GEO & AI Search Glossary

Definitions

Each entry states how the term is used on this site. Where the wider industry disagrees, the disagreement is noted rather than papered over.

Generative engine optimisation (GEO): The practice of making content easier for AI search and answer systems to discover, interpret, verify and cite. It extends technical SEO and editorial quality with passage-level evidence analysis, provenance requirements and citation measurement.
Answer engine optimisation (AEO): Used interchangeably with GEO by most practitioners. Where a distinction is drawn, AEO emphasises being the direct answer to a question, and GEO emphasises being a cited source within a generated response.
Answer capsule: A short, self-contained passage written so it can be quoted by an engine without losing meaning. A capsule states its subject explicitly rather than relying on the heading or the preceding paragraph for context.
Citation: A source URL returned alongside a generated answer. A citation is an observation at a moment in time for a specific query, not proof that a model was trained on the page or that it ranks the page permanently.
Presence rate: The proportion of sampled runs in which a brand or domain appears, across a fixed query set. It is the smallest honest unit of AI visibility measurement and should always be reported with its sample size and confidence interval.
Share of voice: A brand's citations as a proportion of all citations returned across a query set, measured against a named competitor set. Meaningful only when the query set and competitor list are held constant between runs.
llms.txt: A proposed convention: a Markdown file at a domain root giving language models a curated index of the pages the publisher considers authoritative. It is not a ratified standard and not a ranking factor.
Training crawler: A crawler that fetches pages to build model weights, such as GPTBot, ClaudeBot or CCBot. Content gathered this way produces no attribution and no referral traffic to the publisher.
Search crawler: A crawler that indexes pages so an engine can retrieve and cite them in answers, such as OAI-SearchBot or PerplexityBot. Blocking these removes a site from AI answers even when its content is otherwise excellent.
On-demand fetcher: A user-agent that retrieves a page because someone asked a question about it in that moment, such as ChatGPT-User or Perplexity-User. Blocking these is the most direct route to being uncitable.
Retrieval-augmented generation (RAG): An architecture where a model retrieves relevant documents at query time and generates an answer grounded in them. It is why live citation is possible without the source appearing in training data.
Passage retrieval: Matching a query against fragments of documents rather than whole pages. It is the reason a comprehensive guide that never states its answer plainly can lose to a shorter page that does.
Provenance: The visible chain connecting a claim to its origin: who published it, when, based on what evidence, and where that evidence can be checked. Weak provenance is the most common reason a well-written page is not cited.
Evidence class: A label distinguishing what a report measured directly, what it observed at a timestamp, what it estimated from a model, and what cannot be known. Reporting without evidence classes makes inference indistinguishable from measurement.
Hallucination rate: The proportion of sampled answers containing a claim about a brand that contradicts a verified fact base. Measurable only if the fact base is defined in advance and the raw answers are archived.
Canary suite: A small set of pages or checks known to be healthy, run every monitoring cycle. When canaries fail together the fault is in the monitoring or a site-wide deploy, not in many separate pages.
AI Overviews: Google's generated summaries appearing above traditional results, assembled from pages already in the Search index. Eligibility follows from being indexed and retrievable by Googlebot, not from a separate opt-in.
Google-Extended: A robots.txt control governing whether content is used for Gemini model training and grounding. It is independent of Googlebot: blocking it does not remove a site from Google Search or from AI Overviews.

Why the definitions matter commercially

This is not pedantry. The absence of agreed meanings is the main reason buyers cannot compare tools in this category, and it is exploited more often than it is accidental.

Consider presence rate. One vendor samples fifty queries once each; another samples twenty queries five times each; a third counts brand mentions as well as linked citations. All three report a single percentage labelled the same way, and the numbers are not comparable in any direction. A buyer switching tools and seeing their score change has learned nothing about their visibility.

The same applies to anything described as a score. A composite number built from undisclosed weights cannot be audited, cannot be reproduced, and cannot tell an implementation team what to change. It is useful as navigation and misleading as a target.

Terms where the industry genuinely disagrees

Several entries above are stated more confidently than the field warrants, and it is worth flagging which ones.

GEO versus AEO. Used interchangeably by most people, distinguished sharply by a few. Neither usage is wrong; assume nothing about which one a given vendor means without asking.

Visibility. The least stable term in the category. It variously means linked citations, unlinked brand mentions, appearance anywhere in an answer, or a composite of all three. A visibility percentage without a definition attached is not a measurement.

Ranking. Frequently borrowed from SEO and applied to answer surfaces where nothing resembling a stable ranked list exists. Citation order within an answer is not a ranking, is not stable between runs, and should not be reported as one.

Optimisation. Sometimes describes making content genuinely more useful and extractable; sometimes describes attempts to manipulate retrieval. The first is durable and the second is not, and the same word covers both.

The questions that separate real measurement from decoration

When evaluating any claim in this category, including ours, four questions do most of the work.

What exactly was measured? Direct inspection of a page, a timestamped provider response, or an estimate derived from signals. These are different epistemic categories and a report that blurs them is not reporting.

What was the sample? How many queries, how many repeats per query, over what window, against which engines. Without this, a percentage has no confidence interval and no meaning.

Can I see the raw observations? Any aggregate should drill back to the exact archived answer, query, provider and timestamp that produced it. An aggregate that cannot be checked is decoration.

What does this explicitly not know? Nobody outside the model labs knows training corpora composition or internal ranking weights. A vendor who does not say so is either confused or hoping you are.

Primary references

Common questions

Is GEO the same thing as AEO?+

Most practitioners use them interchangeably. Where a distinction is drawn, answer engine optimisation emphasises being the direct answer and generative engine optimisation emphasises being a cited source within a generated response. The underlying work is the same.

Why do vendors define presence rate differently?+

Because there is no standard. Definitions vary in the query set, the number of samples per query, which engines count, and what qualifies as an appearance. Two honest tools can report different numbers for the same brand, which is why the methodology matters more than the figure.

What is the difference between a citation and a mention?+

A citation returns your URL as a source. A mention names your brand in the answer text without linking you. Both have value, but they are different measurements and combining them into one metric obscures more than it reveals.

Is llms.txt a standard?+

No. It is a proposed convention with growing but incomplete adoption. It is worth publishing because it is cheap and removes ambiguity for agents that read it, not because any engine requires it.

What does 'measured, observed, modelled, unknown' mean?+

It is an evidence classification. Measured comes from direct inspection of a page or file. Observed is a timestamped provider response. Modelled is an estimate derived from observable signals. Unknown covers private training data and internal rankings that nobody outside the labs can see.