ChatGPT Visibility and GEO Glossary

The terms

Terms are listed alphabetically. Where a term is central to how we measure, the entry links to the guide or to the audit method page that explains it fully.

AI visibility

How often, how prominently and how accurately a brand appears in answers generated by AI assistants such as ChatGPT, Gemini, Perplexity and Copilot. It is measured per prompt as a mention rate, a position and a citation rate, and summarized per prompt cluster. Unlike a Google ranking, AI visibility is probabilistic: the same prompt gives different answers on different days, so the measure is a rate across many answers rather than a single rank. See how to measure ChatGPT visibility.

Answer engine optimization

Shaping content so that a system which extracts direct answers, rather than listing links, can find, trust and quote it. The term predates chatbots and originally covered featured snippets and voice assistants. In practice AEO work means clear question-and-answer structure, concise definitions near the top of a page, structured data and consistent facts. It overlaps with GEO but focuses on the answer being lifted from your page, not on your brand being recommended among competitors. See GEO vs AEO vs SEO.

Baseline audit

The first full measurement of a brand’s ChatGPT visibility, taken before any work begins, against an agreed prompt cluster in an agreed market. It fixes the prompts, the collection conditions and the classification rules that every later re-audit must copy exactly. Without a baseline there is no honest way to claim movement. In our method the baseline is about 100 real answers per prompt per market, read and classified by hand, as described on the full method write-up.

Blue ocean channel

A marketing channel where buyers already gather but competitors have not yet learned to compete, so attention is available without bidding against everyone else. Google Autocomplete and ChatGPT recommendations are both examples at present: the buyer intent is real and growing, and most brands have no deliberate program for either. The term is borrowed from strategy writing and is used across the Odys Blue Ocean family of services. See ChatGPT recommendations vs Google Autocomplete.

Brand search lift

The increase in people searching your brand name on Google that follows being recommended elsewhere, including in ChatGPT. Buyers rarely click a link inside an AI answer; they take the name and go check it. A 2026 Idea Grove survey found 45% of consumers immediately Google a brand recommended by AI. Brand search volume is therefore the second signal to watch when attributing ChatGPT visibility. See tracking ChatGPT traffic and leads.

ChatGPT Search

The mode in which ChatGPT runs a live web search and composes its answer from the pages it retrieves, showing sources alongside the text. It is triggered automatically for many questions that need current or specific information, and can be forced by the user. Answers in this mode depend on what is on the web today and which pages ChatGPT selects, which rotate heavily. It is the second of the two routes by which a brand gets recommended; the first is model knowledge. See the two routes ChatGPT uses to recommend.

Citation

A link to a specific web page shown as a source for an AI-generated answer. In ChatGPT, citations appear as inline markers or a sources panel beneath the answer. A citation to your own domain means ChatGPT drew on your page; it does not by itself mean your brand was recommended, and a brand can be strongly recommended with no citation at all. Research summarized by everything-pr shows cited sources overlap little with Google’s top ten. See mentions vs citations.

Citation rate

The share of collected answers for a prompt that include a link to the brand’s own domain. It is tracked separately from mention rate because being named and being linked are different outcomes with different causes. A low citation rate alongside a high mention rate usually means ChatGPT knows the brand from third-party sources rather than from your site. It is one of the four fields recorded for every answer in a full audit. See the audit method.

Consensus signal

Agreement across many independent sources about what a brand is, what it does and how good it is. Language models weight claims that recur across reviews, press, directories, forums and the brand’s own site more heavily than claims found in one place. Consensus is why a single well-placed article rarely moves recommendations and why contradictory descriptions of your brand across the web hold you back. It is one of the things ChatGPT rewards, alongside recency and entity consistency. See what happens inside ChatGPT before it names a brand.

Consistency

In measurement, how reliably a brand appears across repeated answers to the same prompt and across the prompts in a cluster. A brand named in most answers for most prompts is consistently visible; one that appears sporadically is not, even if its best single answer looks impressive. Consistency is the quality that turns a lucky screenshot into a dependable channel, and it is what mention rate across about 100 answers is designed to capture.

E-E-A-T

Experience, Expertise, Authoritativeness and Trustworthiness: the qualities Google’s search quality guidelines ask human raters to assess. It is not a ranking factor in the mechanical sense, but it describes the signals that both Google and language models appear to reward: named authors with verifiable credentials, first-hand evidence, references, transparent company information and consistent reputation elsewhere. Pages that would score well with a Google rater tend to be the pages AI engines choose to cite.

