The AIO Lexicon
The vocabulary of AI search, defined in plain language. When an AI engine decides which businesses to name, it runs on machinery with names most marketers have never heard: SBERT, dense retrieval, reranking, embeddings. This is the complete glossary of that machinery, and what each term means for whether AI can find you. Nothing here is left unexplained.
Retrieval & Ranking Mechanics
How an AI engine finds candidate sources before it writes a word. This is where visibility is won or lost.
SBERT Sentence-BERT, sentence transformers
A bi-encoder model that compresses an entire sentence into a single fixed-length vector, so a machine can judge whether two texts mean the same thing by measuring the distance between their vectors. It made large-scale semantic search practical, cutting a task that took plain BERT sixty-five hours down to about five seconds, and it is a foundational mechanism in how AI engines retrieve and cite sources.
Full explainer, with videoEmbeddings vector embeddings
Numerical representations of text (or images, or audio) as points in a high-dimensional space, arranged so that similar meanings land near each other. Embeddings are the common currency of AI retrieval: your content is only findable to the degree that its embedding lands near the questions your customers ask.
Full explainer, with videoCosine Similarity
The measure most systems use to score how close two embeddings are: the cosine of the angle between the two vectors, from -1 (opposite) to 1 (identical in direction). When an AI engine ranks which passages are most relevant to a question, cosine similarity is usually the number doing the ranking.
Full explainer, with videoDense Retrieval
Retrieval by meaning rather than by keyword: the query and every document are embedded into dense vectors, and the system returns the nearest neighbors. Dense retrieval is why a page that never uses the searcher's exact words can still be the source an AI cites, and why exact-match keyword tactics no longer guarantee visibility.
Full explainer, with videoSparse Retrieval BM25, keyword search
The classic approach that matches literal terms and weights them by frequency and rarity, with BM25 as the long-standing standard. It remains fast, transparent, and strong on exact names and codes, which is why modern systems rarely abandon it and instead fuse it with dense retrieval.
Hybrid Search
The combination of sparse (keyword) and dense (semantic) retrieval, blending their scores so a system catches both exact matches and meaning-based matches. Most production AI-search stacks are hybrid, because neither method alone covers every query type.
Reranking cross-encoder reranker
A second, more precise pass that re-scores a shortlist of retrieved passages by reading the query and each passage together. Reranking sharpens which of your pages actually makes it into the answer, but a page can only be reranked if it was retrieved in the first pass.
Full explainer, with videoChunking semantic chunking
The splitting of a page into passages small enough to embed and retrieve individually. How you are chunked decides how you are found: a self-contained passage embeds into a clear, representative point, while a fragment that depends on earlier context embeds into a muddy one that no query matches.
Full explainer, with videoVector Database vector store
A database built to store embeddings and find nearest neighbors at scale, using approximate-nearest-neighbor indexes to search millions of vectors in milliseconds. It is the retrieval memory an AI engine reaches into when it needs sources for an answer.
Full explainer, with videoApproximate Nearest Neighbor ANN
The family of algorithms (HNSW, IVF, and others) that make vector search fast by finding almost-closest matches without comparing every vector. ANN is the engineering trick that lets semantic search run over the whole web in real time.
Retrieval-Augmented Generation RAG
The dominant architecture behind AI answers: retrieve relevant passages from a corpus, then feed them to a language model to compose a grounded, cited response. RAG is why AI visibility is a retrieval problem first and a writing problem second. If you are not retrieved, you are not in the answer.
Full explainer, with videoAgentic RAG
A newer pattern where the model plans, issues multiple searches, and reasons across rounds of retrieval rather than a single lookup. It rewards sources that answer follow-up questions cleanly, not just the first query.
Context Window
The maximum amount of text a model can consider at once, measured in tokens. It bounds how many retrieved passages an engine can read before answering, which is one reason concise, high-signal passages are favored over sprawling ones.
Grounding grounded generation
Tying a model's output to retrieved source material so its claims are supported and attributable, rather than generated from memory alone. Grounding is what produces citations, and citations are the currency of AI visibility.
Retrieval Legibility
Our term for how easily a machine can read, parse, and embed the actual facts on your page. Prices locked in images, key claims buried in scripts, or facts split across interactive widgets are illegible to a retriever, and illegible facts cannot be retrieved or cited, however true they are.
Full explainer, with videoAI Search & Optimization Terms
The surfaces where AI answers appear, and the competing names for the discipline of being chosen by them.
