The AI search glossary.

The practice of getting a business named inside AI-generated answers.

Where classic SEO optimizes a page to rank in a list of links, GEO works to make the business the answer an assistant gives. The difference matters because the levers are different: GEO depends far more on what independent third-party sources say about a business than on anything published on the business's own website.

The share of AI-generated answers to buyer-intent questions that name a given business, measured out of the answers actually received.

A business named in 14 of 96 completed answers has 15% AI visibility. The measurement only means something when the question set is fixed and the denominator is honest, see denominator integrity.

A business's mentions expressed as a fraction of every business mention across a set of AI answers.

It measures how crowded a category is. A business can hold decent visibility and still hold low share of voice, because each answer that names it also names four competitors. Visibility says whether you are in the conversation; share of voice says how much of it you own.

An AI answer that links a business's own domain, as opposed to merely naming the business.

Being mentioned earns a moment of attention. Being cited compounds, because the model returns to that domain as a reference on later questions. Most businesses are mentioned far more often than they are cited, and closing that gap is slower work than getting mentioned in the first place.

Google's AI-generated summary shown above the traditional blue links.

It answers the query directly and typically names only two or three businesses. That is the structural change worth understanding: ten blue links gave a dozen businesses a shot at the click, and ranking eleventh was survivable. An answer that names three businesses has no page two.

Supplying a language model with live retrieved web results at the moment it answers, rather than relying on its training data.

A grounded answer reflects what is on the web today. An ungrounded one reflects whatever the model absorbed during training, which may be a year stale. Any credible AI visibility measurement runs with grounding enabled, because ungrounded answers measure the model's memory rather than the market.

A question phrased the way a real customer would type it into an assistant, used to test AI visibility.

Buyer-intent prompts are full sentences describing a need, not keyword strings, because that is what people actually send to an assistant. "Best med spa in Fort Lauderdale" is one. "med spa fort lauderdale fl best" is a search-box habit that nobody uses in a chat window.

A search in which the customer describes a symptom rather than naming a service.

"My roof is leaking after the storm, who should I call" instead of "roofing contractor". It is the hardest query family to win and usually the most valuable, because the customer has not yet decided what they are buying, which means whoever the assistant names gets to define the purchase.

A plain-text file at a website's root that summarizes the site for large language models, in the way robots.txt addresses crawlers.

It states what the organization does and maps the pages worth reading, so a model can grasp the site from one fetch instead of inferring it from markup. It is a convention rather than a standard, and no engine is obliged to read it, but it costs almost nothing to publish. This site's llms.txt is a working example.

The process by which a search or AI system decides that a name appearing in several places refers to one real organization.

Consistent names, addresses and cross-links between a business's own properties and its third-party profiles are what let a model merge scattered mentions into a single confident entity. Until that merge happens, every mention is diluted across several half-entities, and the model hedges rather than recommending.

Using an identical business name, address and phone number everywhere the business appears online.

Inconsistencies, a name spelled as two words in one place and three in another, a suite number present on one listing and absent on the next, split what should be one entity into several weaker ones. Local answers are the first place the damage shows, because they lean hardest on map and business-profile data.

Excluding failed requests from the total when scoring AI visibility, rather than counting them as misses.

If a scan attempts 100 searches and 4 time out, the score is calculated out of 96, because a timeout is evidence of a network problem, not of invisibility. Scoring failures as misses makes a number look worse than reality and makes week-to-week trends meaningless, which defeats the purpose of measuring at all.