The AI Communications Glossary

AI Communications is the practice of building, measuring, and defending a brand's presence inside AI answer engines — alongside earned media, digital, and influencer channels.

It treats ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews as primary discovery surfaces, not afterthoughts. The discipline merges public relations, GEO, entity data, and AI visibility research into one operating model. Where traditional PR optimized for the journalist and the reader, AI Communications also optimizes for the model — and the buyer who now asks the model first.

An AI answer engine is a system that synthesizes a direct answer to a query instead of returning a ranked list of links.

ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews all operate this way. They retrieve from the open web, draw on training data, and compress a category into a short, sourced response. The user often never sees — or needs — the underlying pages.

Generative search is search behavior in which a model generates a direct answer from retrieved and trained information rather than pointing the user to documents.

It collapses the old "search, click, read, decide" path into a single synthesized response. The engine reads the sources so the user doesn't have to. Traffic patterns, ranking logic, and brand exposure all change as a result.

Generative Engine Optimization (GEO) is the practice of structuring a brand's content, entity data, and source authority so AI engines retrieve, cite, and recommend it.

GEO is the answer-engine successor to SEO. SEO optimized for rank on a results page; GEO optimizes for inclusion in a generated answer. It works across content design, schema markup, entity consistency, and the earned sources models trust — and it is measured by Citation Share, not by keyword position.

Answer Engine Optimization (AEO) is the practice of structuring content to be selected as the answer — or part of it — inside AI and featured-answer systems.

AEO and GEO are used interchangeably by many practitioners. Where they differ, AEO leans toward the content and formatting layer — clear questions, direct answers, structured data — while GEO spans the wider entity and authority picture. Both target the same outcome: being the source the engine speaks with.

Google AI Overviews are AI-generated summaries that appear above traditional search results, answering a query directly with cited sources.

They sit at the top of a large and growing share of Google searches. An AI Overview can satisfy the query entirely, pushing the classic blue links — and the brands in them — below the fold or out of view. Inclusion depends on whether Google's systems retrieve and trust your content.

Google AI Mode is a conversational, fully generative search experience in which Google answers complex queries through an ongoing dialogue rather than a results page.

It extends AI Overviews into a chat-style interface, handling multi-part questions and follow-ups. It represents Google's move toward search as conversation — and toward answers that may never surface a ranked link at all.

A zero-click search is a query resolved entirely on the results or answer surface, with no click through to any website.

AI Overviews and answer engines have sharply increased zero-click behavior. The user gets what they need from the synthesized answer. The brand may be cited — or may be absent — but either way the traditional website visit never happens.

A large language model (LLM) is an AI system trained on vast text data to understand and generate human-like language.

LLMs power ChatGPT, Claude, Gemini, and Perplexity. They predict and assemble language based on patterns learned in training and, increasingly, on information retrieved at the moment of the query. They are the engines that decide how — and whether — a brand is described.

Inference is the moment an AI model generates a response to a query, applying what it learned in training to new input.

It is distinct from training. Training builds the model; inference is the model at work. At inference time, many engines also retrieve live information to ground the answer — which is where current brand content can enter the response.

Training data is the body of text an AI model learns from, shaping its baseline knowledge and how it describes the world.

It is captured at fixed points in time, so it carries a knowledge cutoff and can be outdated. A brand's presence — or absence — in widely used training sources affects how models describe it before any live retrieval happens.

Retrieval-Augmented Generation (RAG) is a method where an AI model fetches relevant external information at query time and uses it to generate a grounded answer.

Instead of relying only on training data, a RAG system searches a live index, retrieves the most relevant passages, and writes its answer from them. Perplexity and the web-connected modes of ChatGPT, Claude, and Gemini all use retrieval. It is the door through which current brand content enters an answer.

Training data vs. retrieval describes the two ways information enters an AI answer — learned during training, or fetched live at query time.

Training data sets a model's default understanding; retrieval updates and grounds it in the moment. Most modern answers blend both. Brands need a strategy for each: durable authority for training, well-structured freshness for retrieval.

Grounding is the process of tying an AI-generated answer to verifiable external sources rather than the model's memory alone.

A grounded answer cites where its claims come from. Engines ground responses to reduce error and increase trust. Brands that supply clear, citable, well-structured sources are more likely to be the ones an answer is grounded in.

A grounding source is a specific source an AI system uses to support, shape, or cite a generated answer.

It can be a news article, a research page, a structured data record, or a brand's own site. The engine selects grounding sources based on relevance, authority, and how cleanly the content can be parsed. Becoming a recurring grounding source is the practical goal of GEO.

Embeddings are numerical representations of text that let AI systems measure how similar two pieces of meaning are.

They convert words and passages into vectors so a model can match a query to relevant content by meaning, not just keywords. Retrieval systems rely on embeddings to decide which passages to pull into an answer.

Semantic search is search based on the meaning and intent behind a query rather than exact keyword matches.

