A glossary of AI citation and AI search

AEO is the practice of writing and structuring content so that systems which answer questions directly — rather than returning a list of links — can extract and attribute a passage from it.

The shift it names is real: when a reader gets an answer instead of ten blue links, being the source of the answer matters more than being the fourth result.

GEO is optimising content for systems that generate an answer from retrieved sources, with the goal of being one of the sources the answer is built from.

In practice GEO and AEO describe the same work. Treat any claimed distinction as vocabulary rather than method.

SEO is the practice of making a page rank highly in a list of search results.

Ranking and being cited are different outcomes with different requirements, and a page can do one without the other. Ranking first is about competing against other pages; being cited is about whether a passage can be lifted out and quoted.

Retrieval is the step where a system searches for documents relevant to a question and pulls back a set of candidate passages before any answer is written.

A page that is not retrieved cannot be cited, no matter how well written it is. Crawler access and topical clarity both act here, before quality is ever assessed.

RAG is the pattern of retrieving relevant documents first and then generating an answer conditioned on them, rather than relying only on what the model learned during training.

It is why fresh pages can be cited by a model whose training ended long before they were published — and why crawler access matters more than training-data inclusion.

Grounding is tying a generated statement to a specific retrieved source, so the claim can be attributed rather than asserted.

Grounded answers are the ones that carry citations. Content that makes clean, self-contained factual statements is easier to ground than content that hedges.

Chunking is splitting a document into smaller passages so that individual sections can be retrieved and quoted independently of the whole page.

This is the mechanical reason structure matters. A section under a descriptive heading survives being separated from its page; a point spread across four paragraphs does not.

Extractability is the degree to which a passage can be lifted out of a page and still make sense on its own.

It is usually the difference between two pages of equal quality where one gets cited. A sentence that depends on the three paragraphs above it cannot be quoted.

Query fan-out is when a system turns one user question into several different searches, then assembles an answer from what all of them return.

It means you are rarely competing for one phrase. A page that answers a cluster of adjacent questions has more chances to be retrieved than one tuned to a single keyword.

An AI crawler is an automated client operated by an AI company that fetches web pages, identified by a user-agent token such as GPTBot or PerplexityBot.

They are not interchangeable. Each operator runs several with different purposes, and a robots.txt rule applies only to the exact token it names.

A training crawler collects pages that may be used to train future models. GPTBot, ClaudeBot, Google-Extended and CCBot are examples.

Blocking one is a legitimate decision about your work being used as training data. It does not affect whether you are cited today.

A retrieval crawler fetches pages so an assistant can cite them in an answer it is composing. OAI-SearchBot, Claude-SearchBot and PerplexityBot are examples.

Blocking one removes the page from consideration for live citations entirely. This is the block that is usually made by accident, in a rule meant to prevent training use.

robots.txt is a file at the root of a domain that tells automated clients which paths they may request, using rules grouped by user-agent.

It is a request rather than a control: well-behaved crawlers follow it and others ignore it. Among matching rules the longest pattern wins, so a broad Disallow can be narrowed by a more specific Allow.

llms.txt is a proposed file at the root of a domain that points language models at the pages a site considers most useful, in Markdown.

Adoption is not universal and it is not a ranking mechanism. Treat it as cheap and harmless rather than as a lever.

AI Overviews are generated summaries that Google places above traditional results for some queries, with links to the sources they drew on.

Appearing in one is a separate outcome from ranking first, and the two do not reliably predict each other.

A citation is a named or linked reference to a specific page, given by an assistant as a source for something it said.

It is the outcome worth measuring, because it is verifiable. A mention of your brand without a link is not a citation and does not send anyone anywhere.

A zero-click result is one where the reader gets what they needed from the answer itself and never visits a source.

It reframes what a citation is worth: often attribution and recall rather than a session. Judge the channel on whether you are named where it counts, not only on referral traffic.

Hedging is qualifying a claim with language like "might", "in some cases" or "it could be argued" so that it no longer asserts anything definite.

Heavily hedged writing is hard to quote, because there is no sentence that says something. Keep the qualifiers a claim genuinely needs and cut the rest.

First-hand evidence is data, testing, or experience the author produced themselves rather than gathered from other sources.

It is the strongest answer to the question an assistant implicitly asks: why cite this page instead of the sources it is summarising? A page that only restates its citations gives no reason to name it.

Synthesis is content assembled from existing sources without adding original data, testing or judgement.

Not automatically a weakness — synthesis with real interpretation is valuable. Synthesis that only rearranges its sources competes directly with them and usually loses.

Quotability is whether a page contains sentences that can be reproduced verbatim, attributed, and still be true and comprehensible.

It rewards specificity: a named figure, a concrete claim, a clear definition. Vague or promotional language is rarely quotable even when it is accurate.

E-E-A-T is the set of qualities described in Google's search quality rater guidelines as signals of content worth trusting.

It is a framework for human raters, not a score in a system. The underlying signals — a named author, visible dates, checkable sources — are worth having regardless of what any one search engine does with them.