GEO glossary

Generative Engine OptimizationGEO

The practice of improving how often, and how accurately, a brand is named inside answers written by generative AI systems. It covers both the measurement of how a brand currently appears and the work that changes it. GEO is not a ranking exercise: there is no position to win, only an answer that either names you or does not.

Answer Engine OptimizationAEO

The work of getting a direct answer surfaced rather than a list of links. The term predates generative engines and covers featured snippets, voice assistants and question boxes. GEO overlaps with it heavily, the difference being that an answer engine selects an existing passage while a generative engine writes a new one.

Retrieval

The mechanism by which an engine runs a search at the moment the question is asked, reads the pages it gets back, and writes an answer grounded in what it just read. Because it happens live, a change to a page can show up within days. Which pages get retrieved is heavily influenced by conventional search visibility.

Parametric memory

What a model absorbed during training and can restate without looking anything up. It has a cutoff date, and models are retrained infrequently, so a change made today may not surface for many months, and may never surface if the brand is not discussed on sources that make it into future training sets.

Retrieval-augmented generationRAG

The architecture that combines the two mechanisms: the system retrieves documents, then conditions the answer it writes on them. Most production assistants now work this way. For a brand it means the fast lever and the slow lever are both live at once, and only the fast one can be acted on inside a quarter.

Grounding

Tying a generated sentence to a specific source the system actually read, rather than to the model's general impression. A grounded answer can be traced and corrected. An ungrounded one about your brand is the harder problem, because there is no document to fix.

Citation

A named reference inside an answer, either to a brand or to a source. The two are not the same thing and are worth counting separately: being named as a recommended provider is the commercial outcome, while being cited as a source is what gets a domain read again the next time a similar question is asked.

Visibility rate

The share of a fixed prompt set in which the brand is named at all. It is the first number to establish, because everything else is conditional on it. A rate of zero is common for a young brand and is not a failure of the method, it is the measurement working.

Share of voice

Among the brands an engine does name for a given question, the proportion of those mentions that are yours. It separates two very different situations that a visibility rate alone confuses: a market where the engine names nobody, and a market where it names your competitors instead of you.

Prompt set

The fixed list of questions a measurement is run against, written to match what buyers actually type rather than what a brand wishes they typed. It has to stay identical between cycles. Change the questions and two reports are no longer comparable, which is the most common way a GEO measurement becomes decorative.

Source set

The domains an engine leans on when it answers a given question. It is usually a short and stable list, and it is the most actionable output of a measurement: it names the handful of places where a brand has to exist before an engine will repeat anything about it.

Description accuracy

Whether the sentence an engine writes about a brand is correct, not merely present. A brand can be named in half the answers and still lose the deal if the engine has the wrong sector, the wrong country or the wrong size. Accuracy is scored separately from visibility for that reason.

Hallucinated attribute

A specific, confident claim about a brand that no source supports: an invented office, a fabricated client, a wrong founding date, a competitor's product credited to you. It spreads because an answer that sounds precise is rarely questioned, and it is the case for publishing a small set of facts a model can find and repeat.

Structured data

Machine-readable statements embedded in a page, usually as JSON-LD following the schema.org vocabulary. It does not make a page rank and it does not force a model to repeat anything. What it does is remove ambiguity about what the page is and which entity it describes, which is exactly the ambiguity that costs a brand its mention.

Entity

A thing a system can hold an opinion about: a company, a person, a product, a place. Entity work is the business of making sure a single consistent thing exists across every source, rather than three half-descriptions a model cannot merge. Two companies sharing a name is the hardest version of this problem.

Knowledge graph

A store of entities and the relationships between them, maintained by a search or AI provider. A brand does not edit it directly. It is influenced by publishing consistent, corroborated statements across independent sources, which is why the same facts have to appear on the site, on the company profiles and in the structured data.

Passage extraction

The unit a generative engine lifts is a passage, not a page. In practice that means a heading stating a question followed immediately by a self-contained answer of roughly a paragraph. Too short and there is no claim worth quoting; too long and the engine has to summarise, which is where the brand name gets dropped.

llms.txt

A plain-text file at the root of a site, proposed as a way to give AI systems a short, curated description of what the site is and where its important content lives. It is a convention rather than a standard, and no engine is obliged to read it. It costs almost nothing to publish and it fails honestly: if nobody reads it, nothing breaks.

AI crawler

The automated readers AI providers send out, among them GPTBot, ClaudeBot, PerplexityBot and Google-Extended. They are declared separately from the classic search crawlers, so a site can be perfectly visible in search and invisible to an assistant simply because robots.txt blocks them. Checking this is the cheapest first move in any GEO engagement.

Search engine optimizationSEO

The work of earning position in a ranked list of links. It shares most of its technical base with GEO, because a page an AI crawler cannot read is a page no engine can quote. What it does not share is the objective: a generated answer has no positions to occupy.

AI Overviews

Google's generated answer shown above the classic results. It is retrieval-driven, so it responds to the same work that earns conventional search visibility, and it is the one surface where SEO and GEO effort overlap almost completely. GenesisHub measures it as one of the five engines it covers.

Prompt engineering

Writing the instruction given to a model in order to get a better output. It is a user-side skill and it is not GEO, which is about what a model says when somebody else does the asking. The two are confused often enough to be worth stating plainly: you cannot prompt your way into another person's answer.