Quick answer

SEO makes content accessible, understandable, and competitive within search systems. GEO, as the industry uses the term, refers to the effort to make a brand and its evidence retrievable and citable in generative answers. AEO aims at answers that are directly usable by both engines and people. These three disciplines overlap, but neither GEO, AEO, nor LLMO is an official standard with a universal metric.

This glossary distinguishes three statuses: official term when a platform or specification defines it; open standard when a public standard governs its meaning; industry usage when the market uses several different definitions. For every term, always ask what object is being observed, the denominator, the platform, the period, and the source. An "AI citation" can mean a displayed link, a source used for grounding, or a simple mention, depending on the provider. Without an operational definition, two dashboards can show incompatible scores while both being mathematically correct.

Key takeaways

  • SEO, GEO, and AEO share the fundamentals of crawling, quality, clarity, evidence, and measurement.
  • A mention, a citation, an impression, a click, and a conversion are five different events.
  • Schema.org is a vocabulary; Google then applies its own eligibility rules for features.
  • The terms GEO, AEO, LLMO, and "AI share of voice" are only useful if their method is stated.
  • A usable definition includes an example, a unit of measurement, and a limitation.

Concept map: from the page to the decision

Layer Question Main objects Common mistake
Discovery Can the system find the resource? link, sitemap, crawler, robots.txt Believing a sitemap forces indexing
Understanding Are the content and its entity explicit? visible text, canonical, structured data, internal links Adding markup that contradicts the page
Selection Does the resource answer the need in this context? intent, relevance, quality, evidence Reducing selection to a keyword
Generation How is an answer built and sourced? LLM, retrieval, grounding, fan-out Assuming a single stable ranking
Observation What event actually happened? mention, citation, impression, click Merging incompatible units
Outcome Did a decision or value follow? conversion, revenue, support Causally attributing every outcome to a citation

Google publicly explains the broad stages of crawling, indexing, and serving, while stating that a compliant page isn't guaranteed to be crawled, indexed, or shown (how Google Search works). That distinction should stay the foundation of every term below.

The 45 definitions

1. Crawling : official term

The process by which a robot discovers and downloads URLs. Discovery can come from links, sitemaps, or already-known URLs. Being crawled doesn't mean being indexed. Measure the activity with server logs and engine reports, accounting for caches, spoofed user agents, and reporting delays.

2. Indexing : official term

The process by which an engine analyzes a resource and may add it to its index. A crawled URL can be excluded for duplication, noindex, quality, error, or another system decision. "Indexed" means neither visible for every query nor well ranked. Check the URL, the retained canonical, and the observation date.

3. Ranking : official term, context required

The ordering of results judged relevant for a query and a context. It varies by location, language, device, freshness, and the engine's own systems. A position tracked by a tool is an observation of its environment, not a permanent property of the page.

4. SERP : established industry usage

Short for search engine results page, the page of results an engine returns. It can contain classic links, ads, cards, videos, rich snippets, and generative answers. Comparing positions without recording which modules were present can hide the real visual space and click opportunity.

5. Query : official term

Text or another input sent to a search system. It can be short, conversational, spoken, or multimodal. The logged query doesn't always express the person's full intent. In an AI panel, keep the exact text, language, account, date, and any rephrasing.

6. Search intent : industry analysis model

The need an analyst attributes to a query: to learn, compare, act, navigate, or resolve a task, for example. It isn't a label the user directly reveals. The same query can carry several intents; validate them against the results page, journey data, and interviews rather than the keyword alone.

7. Keyword : industry usage

An expression chosen to represent a theme or a set of queries. A keyword helps organize research and measurement, but systems don't limit themselves to a literal match. A useful page covers the decision, the entities, and the evidence needed; artificially repeating the phrase isn't optimization.

8. Entity : information-retrieval concept

A distinct object that can be identified: a person, an organization, a product, a place, or a concept. An entity has attributes and relationships, unlike a plain string of text. Disambiguate it with the text, consistent names, reference pages, and, where relevant, structured identifiers. Markup alone doesn't "create" authority.

9. Knowledge graph : technical term

A structure linking entities through relationships and properties. Several engines and organizations maintain their own graphs; there's no single public "Google graph" you can edit directly. The knowledge can come from many sources and change. Use the concept to model relationships, not to promise a display outcome.

10. Canonical URL : official term

The URL considered representative among several identical or near-identical pages. rel="canonical" is a signal, not an absolute directive; the engine can choose a different URL. Align internal links, sitemap, redirects, and content. Check the declared canonical and the one actually selected instead of assuming the tag was honored (Google's guidance on canonicalization).

11. robots.txt : open standard

A file at the root of a host that communicates access rules to compliant robots. It governs crawling, not a guaranteed removal of a URL from the index, and it's public. Never put a secret in it. The protocol is standardized by RFC 9309 (Robots Exclusion Protocol).