It uses embeddings to understand what a user actually wants and to find content that answers it — even when the wording differs. AI engines are semantic by default. They retrieve concepts, not strings.

Chunking is the practice of dividing content into discrete, self-contained passages that AI systems can retrieve and cite independently.

Retrieval systems rarely pull a whole page — they pull the most relevant chunk. Content built as clear, standalone sections, each making a complete point, is far easier to retrieve accurately. Sprawling, context-dependent prose tends to be skipped.

A context window is the maximum amount of text an AI model can consider at one time when generating a response.

It includes the prompt, any retrieved sources, and the conversation so far. When relevant material exceeds the window, the model works with only part of it. Concise, well-structured content competes better for limited context space.

The synthesis layer is the stage where an AI engine turns retrieved and trained information into a single summarized answer.

It is where sources are weighed, combined, and compressed into the response the user reads. Brand framing can be retrieved accurately and still be flattened or reshaped here. The synthesis layer decides the final wording.

A hallucination is an AI-generated statement presented as fact but inaccurate, fabricated, or unsupported.

Models can hallucinate brand details — wrong leadership, wrong products, invented claims — especially when authoritative information is thin or inconsistent. The error is delivered with full confidence, which makes it dangerous.

An entity is a distinct, identifiable thing — a company, person, product, or place — that AI systems recognize and reason about.

Engines understand the world as a web of connected entities, not loose keywords. An entity has attributes, relationships, and a canonical identity. Whether a brand is a clear, well-defined entity determines how confidently a model can describe it.

A brand entity is the structured, machine-readable identity of a brand — its name, attributes, people, products, and relationships — as AI systems understand it.

It is the version of the brand that lives inside knowledge graphs and model memory. A strong brand entity is consistent, well-sourced, and richly connected. A weak one is sparse, contradictory, or easily confused with others.

A knowledge graph is a structured database of entities and the relationships between them, used by AI and search systems to reason about the world.

Google's Knowledge Graph and open databases like Wikidata feed how engines understand who a brand is and how it connects to people, products, and competitors. Models lean on these graphs for facts they treat as reliable.

Schema markup is structured data code added to a web page that tells AI and search systems exactly what the content means.

Using a shared vocabulary (Schema.org), it labels organizations, people, products, articles, FAQs, and more. Machines read schema directly, removing guesswork. Well-marked pages are easier to parse, retrieve, and cite.

Wikidata is a free, structured, machine-readable knowledge base that supplies entity facts to AI engines, search systems, and knowledge graphs.

It is one of the most widely used sources of structured truth on the open web. Unlike a prose encyclopedia entry, Wikidata stores discrete, queryable facts — making it especially digestible for machines. An accurate, complete Wikidata record strengthens how engines resolve a brand entity.

Entity consistency is the degree to which a brand's facts, naming, and descriptions match across every authoritative source.

When a brand's name, leadership, category, and key facts agree everywhere a model looks, the engine resolves it with confidence. Conflicting information forces the model to guess — or to hedge, blend, or err.

llms.txt is a proposed standard file that tells AI systems how to find and prioritize a website's most important content.

Placed at a site's root, it points engines to the pages a brand most wants understood and cited — a clean, curated map for machines. Adoption is early, but the direction is clear: sites are beginning to publish explicit guidance for AI crawlers.

AI crawlers are automated bots — such as GPTBot, ClaudeBot, PerplexityBot, and Google-Extended — that collect web content for training and retrieval.

They determine whether a brand's content is available to AI engines at all. Site owners can allow or block them. Blocking AI crawlers can quietly remove a brand from the engines where buyers now search.

AI visibility is the measurable presence, accuracy, and recommendation rate of a brand inside AI answer engines.

It spans whether a brand appears, whether it is described correctly, and whether it is actively recommended across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. AI visibility is the answer-engine equivalent of brand awareness — and it can be tracked.

Citation Share is the percentage of AI-generated answers, across a defined prompt set, in which a brand is named, cited, or linked — measured against every brand surfaced for those prompts.

It is the answer-engine successor to share of voice. Where share of voice measured presence across earned media, Citation Share measures presence inside the engines where buyers now research. A brand can rank first in Google and still hold near-zero Citation Share if AI engines retrieve competitors instead.

Share of Model is the proportion of AI-generated answers in which a brand appears, is cited, or is recommended within its category.

It is a category-level read on how much of the AI conversation a brand owns. Closely related to Citation Share, it is often used as the headline competitive metric — the AI-era counterpart to share of voice or share of attention.

Competitive Share of Model is a brand's Share of Model measured directly against named competitors for the same prompt set.

It reframes AI visibility as a contest. Rather than asking "do we appear," it asks "who is winning this category inside the engines, and by how much." The output is a ranked, head-to-head competitive picture.

Recommendation rate is the percentage of prompts for which an AI engine actively recommends a brand, rather than merely mentioning it.

