12 terms, defined

The vocabulary of AI search.

Most of the confusion in this field is vocabulary rather than substance. GEO, AEO and SEO get used interchangeably by people who mean different things; “entity” means one thing in a marketing deck and something much more specific inside a resolver. These twelve definitions are the ones we use when we talk about resolving a brand into a single entity, and they are written to be quoted rather than admired.

Each entry states the definition first, then argues with it. The definition is the part that has to survive being read on its own, with no surrounding page and no brand attached.

Generative Engine Optimization(GEO)

The practice of making a brand, its products and its expertise legible to systems that answer a question directly instead of returning a list of links — so the brand is named, quoted or cited inside the answer rather than ranked beneath it.

GEO and SEO share most of their inputs and almost none of their scoreboard. Both want a site a machine can crawl, parse and trust. But a ranking is a position on a page a person can scroll, and a citation is a sentence inside an answer they may never scroll past. Optimising for the first does not automatically buy the second, which is why a site can hold page-one positions and still go unnamed when someone asks an assistant the same question.

The term comes out of academic work published in 2023, and the industry has been repeating one of that paper's headline figures ever since without repeating its method. The study measured a simulated generative engine, not a live one. That distinction matters enough that we wrote it up separately.

In practice the work divides into three parts: making the content retrievable, making the entity resolvable, and making the claims quotable. The first is technical, the second is structural, and the third is editorial — and most programmes that stall have done the first and skipped the other two.

Also called: GEO · generative search optimization

Answer Engine Optimization(AEO)

Optimising content so it can be extracted as a direct answer to a specific question — the discipline that grew up around featured snippets, voice results and People Also Ask, and now feeds AI-generated answers.

AEO predates GEO and is narrower. It is concerned with a single question and the passage that answers it: the definition list, the table, the two-sentence paragraph directly beneath a question-shaped heading. GEO is concerned with whether the brand is a recognised entity in the field at all.

The two are frequently sold as the same thing. They are not, but they are complementary, and the plumbing turned out to be shared: the extraction machinery built for featured snippets is visibly reused by Google's AI answers rather than replaced by them.

AI Overviews

Google's AI-generated summary shown above the traditional results for some queries, assembled from multiple retrieved sources and displayed with links to a subset of them.

An Overview is not a ranking. It is a generated passage with citations attached, and the set of pages cited does not reliably match the set of pages ranking first for the same query. A page can be cited without ranking and rank without being cited.

Google has documented how content is retrieved for these answers in reasonable detail. It has never documented how the generator picks which retrieved sources get named in the visible citation strip, which is the part every publisher actually needs.

Query fan-out

The technique of expanding one user question into several related sub-queries, running them in parallel, and assembling a single answer from the combined results.

Fan-out is why optimising for one exact phrase has stopped paying what it used to. The system that answers 'which GEO agency should I hire in Toronto' may never issue that string as a query. It issues a spread of narrower ones — what the discipline is, what it costs, who provides it locally, what buyers complain about — and builds the answer from whatever each of those returns.

The practical consequence is that coverage beats precision. A page that answers one phrasing exceptionally well competes against a body of pages that collectively answer the whole fan. This is the mechanism that makes topic clusters worth building rather than a stylistic preference.

Retrieval-augmented generation(RAG)

An architecture in which a language model retrieves relevant documents from an external source at query time and generates its answer from those documents, rather than from its training weights alone.

Nearly every consumer AI product that cites sources is doing some form of RAG. Understanding the shape matters because it tells you where the leverage is: the model is not recalling your page, it is being handed a passage of your page moments before it writes. What gets handed over is decided by a retriever, and retrievers are not language models.

This is also where a persistent industry myth breaks down. Structured data is genuinely useful for classical search features and for entity resolution. It is not what most retrieval systems index when they chunk and embed a document, and selling schema markup as the lever that gets you into AI answers overstates it.

Chunking

Splitting a document into smaller passages so each can be indexed, embedded and retrieved on its own — meaning a retrieval system almost never handles your page as a whole page.

Chunking is the reason a long, well-argued article can be retrieved for a question it answers in one paragraph and ignored for the question it answers across four. Each passage is judged largely on its own merits, without the surrounding argument that makes it make sense.