12. noindex : engine directive

An instruction asking that a page not be indexed, usually via meta robots or an HTTP header. The engine must be able to access the resource to see the directive; blocking crawling at the same time can prevent it from being read. Check the final URL, the header, and the render, then allow time for a new crawl.

13. XML sitemap : supported de facto standard

A file listing URLs a site wants to flag to engines, with optional metadata. It helps discovery and tracking, but doesn't guarantee crawling or indexing. Only include canonical, indexable, useful URLs; segment sitemaps to diagnose page families.

A link from one page to another resource on the same site or defined ecosystem. It helps both people and robots discover content and understand relationships. Effective internal linking connects a definition to its guide, its measurement, and its next action; thousands of automated, non-contextual links reduce that value.

A link received from another site. Its value isn't just a tool's score: relevance, context, accessibility, attribute, origin, and naturalness matter. A backlink can bring discovery, audience, or a reputation signal, but no third-party metric guarantees its effect. Links bought to manipulate ranking can violate spam policies.

16. Structured data : standard and platform policies

Information encoded according to a vocabulary, often Schema.org, to express entities and properties in a machine-readable way. Google generally recommends JSON-LD and requires that markup represent the visible content when it powers its features. There's no magic tag for being cited by an AI.

17. Core Web Vitals : official metrics

A set of experience measures: LCP for main loading, INP for responsiveness, and CLS for visual stability. Thresholds are evaluated at the 75th percentile of visits, separately by relevant context. Field and lab data have different uses; passing the thresholds doesn't guarantee a ranking.

18. E-E-A-T : Google quality framework

Short for experience, expertise, authoritativeness, and trust(worthiness), used in the quality rater guidelines. Trust sits at its center. E-E-A-T isn't a public score or a tag. Show who wrote the piece, how facts were established, why the source is competent, and how an error gets corrected.

19. YMYL : Google risk category

"Your Money or Your Life" describes topics that can strongly affect health, safety, financial stability, or society. The level of evidence and control should rise with the possible harm. It isn't a closed list or a schema. A sensitive page deserves primary sources, qualified review, a date, and explicit limitations.

20. Information gain : an evaluation concept, not a public SEO score

New, verifiable value given to the reader beyond what they could already find: data, a test, a calculation, a failure case, a contradictory synthesis, or a tool. The term is sometimes used without a method. State the original unit, its provenance, and its reproducibility; a longer rewording doesn't automatically create a gain.

21. Topic cluster : industry architecture model

A set of pages linked around a subject, often with a hub and specialized guides. Its value comes from covering distinct tasks and useful internal linking, not the page count. Every URL should have its own intent, evidence, and next action, or the cluster can become redundant.

22. Programmatic SEO : industry usage

Producing pages from data and templates. The method is legitimate if every generated state is a useful, accurate, controlled, and maintainable page. Swapping in a city or product name with no distinctive information creates weak pages. Define the URL inventory, indexing rules, QA, empty-data handling, rollback, and removal.

23. Large-scale content abuse : official Google policy

Creating many pages primarily to manipulate ranking rather than help people, whatever the mix of human, automation, or AI involved. Scale alone isn't the violation; intent and low value are central. Google's spam policies give examples and can evolve.

24. LLM : technical term

A large language model predicts and generates sequences from its training and the supplied context. It doesn't necessarily consult the web, doesn't have a guaranteed factual base, and can produce plausible errors. Distinguish model, application, search tools, memory, date, and parameters in any measurement.

25. Generative engine : industry usage

A system that composes an answer instead of only presenting a list of links. It can combine a model, search, document retrieval, tools, and citations. Two interfaces using a similar model family can produce different observations. Don't attribute all the behavior to the underlying LLM.

26. AI Overviews : Google product name

A generative summary that can appear in Google Search for some queries and users, accompanied by links. Its triggering and composition vary. The same SEO fundamentals still apply; Google states that no special markup or AI file is required to appear there (AI features and your website).

27. AI Mode : Google product name

Google's conversational search experience, allowing more complex questions and follow-ups, with links to the web. The interface, availability, and reporting can evolve by country and account. Document the tested surface rather than aggregating its answers with AI Overviews as an identical measurement.

28. RAG : technical architecture

Retrieval-augmented generation: an application retrieves documents or passages and then feeds them to the model to generate an answer. Quality depends on the query, the index, retrieval, context, and generation. RAG can reduce certain errors without guaranteeing accuracy, completeness, or faithful citation.

29. Grounding : platform term

The process of tying an answer to retrieved information, tools, or sources. Providers may measure "grounding sources" or supporting passages using different methods. A document that was used isn't necessarily shown as a link, and a displayed link doesn't prove every sentence came from it.

