AI search glossary: the terms, defined plainly

This category generates acronyms faster than it generates evidence. Here is what each term actually means, written so you can read a vendor proposal without a decoder ring.

Terms are grouped by what they describe rather than alphabetically, because the relationships matter more than the spelling.

The optimisation acronyms

SEO, Search engine optimization. The original discipline: making pages crawlable, relevant and authoritative enough to rank in a list of results. Still the largest source of commercial organic traffic for most businesses.

AEO, Answer engine optimization. Optimising to be the single direct answer rather than one of ten links. Predates the current AI wave, it grew up around featured snippets, “People also ask” and voice assistants.

GEO, Generative engine optimization. Optimising to be cited inside an answer that a model composes from several sources, rather than extracts from one.

LLMO, Large language model optimization. A newer label, broadly interchangeable with GEO. Mostly a marketing distinction.

AIO, AI Overviews. Google’s generated summary that appears above the conventional results. Also used loosely as shorthand for “AI optimisation”, which causes regular confusion. See how to show up in AI Overviews and AI Mode.

AI Mode. Google’s conversational search experience for questions that need exploration, reasoning or comparison. It returns a generated answer with links to supporting sites. Google says no special optimization is required to appear in it beyond being indexed and eligible for a snippet.

Query fan-out. The technique behind AI Overviews and AI Mode: one question is split into several related searches across subtopics, and the answer draws on the best source for each part. It is why a set of focused pages, each answering one follow-up question, competes better than one general page.

If a proposal prices SEO, GEO, AEO and LLMO as four separate workstreams, ask which specific tasks belong to each. There are genuine differences but they are a minority of the total work.

How the systems work

LLM, Large language model. The model that generates text. Trained once on a large corpus, then updated periodically. On its own it has no knowledge of anything after its training cutoff.

Grounding. Attaching a model’s answer to retrieved source documents rather than letting it answer purely from training. Grounded answers are the ones that carry citations and the ones your content can influence.

RAG, Retrieval-augmented generation. The common technical pattern behind grounding: search for relevant documents first, then have the model write an answer using them. Most cited AI answers you see are some variant of this.

Training data vs retrieval. Two different routes into an answer. Training data is fixed until the next model version and effectively impossible to influence on any useful timescale. Retrieval happens live at query time and it is the one that responds to work you do this quarter. Almost all practical AI visibility work targets retrieval.

Hallucination. A model stating something false with confidence. Relevant to you because models hallucinate about businesses, wrong services, wrong locations, wrong pricing. Correcting the public record is a real part of this work.

Knowledge cutoff. The date after which a model has no training knowledge. Why a model may describe your business as it was two years ago unless it retrieves something current.

Crawling and access

AI crawler. A bot that collects web content for an AI system. Governed separately from Googlebot.

The ones worth knowing:

User-agent	Operator	Purpose
GPTBot	OpenAI	Collects content for model training
OAI-SearchBot	OpenAI	Collects content for ChatGPT search results
ChatGPT-User	OpenAI	Fetches a page when a user’s request requires it
ClaudeBot	Anthropic	Collects content for Claude
PerplexityBot	Perplexity	Indexes content for Perplexity answers
Perplexity-User	Perplexity	Fetches a page when a user’s question requires it
Google-Extended	Google	Controls use in Google’s AI products, separately from search indexing

robots.txt. The file that tells crawlers what they may fetch. Two things people get wrong: it controls crawling, not indexing, a blocked URL can still appear in results and it is honoured voluntarily. Treat it as a signal, not a lock.

llms.txt. A proposed convention for a plain-text file summarising a site for language models. It is a community proposal, not an adopted standard and no major AI platform has committed to reading it. Harmless to publish; be sceptical of anyone presenting it as a requirement or a shortcut to citations. More in llms.txt explained.

Measurement

Share of AI voice. The proportion of your tracked prompts where your brand is named, usually relative to named competitors. Only meaningful if the prompt set and run count are disclosed.

Prompt set. The defined list of questions used for measurement. The equivalent of a keyword list but written the way people actually talk to assistants.

Citation rate. How often you are not just mentioned but linked or explicitly attributed. Varies enormously by platform, Perplexity cites heavily, others less consistently.

AI referral traffic. Visits arriving from an AI assistant, identifiable by referrer in analytics. A floor, not a measurement: it counts only those who clicked and attribution from these platforms is unreliable.

Entity and structure

Entity. The thing a search system believes your business is, distinct from the words on your page. Search and AI systems reason about entities, not strings.

Entity consistency. Describing your company the same way across your site, your profiles and third-party listings. Inconsistency makes a model less willing to assert anything specific about you.

Structured data / schema markup. Machine-readable annotations, usually JSON-LD, describing what a page is about. Must match the visible content, marking up claims that are not on the page is a guideline violation, not a clever trick.

Knowledge panel. The brand box in Google results. A visible output of entity recognition, not something you can buy or directly configure.

E-E-A-T. Experience, Expertise, Authoritativeness, Trustworthiness. Concepts from Google’s quality rater guidelines. Not a score in the algorithm and not a ranking factor you can set, a description of what the systems are trying to approximate.

Terms to treat carefully

“Guaranteed AI citations.” Not possible. These systems do not publish ranking factors, change frequently and personalise results.

“AI visibility score.” Fine as shorthand, meaningless unless the vendor discloses the prompt set, the assistants tested and the number of runs.

“Optimised for ChatGPT.” ChatGPT has no published optimisation specification. What is real: making your content retrievable, parseable and quotable. What is not: a secret configuration.

---

Next in the Academy: What generative engine optimization actually involves.