AIO AI Optimization
The practice of making a business easy for AI systems to find, understand, trust, and recommend. Note the ambiguity worth owning: some use AIO to mean Google's AI Overviews. We use it in the broader, discipline-level sense, the AI-era successor to SEO.
Full explainer, with videoAI Overviews AIO, formerly SGE
Google's AI-generated answer that appears above the traditional blue links, synthesizing multiple sources into a direct response with citations. Being cited in an AI Overview is now one of the most valuable positions in search, and it is earned through retrievability and authority, not ad spend.
Full explainer, with videoAI Mode
Google's fully conversational search experience, where a user carries on a dialogue and the engine answers across turns rather than returning a page of results. It pushes ranking further away from ten blue links and toward being the source a conversation is built on.
SGE Search Generative Experience
Google's original name for its generative search experiment, since evolved and rebranded into AI Overviews and AI Mode. You will still see SGE in older strategy documents; treat it as the predecessor term.
Answer Engine
Any system that returns a direct, synthesized answer instead of a list of links: ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot. The shift from search engines to answer engines is the entire reason AI visibility is now its own discipline.
Full explainer, with videoAEO Answer Engine Optimization
Optimizing to be the source a direct answer is built from: clear, self-contained, question-shaped content that an engine can lift and attribute. AEO overlaps heavily with GEO and LLMO; the differences are mostly emphasis and vintage.
Full explainer, with videoGEO Generative Engine Optimization
The discipline, named in 2023 research, of improving a brand's visibility inside generative engine answers. GEO reframes optimization around being retrieved, cited, and synthesized rather than ranked, and it is fast becoming the headline term for the whole field.
Full explainer, with videoLLMO Large Language Model Optimization
Optimizing specifically for how large language models represent, recall, and recommend your brand, including the knowledge they carry in their weights and the sources they pull at runtime. A near-synonym for GEO and AEO, favored by those who emphasize the model itself.
Full explainer, with videoZero-Click Search
A search that ends without a click, because the answer was delivered on the results page itself. AI answers accelerate zero-click, which raises the stakes of being cited: often the citation is the only visibility, and the only click, you will get.
Citations inline citations, grounding links
The source links an answer engine attaches to its claims. Citations are the measurable outcome of AI visibility: being named and linked inside the answer is the AI-era equivalent of ranking first.
Entity & Knowledge
How machines model your business as a distinct, real-world thing, and connect it to what it does and where.
Entity named entity
A distinct, identifiable thing (a business, person, place, or product) that a machine can recognize and reason about, separate from the words used to describe it. AI systems increasingly think in entities, not keywords, which is why being an unambiguous, well-defined entity matters more than any single phrase.
Knowledge Graph
A structured network of entities and the relationships between them, the substrate behind knowledge panels and much of how engines understand who you are. Being a clean, connected node in the knowledge graph is a prerequisite for confident AI recommendation.
Full explainer, with videoEntity SEO
Optimizing so search and AI systems recognize your business as a well-defined entity: consistent naming, clear relationships, and corroboration across authoritative sources. It is the bridge between old-world SEO and entity-driven AI understanding.
Full explainer, with videoEntity Salience
How central and important an entity is within a given piece of content, as assessed by language models. High salience for the right entity tells an engine your page is genuinely about your business and its category, not merely mentioning it in passing.
Entity Disambiguation
The process of deciding which real-world entity a name refers to when several share it. Businesses with generic or shared names must supply strong disambiguating signals, or risk being confused with, or absorbed into, someone else.
Co-occurrence
The pattern of which entities, terms, and topics appear together across the web. Consistent co-occurrence of your brand with its category and location teaches AI systems what you are and where you belong.
sameAs
A schema.org property that links your entity to its authoritative profiles elsewhere: Wikipedia, Wikidata, LinkedIn, official social accounts. It is one of the most direct ways to tell a machine, explicitly, that these references are all the same you.
Knowledge Panel
The information box that appears for a recognized entity, drawn from the knowledge graph. Earning one is strong evidence that engines model you as a real, verified entity worth surfacing.
Brand SERP
The results page for a search of your own brand name, the clearest mirror of how engines understand you. A clean, controlled brand SERP is foundational; if engines are confused about you here, they will be confused everywhere.
Machine Legibility
The structured signals and access rules that decide whether a machine can read you at all.
Structured Data Schema.org, JSON-LD
Machine-readable markup that states facts about your business explicitly: name, type, address, ratings, services. Delivered as JSON-LD, structured data removes ambiguity, turning things a machine would have to infer into things it can simply read.
