VercatusGEO, AEO and LLMO: An AI Search Glossary with Measurement Definitions
Understand AI search terminology and distinguish access, coverage, retrieval, mentions, citations and business outcomes. Includes Vercatus's own terms: Answer Share and verdicts.
By Norbert Szabo, founder of Vercatus
Published · Updated · 6 min read
GEO, AEO and LLMO are overlapping industry labels for work intended to improve how content appears in AI answers. They are not interchangeable measurement standards. When evaluating a plan or report, ask what was observed, under which conditions, and how the metric was calculated.
How to use this glossary
- Terms are defined by what you can observe, not by what a vendor hopes to sell.
- Where a term has no standard meaning, the entry says what to ask.
- Vercatus’ own terms are marked as such.
SEO, GEO, AEO and LLMO
SEO: search engine optimization. Work on discoverability, indexing, relevance and the usefulness of pages in search. It includes technical delivery and content quality as well as measurement.
GEO: generative engine optimization. A label commonly used for improving visibility or source inclusion in generated answers. Check whether a particular provider means brand mentions, citations, referrals or content preparation.
AEO: answer engine optimization. Often used for work on answer-oriented search and assistants. Its scope overlaps GEO and SEO; the acronym alone does not define a product’s evidence or capabilities.
LLMO: large language model optimization. A less consistently defined label for influencing how model-based systems describe or retrieve information. Ask whether it concerns live web search, an internal retrieval system or model behavior without search.
Google’s guidance discusses GEO and AEO while treating optimization for its generative search experience as part of SEO. There is no separate acronym that exempts content from ordinary usefulness and quality requirements.
Access and readable content
Crawler access means a requester can fetch the intended content under its policy and delivery conditions. Specify the identity, origin, URL and date. A copied user-agent probe is not a verified vendor visit.
Readable content is a practical description of whether a system can obtain and interpret the page’s useful information. Access, extraction and content checks contribute evidence about that question; it is not a universal engine metric.
Render delta describes a difference between content obtained without scripts and after rendering under a stated test. It can reveal missing initial-HTML passages. The JavaScript guide explains its measurement limits.
Coverage, retrieval and citations
| Term | Meaning to establish before reporting |
|---|---|
| Answer coverage | Whether the available page content answers the question, under the test’s grading method |
| Retrieval | A page or passage was found or made available to a system |
| Mention | The answer names a brand or entity |
| Citation | The observed answer references or links a source |
| Owned-site citation | The cited source is on the site’s own domain |
| Third-party brand citation | Another publisher’s page about the brand is cited |
Retrieval is not citation, and a brand mention can occur without any owned-site source. A content simulation’s coverage score does not establish what a consumer assistant answered. See the measurement guide for evidence and denominators.
Prompt panels and share measures
Prompt panel means the selected questions repeatedly observed under defined settings. Its wording, market, engine, schedule and failed requests affect interpretation.
Cited share should identify its denominator, such as successful comparable answer observations. Other tools may use different populations.
Share of answer or voice needs a local definition. It might describe brand mentions among competitor mentions or the share of answers containing a brand. Never compare percentages before comparing those formulas.
Query fan-out refers to related searches generated to gather supporting information for a question. An editorial subtopic map is not an observed engine trace. The fan-out guide separates those concepts.
Agent surfaces and product-specific scores
An agent surface is an interface through which a system can read information or perform a task, such as a public page or a supported tool endpoint. Access to a page does not establish authorization to transact.
Answer Share, Vercatus’s measure, is the share of AI answers on a set of topics that name or cite a brand, shown with its range. It is not Google’s ranking score or a universal probability of being cited. The methodology explains its design.
Verdict, in Vercatus, is the reading of a shipped change against topics left alone: Won, Lost ground, No effect, Inconclusive or Too early. A win in answers does not prove incremental revenue. Referrals, qualified leads and causal business value belong in a separate outcome model.
More terms, defined by what you can check
Grounding. An answer engine’s use of retrieved web sources to support an answer. Engines expose two different things: the sources they retrieved or considered, and the sources a passage of the answer is attributed to. A source that was retrieved but not attributed was not cited for that passage (grounding and attribution, grade A).
Retrieval-augmented generation (RAG). A pattern where a system retrieves passages and gives them to a model to write the answer. Retrieval works on passages, not whole pages, which is why a passage that names its subject and states the answer early can be matched on its own (self-contained passages, grade B).
Search bot, training bot and user fetcher. Three purposes with three different tokens at most vendors: search visibility, model training and fetching a page because a user asked. Each robots.txt token is independent (search vs training bots, grade A). See the crawler guide.
Web Bot Auth. A way for an automated client to prove its identity by signing requests with HTTP Message Signatures. The signature identifies the client, not the page content (Web Bot Auth, grade A).
llms.txt. A proposal for a markdown file that points agents to LLM-friendly versions of key pages. Google’s AI optimisation guide says Search does not use such files and that they neither help nor hurt visibility (llms.txt, grade A). Whether another system reads one is a separate question to verify.
Entity home. The single authoritative page that defines an entity, a practitioner term; no published study shows how engines choose a defining page (entity home, grade D). The related, documented thing is where Google asks for organisation markup.
Earned media. Coverage by independent publishers, as opposed to paid placement. One analysis of citations found most went to earned media (earned media, grade C).
Vercatus terms. Topic: a cluster of buyer questions with the AI demand behind it. Engine Brief: what an engine says about a brand when asked, shown as perception. Play: the exact change with a guide, a draft and an acceptance check. Answer Share by engine: named and cited figures per engine, kept apart. Each is explained in the docs.
Questions to ask when someone promises “better GEO”
- Which question, which engine, which market and which date was observed?
- How was the percentage calculated, and what is the denominator?
- What changed on the site, and when did the engine see it?
- What would count as failure?
- Which business outcome is this meant to improve, and how is it measured?
Why the labels differ
The labels are new, they overlap, and they are used by people with different products to sell. Google’s guidance treats optimisation for its generative search experiences as part of SEO. Do not let an acronym stand in for a definition: ask for the observation behind the claim.
Use these definitions to make briefs and dashboards explicit. When someone promises “better GEO,” ask which task, source, observation and business outcome the work is intended to improve.