AEO and GEO glossary
Twenty-four terms from answer-engine optimization, defined plainly and without the vendor spin. If a definition here is wrong, tell us and we will fix it.
Last updated · 24 terms
- Answer engine optimization AEO
- The practice of getting a brand named and cited inside AI-generated answers, rather than ranked in a list of links. The defining constraint is that answer engines return a handful of options and omit everything else, so the goal is inclusion rather than position.
- See also GEO, answer engine.
- Generative engine optimization GEO
- A synonym for AEO, more common among practitioners who came from traditional SEO. There is no settled distinction between the two terms in practice; treat anyone drawing a sharp line between them as selling a taxonomy rather than a service.
- Answer engine
- Any system that responds to a query with a synthesised answer instead of a ranked list. In 2026 the ones that matter commercially are ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini, Claude, and Copilot.
- AI share of voice AI SOV
- The percentage of AI-generated answers in which a brand appears, relative to all brand mentions in its category across a sampled prompt set. The industry's default metric, and a flawed one — the sample is almost never disclosed.
- See hidden denominator for why this matters.
- Recommendation Rate
- The share of shortlist-class prompts in which a brand is explicitly named. Distinct from share of voice because it counts only the prompts where a buyer is actually asking to be given options — the moment where being omitted costs a deal.
- Narrative Position
- The framing a model applies when it names you: category leader, credible challenger, budget option, legacy incumbent, or niche specialist. Rarely measured and frequently the expensive problem — being named consistently as "the cheap one without enterprise controls" is worse than not being named at all in some deals.
- Prompt panel
- A fixed, documented set of prompts used as the stable denominator for repeatable measurement. A usable panel is versioned, dated, agreed before measuring, and published alongside every result derived from it.
- Shortlist prompt
- A prompt that explicitly asks the model to name options — "best X for Y", "top tools for Z". The highest-value prompt class, because the model is constructing the consideration set in real time.
- Head-to-head prompt
- A direct comparison query — "A vs B for use case C". Valuable because the answer usually establishes Narrative Position as well as presence, and because comparison content is disproportionately cited.
- Brand prompt
- A query naming the brand directly. Excluded from serious panels: you handed the model the answer inside the question, so a correct response measures nothing about discoverability.
- Grounding
- Anchoring a generated answer to retrieved source documents so claims can be attributed to something. Ungrounded answers come purely from the model's weights and are far harder to influence.
- Retrieval-augmented generation RAG
- An architecture where the system retrieves documents at query time and writes its answer from them rather than from memory alone. Where RAG is in play, fresh well-structured public content can affect answers within days.
- Parametric knowledge
- What a model recalls from its training weights without retrieving anything. Roughly 60% of ChatGPT answers are reported to come from parametric recall rather than live retrieval, which is why long-run reputation across the open web matters as much as recent publishing.
- Citation vs mention
- A mention is your brand name appearing in the answer text. A citation is a linked source the model attributes. You can be mentioned without being cited, and cited without being mentioned. They behave differently and should never be collapsed into one metric.
- Entity clarity
- How unambiguously a model can determine what your organisation is, who it serves, and what it competes with. Poor entity clarity is the most common fixable cause of omission — the model cannot recommend you for a job it does not know you do.
- Extractability
- How easily a passage can be lifted from a page and reused in an answer without losing meaning. Short self-contained answers, comparison tables, clear headings and definition blocks are extractable; long unstructured narrative is not.
- llms.txt
- A proposed plain-text file at a site's root giving language models a curated map of its most useful content, loosely analogous to robots.txt or a sitemap. Adoption by model providers is not guaranteed; the cost of publishing one is close to zero, so the expected value is positive regardless.
- Ours is at /llms.txt.
- Variance band
- The reported spread around a mean visibility figure. Model outputs are non-deterministic, so the same prompt run five times can name different vendors. A figure without a band is a sample of size one dressed as a fact.
- Model-version stamping
- Recording which engine and model version produced each data point, so that provider updates show up as labelled platform events rather than as unexplained client performance changes.
- Freshness signal
- Recency of update as an input to retrieval ranking. Content refreshed within the last 30 days has been reported to attract roughly 3.2× the citations of older material, making a refresh cadence one of the cheapest available levers.
- Zero-click answer
- A query resolved entirely inside the answer surface, with no visit to any source. The commercial consequence is that a mention without a click can still be the decisive event in a purchase, which breaks traffic-based reporting.
- AI presence score
- A composite 0–100 index used in published benchmark research to summarise how visible a brand is across answer engines. Useful for cross-category comparison; too coarse to manage a programme against, which is why we report three separate metrics instead.
- See the benchmark data.
Next step
Now see where you actually sit
A free finding: one real buyer prompt for your category, the verbatim answer, and who got named.