Earned media

Coverage of a brand that it did not pay for or publish itself: press articles, independent reviews, analyst mentions, podcast appearances, forum discussions and listicles written by third parties. A University of Toronto audit found earned media made up 57% of GPT-4o citations, far more than brand-owned pages. Earned media is the main raw material from which ChatGPT builds its sense of which brands are credible in a category.

Entity

A distinct thing that a system can identify and attach facts to: a company, a product, a person, a place. For a language model, your brand is an entity only if the model has a stable picture of what it is, what it does, where it operates and how it relates to other entities. Brands with weak or fragmented entity signals are often confused with similarly named companies or simply not retrieved. See LLM entity consistency.

Entity consistency

Describing your brand the same way everywhere: the same name, the same category, the same one-sentence description, the same founding facts, the same locations, across your site, directories, social profiles, press and reviews. Inconsistency fragments the entity in a model’s view and lowers confidence in any single claim about it. Entity consistency is among the most controllable things a brand can fix and is covered in depth in the guide on making your brand legible to language models.

Generative engine optimization

The practice of increasing how often and how favorably a brand is named in answers generated by AI engines such as ChatGPT. GEO draws on SEO, PR, review management and content strategy, but its success measure is different: a mention rate and position inside generated answers rather than a rank in a list of links. The term is often abbreviated GEO. See the full guide, what GEO is and what actually moves ChatGPT, and the GEO vs SEO comparison.

Grounding

Tying a generated answer to specific retrieved sources rather than to the model’s general training alone. When ChatGPT searches the web and composes an answer from the results, the answer is grounded in those pages. Grounding reduces but does not eliminate error: a Stanford study found 50 to 90% of LLM responses were not fully supported by their cited sources. For brands, grounded answers are the ones most influenced by what is on the web this month.

Hallucination

A confident statement by a language model that is false or unsupported: a product feature you do not offer, a location you do not serve, a price that was never yours, or a brand that does not exist. Hallucinations about your brand are a reputation risk and are more likely when the model has little consistent information to draw on. The remedy is more and clearer facts in more places, not complaints to the model. See how ChatGPT answers the safety question about a brand.

Knowledge graph

A structured database of entities and the relationships between them, such as “Company A is headquartered in City B and makes Product C”. Search engines maintain large knowledge graphs, and language models learn similar relationships implicitly from text. Having your brand’s facts represented clearly, through structured data on your own site and consistent third-party listings, helps both kinds of system place you correctly in a category and alongside the right competitors.

llms.txt

A proposed plain-text file placed at the root of a website that summarizes what the site is and points to its most important pages, intended to help AI systems understand the site efficiently. Adoption by the major engines is uncertain and no engine has committed to reading it as a ranking input. It is cheap to add and harmless, but it is a tidy-up rather than a strategy; the facts on your actual pages matter far more.

Mention rate

The share of collected answers for a given prompt in which the brand is named at all. It is the base measure of ChatGPT visibility. Because answers vary, mention rate is only meaningful across a large sample: with about 100 answers it comes with a confidence band of several points either side, which the full method write-up explains. Mention rate is tracked per prompt and summarized per cluster. See how to measure ChatGPT visibility.

Model knowledge

What a language model has absorbed during training and can recall without searching the web. When ChatGPT answers a buying question without triggering a search, the brands it names come from this knowledge, which reflects how often and how consistently each brand appeared in its training data up to the training cutoff. Model knowledge changes only when the model is retrained or updated, so it moves slowly compared with ChatGPT Search. See the two routes ChatGPT uses to recommend.

Position

Where a brand appears in the order ChatGPT names brands within an answer, with the first named as position one. Buyers shortlist from the top, so a brand named first or second in most answers has a very different outcome from one named sixth. Position is recorded in every answer we classify and reported as an average and a median per prompt. See why only the first three names count.

Prompt cluster

A set of related prompts, typically 20 to 40, that express one commercial intent in the different ways real buyers phrase it: direct asks, comparisons, situational descriptions, local and trust questions. Visibility is measured and reported at the cluster level because no single prompt represents how a market asks. The cluster is agreed before a baseline audit and held fixed for every re-audit. See prompt clusters: how buyers actually phrase the question.

