The GEO & AEO glossary

Also: GEO

Generative Engine Optimization (GEO) is the practice of improving how often and how favourably a brand is mentioned, cited, and recommended inside AI-generated answers.

The term was coined by Aggarwal et al. in “GEO: Generative Engine Optimization” (KDD 2024), which framed content creators as a third stakeholder with “little to no control over when and how their content is displayed” in generative engines. Note that Google's official position is that GEO is not a distinct discipline — its AI features run on core Search ranking, so optimizing for them “is still SEO”. That objection is strongest for Google's own surfaces and weakest for ChatGPT, Claude and Perplexity, which don't use Google's ranking at all.

ProvenanceAggarwal et al., KDD 2024 (arXiv:2311.09735)

Answer Engine Optimization (AEO)Answer engineSearch Engine Optimization (SEO)

Also: AEO

Answer Engine Optimization (AEO) is the practice of optimizing content to be used in a direct answer rather than a ranked list of links. In current usage it is effectively a synonym for GEO.

AEO and GEO are used interchangeably by most practitioners; no authority distinguishes them. Where people do draw a line, AEO is framed as the broader idea of optimizing for any direct-answer surface (including featured snippets and voice assistants, which predate LLMs), while GEO is specific to generative engines. Treat any vendor drawing a sharp distinction as marketing.

Generative Engine Optimization (GEO)Answer engine

Also: Generative engine

An answer engine is a system that responds to a query with a synthesized natural-language answer, optionally citing sources, instead of returning a ranked list of links.

Examples include ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews / AI Mode. The GEO paper defines the related “generative engine” as a system that combines retrieval with generative models to produce answers. The practical distinction from a search engine is that the user is shown a synthesis, not a set of options — so there is no “page two” to rank on.

ProvenanceAggarwal et al., KDD 2024

Generative Engine Optimization (GEO)GroundingRetrieval-Augmented Generation (RAG)

Also: SEO

Search Engine Optimization (SEO) is the practice of improving a site's visibility in a search engine's ranked, organic results.

Relevant here because Google states that optimizing for its generative features is still SEO, and because C-SEO Bench (NeurIPS 2025) found that traditional retrieval/ranking work outperformed most GEO-specific content tactics at winning citations. The honest framing: SEO gets you retrieved; GEO concerns what happens once you are.

Generative Engine Optimization (GEO)Retrieval

A zero-click search is a query that is resolved on the results surface itself — by an AI answer, snippet, or knowledge panel — without the user clicking through to any source.

The term predates AI Overviews (it described featured snippets and knowledge panels) but has become the main mechanism by which AI answers reduce referral traffic: your content is used to build the answer, and the visit never happens. This is why traffic can fall while rankings hold steady.

Answer engineCitation

Also: AI SOV · AI visibility score

AI Share of Voice is a vendor-defined metric estimating how often and how prominently a brand appears in AI answers for a defined prompt set, relative to named competitors.

This is a vendor coinage — including ours. There is no standards body, no agreed formula, and no cross-vendor comparability: two vendors' “AI Share of Voice” figures are generally not measuring the same thing, and should never be compared directly. Ours weights mention rate, position in the answer, sentiment, and recommendation strength across a fixed prompt corpus and five engines; the formula is published in our methodology so you can audit it. Treat any unaudited SOV number — ours included — as internal to the vendor that produced it.

ProvenanceVendor coinage (including Word of GPT). No authoritative definition.

Mention rateCitation ratePrompt corpus

Mention rate is the share of sampled AI responses, for a defined prompt set, in which a brand is named at all.

The simplest and most robust AI-visibility metric, because it requires no judgement about position or sentiment — the brand is either named or it isn't. Only meaningful when computed over repeated samples: AI answers vary run to run, so a mention rate derived from a single pass is noise.

AI Share of VoiceRepeated samplingCitation rate

Citation rate is how often a specific domain or URL is linked as a source in AI answers for a defined prompt set.

Distinct from mention rate: an engine can recommend your brand without citing your site (drawing the claim from a third party such as Reddit or G2), and it can cite your site without recommending you. Brand-level presence is the more reliable KPI, because cited URL sets are unstable between runs.

Mention rateCitation source map

Repeated sampling is querying each prompt multiple times, across engines and over a window, so that a visibility metric reflects a distribution rather than a single generation.

This is not methodological fussiness — it is the difference between measurement and anecdote. Day-to-day source overlap in AI answers runs roughly 0.34–0.42, and in one analysis 57.8% of ChatGPT runs never triggered a web search at all. A single prompt check can show you as category leader or invisible purely by chance. We sample every prompt multiple times across a 14-day window for this reason.

Prompt corpusMention rateAI Share of Voice

A prompt corpus is the fixed, categorised set of buyer-intent prompts a brand's AI visibility is measured against.

The corpus is the measurement instrument: change it and the score changes, so it must be fixed, documented, and held constant period over period for trends to mean anything. A useful corpus spans intent stages — problem-aware, solution-aware, brand-aware, and competitor-comparison — rather than only branded queries, which flatter the brand.