Google's public guidance has told publishers not to write for chunking. Its own retrieval systems chunk anyway. Both things are true at once, and the resolution is not to write in disconnected fragments but to make sure each section can stand up if it is lifted out — a self-contained claim, its own context, and no pronoun pointing at something three headings earlier.

Also called: passage retrieval · passage indexing

Entity

A thing a search or AI system can identify and hold facts about — a company, a person, a place, a product — as distinct from the words used to refer to it.

The distinction is the whole point. 'LumiRank' is a string. The company it names is an entity, and a resolver's job is to decide which entity a given string refers to, or whether it refers to one it already knows at all. Get that wrong and every signal you earn gets filed against the wrong thing, or against nothing.

Most brands that are invisible inside AI answers are not being penalised. They are unresolved. The system has no confident basis for merging the company named on the website, the one named in a directory, and the one named in a press mention into a single thing it knows facts about.

Knowledge graph

A store of entities and the typed relationships between them — not a list of documents, but a network of things and the facts connecting them.

A graph is what lets a system answer a question no single document answers, by traversing relationships: this company provides this service, in this city, founded by this person, who wrote this article. Each edge is cheap on its own and expensive to fake in aggregate.

The work on the site's own side is unglamorous: declare each entity once, give it a stable identifier, and make every later reference point at that identifier instead of restating the entity inline. Two descriptions of the same company with no shared identifier give a resolver no reason to merge them, which is the most common way a carefully marked-up site still fails to resolve.

Structured data

Machine-readable statements embedded in a page — most often as JSON-LD using the schema.org vocabulary — that name what the page is about and how it relates to other things.

Structured data is a set of claims, not a ranking input. It says 'this page is about a service, provided by this organisation, serving this area'. It earns its keep by removing ambiguity a parser would otherwise have to guess at, and by making entity identifiers explicit.

It is routinely oversold. Markup will not make thin content retrievable and it is not the mechanism by which passages get pulled into AI answers. It is also routinely under-used in the one place it genuinely decides outcomes: consistent identifiers across every page, so the site describes one organisation rather than forty unrelated ones.

Also called: schema markup · JSON-LD

Semantic authority

The degree to which a site is treated as a credible source across an entire subject area rather than for individual keywords — earned by covering a topic completely, consistently and verifiably.

Authority in this sense is measured against a topic, not a phrase. A site with forty pages that each answer one narrow question, cross-link honestly, and agree with one another reads very differently from forty pages that each chase a separate keyword and contradict each other on price, scope and claim.

It is also the slowest of the levers, and the one most often abandoned halfway. Coverage compounds only if the pages stay consistent as they multiply, which is why we treat internal contradiction — two pages quoting different prices for the same thing — as a technical defect rather than a copywriting preference.

Also called: topical authority

See also query fan-out, entity, ai citation.

AI citation

A reference to a source inside a generated answer — the link or brand name an assistant shows as the basis for what it just said.

A citation is the closest thing GEO has to a ranking, and it behaves differently from one. Citations are unstable across repeated runs of the same prompt, vary by engine, and can name a source whose page never ranked for the equivalent query. Any measurement that samples a prompt once is measuring noise.

Being cited is also not the same as being recommended. An assistant may cite your page as evidence while naming a competitor as its answer. Both are worth tracking, and conflating them flatters the numbers.

Zero-visit search

A search that is fully resolved in the interface — the person gets their answer without opening any result — so the source that informed the answer records no visit.

Zero-click was the older framing and it undersold the problem. A zero-click result at least put a brand name on screen. The sharper version is a session that never reaches a search results page at all, because the question was asked of an assistant, answered, and closed.

The measurement consequence is the one that catches teams out: analytics cannot see it. A brand can be losing ground in the place buyers actually form opinions while every dashboard it owns reports flat. That gap is the entire argument for tracking engine answers directly rather than inferring them from traffic.

Also called: zero-click search

Knowing the words is not the same as being named

If you want to know whether the engines currently name you when someone asks a question in your category, that is a measurement, not an opinion — and it is the first thing we run.

Ask us what the engines say about you The discipline behind the vocabulary