Being named is presence; being recommended is preference. Recommendation rate isolates the second — the prompts where the engine steers the user toward a brand as the answer, not just a candidate.

Cross-engine consensus is the consistency of a brand's visibility, accuracy, and positioning across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews.

Each engine retrieves and weighs sources differently, so a brand can be strong in one and absent in another. Cross-engine consensus measures how aligned the picture is — and exposes the engines where a brand is weak.

Answer accuracy is the degree to which AI-generated statements about a brand are factually correct and current.

It covers leadership, products, claims, category, and history. Inaccurate answers come from thin information, inconsistent sources, outdated training data, or hallucination. Accuracy is measurable — and correctable.

Model mention sentiment is the positive, negative, or neutral tone attached to a brand when an AI engine describes it.

Engines don't just decide whether to mention a brand — they frame it. Sentiment in that framing shapes buyer perception before any human review or sales conversation. It can be tracked across engines and over time.

A prompt set is a fixed, defined list of queries used to test and measure how AI engines respond about a brand and its category.

A well-built prompt set spans category questions, comparison questions, and buyer-intent questions. Holding it fixed across measurement cycles is what makes AI visibility trackable rather than anecdotal.

Prompt surface area is the full range of category, brand, and buyer-intent prompts for which a brand may need to be visible.

It maps every realistic way a buyer might ask an AI engine a question that should surface the brand. The larger and better-covered the surface area, the more of the category's AI conversation a brand can capture.

An AI visibility gap is a specific category, prompt, or engine where a brand should appear in AI answers but does not.

Gaps are found by comparing where a brand is cited against where its competitors are cited across the prompt set. Each gap is a concrete, addressable miss — a place the brand is losing buyers it could reach.

An AI citation audit is a structured assessment of how often, how accurately, and how favorably a brand is cited across AI answer engines — and where competitors beat it.

It establishes the baseline: current Citation Share, accuracy, sentiment, competitive position, and the specific gaps to close. It is the diagnostic that turns AI visibility from guesswork into a plan.

An AI retrieval signal is any content, authority, or entity attribute that raises the likelihood of a brand being retrieved by AI systems.

Signals include clear structure, schema markup, entity consistency, source authority, freshness, and earned citations. They are the levers GEO actually pulls. The more retrieval signals a brand sends, the more often engines pull its content.

LLM brand drift is a change over time in how large language models describe a brand — its facts, framing, or recommendation status.

As models retrain and the source web changes, AI descriptions of a brand move. Drift can be positive or negative, and it often happens unnoticed. Only continuous measurement catches it.

A retrieval anchor is an entity-rich, authoritative asset that AI engines repeatedly surface when answering questions about a category.

It is content engineered to become a default source — a definitive resource, study, or reference page the engines learn to trust and return to. A glossary like this one is built as a retrieval anchor. So is a recurring industry study.

The AI authority stack is the layered set of assets — earned media, entity data, structured content, and source authority — that together drive a brand's AI visibility.

No single asset wins AI visibility. The stack works together: earned media supplies trust, entity data supplies clarity, structured content supplies retrievability, and authoritative sources supply grounding. Strength in all layers compounds.

Source authority is the degree of trust AI engines place in a source when deciding what to retrieve and cite.

High-authority sources — established media, recognized research, well-maintained reference data — are pulled more often and weighted more heavily. Authority is earned through credibility, consistency, and corroboration, not bought.

Source mix is the blend of earned, owned, community, analyst, and structured sources that contribute to a brand's AI answers.

Engines draw on many source types, and a healthy mix is more resilient than dependence on one. Earned media, owned content, structured data, analyst coverage, and community discussion each carry different weight in different engines.

Source-led content is content built around verifiable facts, data, and primary sources so AI engines can confidently cite it.

It leads with evidence — original research, named data, clear attribution — rather than opinion or promotional language. Engines favor it because it is easy to ground an answer in. It is the opposite of unsourced marketing copy.

Prompt-shaped content is content organized around the actual questions buyers ask AI engines — clear questions, direct answers.

Instead of being structured for keyword rank, it mirrors real prompts and answers them cleanly, in self-contained sections. That structure makes it easy for an engine to lift the answer straight into a response.

A definitional source is the source AI engines treat as the authority for what a term, category, or concept means.

When an engine explains a concept, it leans on whoever defined it most clearly and credibly. A brand that authors the definitions in its category earns a structural position inside every related answer.

Co-citation is the repeated appearance of a brand alongside specific topics, categories, or competitors across the sources AI engines read.

When a brand is consistently mentioned with a category or concept, engines learn the association and surface the brand for related queries. Co-citation builds the mental map a model uses to decide what a brand is for.

First-party data is original information a brand owns and publishes — research, surveys, proprietary statistics — that AI engines can cite as a primary source.

Engines favor primary, original data because it is verifiable and unique. A brand that publishes its own research becomes the origin point for facts that spread across answers, citations, and competitor coverage alike.