Repeated samplingBuyer-intent promptAI Share of Voice

A buyer-intent prompt is a question a prospective customer would realistically ask an AI engine while researching or choosing a vendor.

The classes that matter commercially are vendor-shortlist prompts (“best X tools for Y”) and head-to-head comparisons (“A vs B”), because these are where consideration sets are formed — increasingly before any contact with a salesperson.

Prompt corpusZero-click search

Recommendation strength is how firmly an AI answer endorses a brand — from a passing mention, to one option among several, to the explicit top pick.

The distinction that matters commercially: being named is not the same as being recommended. Worth tracking alongside mention rate, though note it requires a judgement call (usually a classifier's), which makes it inherently softer than the binary 'were we named or not'.

Mention ratePosition in answerAI Share of Voice

Position in answer is where a brand appears within an AI response — named first, listed mid-answer, or buried at the end.

A proxy for prominence, on the assumption that earlier mentions carry more weight with the reader. Worth measuring, but note that the academic word-count-share metrics used in the original GEO paper have been criticised (by C-SEO Bench) for not reflecting actual model preference the way citation ranking does.

AI Share of VoiceMention rate

Also: RAG

Retrieval-Augmented Generation (RAG) is an architecture in which a system retrieves relevant documents and supplies them to a language model as context, so the answer is generated from fetched sources rather than memory alone.

Introduced by Lewis et al. (2020). It's the mechanism underneath most cited AI answers, and it explains why GEO is largely a retrieval problem: if your page is never fetched into the context window, nothing you wrote on it can influence the answer. C-SEO Bench found that getting a document into the context — and high in it — mattered more than any content-modification tactic.

ProvenanceLewis et al., 2020 (arXiv:2005.11401)

RetrievalGroundingContext windowParametric memory

Retrieval is the step in which an answer engine selects which documents to fetch and place in the model's context before generating a response.

The dominant lever in AI visibility, and the least discussed by GEO vendors — probably because it looks a lot like ordinary SEO and technical hygiene. If you are not retrieved, you cannot be cited, regardless of how extractable your prose is.

Retrieval-Augmented Generation (RAG)Context windowPer-engine crawler access

Grounding is the practice of constraining a model's answer to retrieved source material so that claims can be attributed to documents rather than generated from memory.

Grounded answers are the ones that carry citations — and therefore the ones GEO can influence. Ungrounded answers come from the model's parametric memory, where your only leverage is long-run presence in training data and the wider web.

Retrieval-Augmented Generation (RAG)Parametric memoryCitation

Parametric memory is what a model 'knows' from training, as opposed to what it retrieves at query time.

Commercially important and routinely ignored: in one analysis 57.8% of ChatGPT runs never triggered a web search, meaning the answer came from memory with no citation and no retrieval step to influence. For those answers, no amount of on-page optimization helps — only being sufficiently present across the web that the model learned you.

GroundingRetrieval-Augmented Generation (RAG)Entity

The context window is the working set of text a model can consider at once, including any retrieved documents.

Relevant to GEO because position within the context matters: C-SEO Bench reports that making a document first in the LLM context produced far greater citation gains than any content-modification method it tested.

Retrieval-Augmented Generation (RAG)Retrieval

Also: Fan-out queries

Query fan-out is when an answer engine decomposes one user question into several internal search queries and synthesizes the results into a single answer.

Described by Google in the context of AI Mode. It means the query you think you're being measured on is often not the query the engine actually ran — which is a structural argument for measuring brand presence across a broad prompt corpus rather than optimizing for individual keywords.

ProvenanceDescribed by Google for AI Mode

Prompt corpusAnswer engine

A hallucination is a confident, fluent model output that is factually wrong or unsupported by any source.

The reputational face of AI visibility: an engine can state your pricing, features, or leadership incorrectly, at scale, with no byline to correct. Remediation means fixing the sources the model relies on, not appealing to the model.

GroundingEntity

Per-engine crawler access is the practice of allowing or blocking each AI company's individual user agents — separately for training, live retrieval, and search indexing — in robots.txt.

On the current evidence this is the most concrete, primary-source-documented technical lever available to a GEO practitioner. Each major engine runs multiple separately-blockable agents, and a blanket “block AI bots” rule is a common misconfiguration that silently removes a site from the citable surface — you lose live citations, not just training inclusion. Unglamorous, free, and more consequential than most of what this industry sells.

ProvenanceOpenAI, Anthropic, Perplexity and Google crawler documentation

llms.txtRetrieval

llms.txt is a proposed Markdown file at a domain's root offering AI engines a curated map of its key content. As of 2026 no major engine is documented as reading it.

Google has explicitly stated it ignores llms.txt. OpenAI's, Anthropic's and Perplexity's crawler docs describe robots.txt-based controls and never mention consuming it from third-party sites. Several of those companies publish an llms.txt for their own documentation — publishing is not consuming. Ship it if you like: it's an hour's work and harmless. Do not buy it as a visibility lever.