Prompt variation

A rewording of a prompt that keeps the same intent: “best CRM for a small agency”, “which CRM should a 10-person agency use”, “CRM recommendations for small marketing firms”. Variations matter because ChatGPT can name different brands for different phrasings of the same need. A good prompt cluster deliberately includes variations so the mention rate reflects the whole intent rather than one lucky or unlucky wording. The ChatGPT prompt generator tool produces five variations for a category, location and buyer type.

Re-audit

A repeat of the baseline audit under identical conditions: same prompts, same market, same connection type, same classification rules, similar sample size. In our method the first re-audit runs at day 30 and then monthly. Changes are reported with their confidence band so a real shift can be separated from noise. A re-audit that changes any condition is not comparable and should not be presented as progress. See the audit method.

Recency

How recently a source was published or updated. AI engines favor fresh content when they search: a University of Toronto audit found a median cited-source age of 62 days for Claude versus 130 days for Google in consumer electronics. For brands this means that coverage from two years ago does less work than coverage from this quarter, and that maintaining a steady flow of current third-party mentions matters more than one old feature.

Retrieval

The step in which a system fetches relevant documents from the web or an index before composing an answer. In ChatGPT Search, retrieval decides which pages the model reads, which in turn largely decides which brands it can name. Retrieval is query-dependent and changes as the web changes, which is one reason cited sources rotate so heavily. If your pages are not retrieved for buying questions in your category, no amount of good writing on them will be seen.

Retrieval augmented generation

The architecture in which a model retrieves documents first and then generates an answer conditioned on them, usually abbreviated RAG. ChatGPT Search is a form of it. RAG lets a model answer about things after its training cutoff and cite sources, at the cost of making answers depend on what retrieval returns on a given day. For brands, RAG is why what is on the web this month can change recommendations faster than model knowledge alone would.

Sample size

The number of answers collected per prompt before a rate is computed. Because ChatGPT answers vary, a small sample gives an unreliable rate: three answers cannot distinguish 30% from 70%. Our audits collect about 100 answers per prompt per market, which yields a confidence band of several points either side rather than a precise figure. Larger samples narrow the band at higher cost. See the limits section of the full method write-up.

Share of voice

A brand’s portion of all brand mentions across the answers in a cluster, measured against every competitor named. If ChatGPT names four brands per answer on average and yours is one of them in most answers, your share of voice is high even if you are rarely first. Share of voice shows who you are displacing and who is displacing you, which mention rate alone does not. It is reported for the five most-named competitors in each cluster.

Sponsored card

The labeled advertising unit OpenAI introduced in ChatGPT in 2026, shown beneath the answer’s sources for logged-in US adults on the free tiers, with paid tiers excluded, according to third-party tracking. The recommendation inside the answer itself is organic and not for sale; the Sponsored card sits below it. See how ChatGPT ads differ from organic recommendations.

Structured data

Machine-readable markup, usually schema.org in JSON-LD, that states facts about a page or an organization explicitly: name, address, founding date, products, prices, reviews, authors. Structured data helps search engines and, by extension, the retrieval systems that feed AI answers to read your facts without guessing. It is one of the more controllable entity-consistency measures and one of the few places where you state your own facts in a form machines are built to trust.

Training cutoff

The date after which a language model has seen no new data. Anything that happened after the cutoff is unknown to model knowledge and can only reach an answer through live retrieval. A brand founded, renamed or repositioned after the cutoff will be absent or described in its old form unless ChatGPT searches. This is why newer brands depend more heavily on ChatGPT Search and on the recency of third-party coverage than established ones do.

utm_source=chatgpt.com

The tracking parameter ChatGPT appends to links it shows, which lets analytics tools such as GA4 attribute a visit to ChatGPT. It is the first attribution signal to watch. Referral volume from AI is still small, 0.32% of all website traffic in 2026 according to SE Ranking, so the second signal, brand search lift, usually carries more of the value. See our ChatGPT attribution guide.

Variance

The natural spread in answers to the same prompt, caused by the model sampling from a range of plausible responses and by the retrieval layer changing day to day. Variance is why one answer proves nothing and why mention rate needs a confidence band. It also means a brand’s true visibility can look like it moved a few points between months when nothing changed. Honest reporting names the band; the full method write-up explains how we report it.

Zero-click

A search or question that ends without the user clicking any link, because the answer itself was enough. AI answers are zero-click by design: the buyer gets a shortlist and leaves. The commercial consequence is that the recommendation, not the visit, is the product, and the next observable step is often a Google search for the recommended brand. See how the ChatGPT buyer journey differs from Google’s.