": "ListItem", "position": 1, "name": "Home", "item": "https://artlogic.ca" },
{ "@type": "ListItem", "position": 2, "name": "GEO", "item": "https://artlogic.ca/geo" },
{ "@type": "ListItem", "position": 3, "name": "Glossary", "item": "https://artlogic.ca/geo/glossary" }
]
}
]
}
Last reviewed 6 August 2026. Where a figure comes from a vendor rather than an independent study, the entry says so.