Full explainer, with videollms.txt
A proposed standard file that offers AI systems a clean, curated map of your most important content in plain markdown. Early but rising, it is a direct invitation to language models to understand your site on your terms.
Full explainer, with videoAI Crawlers GPTBot, ClaudeBot, PerplexityBot, Google-Extended
The bots that fetch web content for AI training and live retrieval. Whether you allow or block them, in robots.txt and at the edge, directly determines whether your business can appear in AI answers at all. Blocking them is an invisibility switch many sites flip by accident.
Full explainer, with videorobots.txt
The file at the root of your site that tells crawlers what they may access. In the AI era it carries new directives (such as those for Google-Extended and GPTBot) that govern AI use specifically, and misconfiguring it can silently remove you from AI answers.
Semantic HTML
Markup that uses meaningful elements (headings, lists, articles, tables) so structure conveys meaning, not just appearance. Semantic HTML makes your content easier for machines to parse, chunk, and understand.
Fact Legibility
Whether your concrete facts, prices, hours, credentials, numbers, exist as readable text rather than being locked inside images or rendered only by scripts. Facts a machine cannot read are facts it cannot cite.
DefinedTerm
The schema.org type used to mark up glossary entries and terminology, like this lexicon itself. Using it signals to machines that a page is an authoritative definition, the kind of source AI answers prefer to cite.
Model Fundamentals
The core machinery of language models, in the terms you will meet everywhere else.
LLM large language model
A model trained on vast text to predict language, and in doing so to summarize, reason, and answer. LLMs are the engines behind every AI answer surface, and their behavior is what AI visibility ultimately aims to influence.
Transformer
The neural network architecture, built on attention, that underpins BERT, SBERT, GPT, and essentially all modern language and embedding models. It is the foundation the entire field is built on.
Token / Tokenization
The unit a model actually processes, roughly a word-piece, and the act of splitting text into those units. Tokens are how model limits, costs, and context windows are measured.
Fine-Tuning
Further training a pretrained model on specific data to specialize it, the step that turns general BERT into task-tuned SBERT, for example. It is how broad models become sharp tools.
Distillation
Training a smaller, faster model to mimic a larger one, preserving most of the quality at a fraction of the cost. Distillation is why fast, cheap embedding and answer models can run at web scale.
Hallucination
When a model states something fluent but false, typically because it is generating from memory rather than grounded in retrieved sources. Strong grounding, and being the clear authoritative source, reduces the chance an engine invents something wrong about you.
Temperature
A setting that controls how random or deterministic a model's output is, higher for creative variety, lower for consistency. Relevant to why the same AI question can yield different answers on different runs.
System Prompt
The hidden instructions that shape how an AI assistant behaves before it sees the user's message. It frames the entire interaction, including how the model weighs and presents sources.
MCP Model Context Protocol
An open standard for connecting AI models to external tools and data sources in a structured way. As agents adopt it, MCP shapes how models reach live information, including information about your business.
Trust, Authority & Measurement
The signals that make a source trustworthy, and the ways AI visibility is measured.
E-E-A-T Experience, Expertise, Authoritativeness, Trust
Google's framework for evaluating source quality, now a useful lens for AI trust generally. Engines favor sources that demonstrate real experience, expertise, authority, and trustworthiness, which is exactly the standing AI visibility is built to earn.
Full explainer, with videoTopical Authority
Depth and breadth of credible coverage across a subject, which makes a site a go-to source in its category. This lexicon is itself a topical-authority play: comprehensive coverage of a field signals to AI that you are a source worth citing.
Corroboration
Independent confirmation of your claims across third-party sources, directories, press, and profiles. AI systems trust what multiple credible sources agree on, so corroboration is one of the strongest trust signals you can build.
Sentiment
Whether the way you are described, by sources and by the model itself, skews positive, neutral, or negative. Being cited is necessary but not sufficient; how you are characterized matters just as much.
AI Visibility Score
A composite measure of how findable, understandable, and trustworthy your business is to AI systems. AIOInsights computes one deterministically across six pillars and forty observable signals, so the same site always produces the same score.
Deterministic Scoring
Scoring where the same input always yields the same output, with every point traceable to an observable signal. It is the opposite of a black-box number that drifts between runs, and it is the standard every AIOInsights evaluation is held to.
New term entering the field? It belongs here. This lexicon is maintained so that no vocabulary of AI search goes unexplained on AIOInsights. If a term is missing, it is a gap we close, not a surprise.
Understand the vocabulary, then measure where you stand across all six pillars.