ProvenanceProposal at llmstxt.org; not adopted by any major engine

Per-engine crawler accessrobots.txt

robots.txt is the long-standing file at a domain's root that tells crawlers which paths they may request, by user agent.

The actual control surface for AI crawlers, as opposed to llms.txt. Compliance is voluntary and honour-based; Cloudflare has published findings of at least one AI company circumventing it, and no equivalent independent audit exists for most engines.

Per-engine crawler accessllms.txt

Also: Schema markup · JSON-LD

Structured data is machine-readable markup that states explicitly what a page is about — its entities, author, prices, and relationships — most commonly as schema.org vocabulary in JSON-LD.

Worth shipping for conventional rich-result eligibility and entity clarity. It is not, per Google, a requirement for or a lever on generative AI search: no schema type is specific to AI features, and Google fully removed FAQ rich results on 2026-05-07. Whether markup affects citation in non-Google engines is an open question with no good public evidence.

Provenanceschema.org; Google Search Central documentation

EntityExtractability

An entity is a distinct thing — a company, person, product, or place — that a search or answer engine can identify and reason about, as opposed to a string of characters.

Foundational to AI visibility: if an engine cannot resolve your brand to a distinct entity, it will resolve your name to the nearest entity it does know, which may be a competitor or an unrelated product with a similar name. Disambiguation signals — consistent naming, sameAs links to authoritative profiles, schema, and corroborating third-party references — are how you become resolvable.

Entity disambiguationStructured data (schema.org / JSON-LD)Parametric memory

Entity disambiguation is the work of making an engine resolve your brand to you, rather than to a different entity with a similar name.

The failure mode is specific and common for new brands: with no corroborating references, a model resolves the name to whatever it already knows. Levers include an explicit disambiguatingDescription in schema, consistent naming, sameAs links to authoritative profiles, and third-party references that co-occur your name with your category. We had this exact problem — see our about page.

EntityStructured data (schema.org / JSON-LD)

An LLM-trusted source is a domain that answer engines cite disproportionately often within a given category — commonly Reddit, Wikipedia, review platforms, or a handful of sector publications.

A useful shorthand rather than a formal category: no engine publishes a list of trusted domains, and the set is category-specific and shifts over time. What is well evidenced is that the cited set diverges sharply from organic results, so the domains deciding your category's answers are often ones you don't own and may not track.

Citation source mapThird-party presenceCitation rate

Extractability is how easily a model can lift a clean, quotable claim out of a page — via clear statements, statistics, named entities, and question-and-answer structure.

Treat as hygiene, not strategy. The GEO paper reported gains for tactics like adding statistics and quotations, but C-SEO Bench (NeurIPS 2025) re-tested seven of those methods and found most “largely ineffective” and “frequently” negative for citation ranking, while traditional retrieval work was significantly more effective — and found the gains congested and zero-sum as adoption spreads. Write clearly because it serves readers and cannot hurt; don't buy it as a moat.

ProvenanceAggarwal et al. (KDD 2024) vs. C-SEO Bench (NeurIPS 2025)

Structured data (schema.org / JSON-LD)RetrievalGenerative Engine Optimization (GEO)

A citation source map is an inventory of which URLs and domains an answer engine actually cites across a category's prompt set.

The most useful artefact in AI-visibility work, because the cited set diverges sharply from the ranked set: roughly half of AI Overviews citations come from domains outside the organic top 10, and cross-engine URL overlap is low. It tells you which third-party properties — often Reddit threads, review sites, or a handful of publications — are actually deciding your category's answer.

ProvenanceVendor coinage (including Word of GPT); the underlying divergence is well evidenced.

Citation rateThird-party presence

A content gap is a topic or buyer question where competitors are mentioned or cited in AI answers and you are not.

Useful as a prioritisation tool: it tells you which prompts you're losing and to whom. Treat the implied remedy with care, though — publishing a page on the topic does not reliably win the citation, because the gap is often held by a third-party source rather than the competitor's own site. Diagnose the gap, then fix the actual cause.

Citation source mapPrompt corpusThird-party presence

Third-party presence is a brand's footprint on sites it does not own — review platforms, communities, directories, and publications — that answer engines cite.

Usually the binding constraint on AI visibility, and the part no on-site work can fix. Because the cited source set diverges so far from organic results, being recommended often depends on what other people's pages say about you rather than on your own.

Citation source mapEntity

Programmatic SEO is generating a large set of pages from structured data against a template, so that each page covers a distinct entity or query.

Relevant to GEO because it scales the surface an engine can retrieve. The failure mode is thin, near-duplicate pages that add no per-page value; the discipline is making each generated page genuinely distinct and useful.

RetrievalStructured data (schema.org / JSON-LD)

A citation is a link or attribution an answer engine attaches to a claim, pointing to the source it drew from.

The unit of account in GEO. Note the asymmetry that makes measurement subtle: being cited is not the same as being recommended — an engine may cite your page while recommending a competitor, or recommend you while citing someone else's page about you.

Citation rateMention rateGrounding