30. Query fan-out : generative search term

Breaking down or expanding a request into several related searches on sub-topics, sources, or data sets, then synthesizing the results. Google uses this term for its AI features. The exact sub-queries generally aren't all exposed, so a consultant's "fan-out queries" remain hypotheses to test, not official logs.

31. Prompt : AI interaction term

The input given to a generative system: an instruction, a question, context, and sometimes files or history. A prompt alone doesn't describe the experience; account, memory, location, model, tools, and date all influence the output. For a benchmark, version the exact text and repeat the runs.

32. Hallucination : common technical usage

A plausible but unsupported, incorrect, or fabricated output. The word covers different defects: an invented source, a wrong figure, a misattribution, or a conclusion that goes beyond the evidence. Measure them separately. An answer with citations can still hallucinate if the sources don't confirm the claims.

33. Brand mention : operational definition

The presence of the brand's name or a validated variant in an observed answer. It requires neither a link, nor a recommendation, nor positive sentiment. Define homonyms, misspellings, products, and parent companies before counting. The mention rate must state eligible prompts, platforms, runs, and period.

34. AI citation : ambiguous industry usage

In this glossary, a URL or domain explicitly shown as a source in an observed generative answer. Some tools use "citation" for an unseen grounding source or for a mention; check their definition. A citation proves neither reading, nor a click, nor causal influence.

35. AI share of voice : non-standard industry metric

The share of mention events attributed to a brand among a defined set of competitors, prompts, platforms, and runs. The result changes with the panel and the denominator. Publish the formula and keep platforms separate. A proprietary score with no raw events isn't comparable to another one.

36. AI visibility : industry measurement category

A generic term sometimes grouping mentions, citations, prompt frequency, page presence, and referral traffic. It has no universal unit. Replace "visibility +20%" with the exact observations: for example, 18 mentions out of 120 runs of a versioned panel.

37. GEO : non-standard industry usage

Generative Engine Optimization refers to improving the discoverability, understandability, and citability of a brand or its content in generative answers. The practice overlaps with SEO, PR, data, content, and measurement. There's no engine certification or recipe that guarantees a citation.

38. AEO : non-standard industry usage

Answer Engine Optimization refers to designing clear, retrievable answers for answer engines and their users. The term predates the rise of LLMs and can cover featured snippets, voice assistants, and AI answers. It doesn't replace accessibility, evidence, or site quality.

39. LLMO : non-standard industry usage

Large Language Model Optimization is used for efforts aimed at a brand's presence within LLM-based applications. Its scope varies: training, web retrieval, citations, or reputation. Prefer describing the intended surface and outcome rather than presenting LLMO as a standardized discipline.

40. Zero-click : analytics metric

A results session that doesn't lead to a click to the open web, according to the provider's own definition and data. The event can correspond to a satisfied answer, a rephrasing, an internal click, a close, or an incomplete measurement. Always state device, country, period, panel, and how clicks to the engine's own properties are treated.

41. AI crawler : ambiguous technical category

A robot associated with training, real-time search, an assistant, or a specific feature. A provider can operate several user agents with distinct purposes and controls. Check its official documentation and your logs; a name that resembles a known bot doesn't authenticate the request.

42. llms.txt : community proposal, not a ranking standard

A text file proposed to present models with a curated set of machine-readable resources. As of July 2026, Google states that no special AI file is needed for its search features. Treat llms.txt as a documented experiment, never as a replacement for robots.txt, a sitemap, accessible HTML, or structured data.

43. Impression : platform metric

An event counted when a result or link is considered displayed under the platform's rules. The definition varies by module and report. An impression is neither an attentive read, nor a brand mention, nor a citation. Keep the source, the surface, and the counting rule.

44. CTR : calculated metric

Click-through rate, generally clicks / impressions × 100. A change in CTR can come from position, the result module, the brand, intent, or the measurement itself. Don't compare CTRs whose impressions come from different surfaces, and don't infer causality from a simple variation.

45. Assisted conversion : attribution model

A conversion credited to an earlier touchpoint without it necessarily being the last click. The outcome depends on the window, identity, consent, and the attribution model. An AI exposure with no click is hard to tie to a conversion; use surveys, experiments, or aggregate analysis, and present the uncertainty.

Procedure: making a term measurable

  1. Identify the authority. Look first for product documentation, a standard, or an official policy. If none exists, label the term "industry usage."
  2. Define the observed object. An answer, a URL, a run, a user, an impression, or a conversion aren't interchangeable.
  3. Write the inclusion rule. Specify brand variants, failures, duplicates, redirects, multi-citations, and empty answers.
  4. Fix the denominator. Runs sent, eligible runs, unique prompts, impressions, or competing events.
  5. Bound the context. Platform, surface, account, language, country, device, date, and model where available.
  6. Attach the evidence. Keep the raw output, URL, export, screenshot, or log, and the definition's version.
  7. Document the limitation. Sampling, personalization, missing data, causality, or unknown coverage.
  8. Test reproducibility. Another person should be able to recompute the number from the observations.

Worked example: a chain with no causal shortcut

Someone asks: "Which tool should I use to track my brand's presence in AI answers?" The system can break this prompt down into sub-searches across the platforms covered, the methodology, the exports, and the prices: that's query fan-out. It retrieves several documents: that's the retrieval layer. A SEOryon page might be used to ground part of the answer: that's grounding. If its URL is shown, the protocol counts a citation; if the name appears, it counts a mention.

A platform might then report an impression, and if the person opens the link, a click. A demo request afterward could be an assisted conversion, depending on the attribution model. Each step has different evidence. The citation doesn't license a claim that it caused the sale; the click doesn't license a conclusion that the person read the whole page.

What the data proves and doesn't prove

Google announced in June 2026 a Search Console report devoted to generative performance, initially available to a subset of sites. The report shows, among other things, impressions, pages, countries, and devices; the data also stays aggregated in overall reports (Google's announcement and scope). That proves the report's stated definition and scope, not a sales attribution or an observation of every AI application.

Engine documentation describes their systems at the level needed for use, but doesn't publish every sub-query, weighting, or evaluated document. A consultant can formulate plausible fan-out queries; they must be presented as hypotheses. Likewise, an industry definition repeated by many publishers doesn't automatically become a standard.

Common mistakes and stopping conditions

  • Calling every direct or referrer-less session "AI traffic."
  • Comparing a ChatGPT mention, a Google impression, and a Bing citation within the same percentage.
  • Presenting GEO, AEO, or LLMO as an official certification.
  • Claiming a schema, llms.txt, or a content score guarantees a citation.
  • Forgetting failed answers in the denominator.
  • Using "AI ranking" without defining prompt, platform, repetition, and observed order.
  • Confusing a correlation between presence and traffic with a causal effect.

Stop an analysis if the provider doesn't give its metric's definition, if the raw observations don't allow recomputing the score, or if platforms and periods were aggregated with no stated rule.

Reusable asset: SEOryon's metrics dictionary

Create a table with the columns term, status, definition, object, numerator, denominator, platform, source, version, limitation, owner, and review_date. For any published statistic, add a row to assets/registre-preuves.csv.

A definition change should produce a new version, never silently rewrite the history. If "citation" moves from displayed domain to exact URL, recompute the series or place a visible break.

Where to go next

How SEOryon fits in

SEOryon uses these terms as a clarity contract, not a commercial promise. Every public page or feature should make the unit, scope, and limitations of its metrics accessible. When you evaluate SEOryon, ask for the same level of evidence you'd ask of any other provider: versioned definitions, exportable events, and a reproducible calculation.

Measurable exercise: fix ten ambiguous metrics

Take an existing SEO or GEO report. Pick ten terms from the 45 and fill in, for each, the object, numerator, denominator, platform, period, evidence, and limitation. Then have someone who didn't build the report recompute two of the percentages.

Deliverable: a versioned dictionary and the independent calculation. Success criterion: the second person finds exactly the same values, no industry term is presented as an official standard, and every important limitation appears next to its metric.

FAQ

Does GEO replace SEO?

No. Generative answers still rely on accessible, understandable, and trustworthy content. GEO adds surfaces and measurements but keeps the technical, editorial, and authority fundamentals.

What's the difference between AEO and GEO?

AEO emphasizes phrasing usable answers; GEO emphasizes presence in generated, sourced answers. In practice, their activities overlap. Define the outcome you want rather than debating a non-standardized boundary.

Not necessarily. It can be a visible link, an unfollowed reference, or a grounding source depending on the platform. Measure observed citation, click, link indexability, and business outcome separately.

Does structured data make a page appear in AI Overviews?

It can clarify entities and make a page eligible for certain features, but Google doesn't describe any special schema that guarantees a presence in AI Overviews or AI Mode.

How do you track a term whose meaning changes?

Version the definition, date the change, and keep the old series. If the units are no longer compatible, show a break rather than a misleading continuous trend.

References

  1. Google Search Central: How Google Search works2. Google Search Central: Canonicalization3. RFC 9309: Robots Exclusion Protocol4. Google Search Central: Spam policies5. Google Search Central: AI features and your website6. Google Search Central: AI features optimization guide7. Schema.org: Latest vocabulary release8. Schema.org: DefinedTermSet9. Google Search Central: Generative AI performance reporting

Method and update note

Glossary reviewed 22 July 2026, translated and edited from the French original (reviewed 16 July 2026). The "official," "standard," and "industry usage" statuses were assigned from the listed sources and how the term is actually used; they don't claim to fix the language permanently. Product names, interfaces, and reports evolve. Check the Google, Schema.org, and RFC documentation each quarter, then version any change that alters a metric or a history.