French-Language AEO / SEO / GEO Glossary

The French-language reference for AEO, SEO, and GEO terminology — structured definitions (claim, evidence, warrant, and limitations) at three levels of depth, with transparent terminology status for every term: recognized standard, emerging convention, or marketing positioning.




—

AEO Emerging convention
Beginner
Claim
AEO means optimizing content so that an answer engine such as Google SGE or Alexa uses it directly as an answer, a practice that predates the rise of mainstream generative AI.
Example
A page that answers “What is the capital of Portugal?” in one clear sentence and appears in Google's answer box is an example of AEO.
Intermediate
Claim
AEO is the historical precursor to GEO, focused on featured snippets, voice assistants, and direct answers before conversational LLMs became widespread.
Evidence
AEO techniques — an answer at the top of the page, a question-and-answer structure, and concise wording of 40–60 words — remain technical prerequisites for GEO, which extends them to more complex generative engines.
Warrant
An answer engine such as SGE or Alexa and a generative engine such as ChatGPT share the same need: extracting a short, unambiguous answer — hence the methodological continuity between AEO and GEO.
Expert
Claim
AEO is the discipline of optimizing content for direct-answer systems — featured snippets, voice assistants, and SGE — based on extracting a short, self-contained passage rather than ranking an entire page.
Evidence
Historical answer engines use passage-extraction heuristics similar to extractive QA rather than free-form generation — performance depends on a direct match between the question and the answer provided on the page.
Warrant
Question-and-answer content with a 40–60-word answer immediately after the heading maximizes the probability that these systems will extract it correctly.
Backing
This practice stems from SEO experience with Google featured snippets from 2015 to 2023, documented by the industry but never formalized by a standards body.
Scope
AEO remains relevant to traditional passage-extraction systems, but it is an insufficient subset for generative engines that reformulate and synthesize multiple sources instead of extracting a single passage.
Rebuttal
Relying solely on AEO techniques (short answers and Q&A structure) does not guarantee citation by a generative engine — but not for the reason one might assume: Google explicitly states that no special markup or AI-specific file is required to appear in its AI search features. AEO's real limitation lies elsewhere: it targets the extraction of a single passage, whereas a generative engine may synthesize and cite multiple sources simultaneously.

Voir aussi : GEO (Generative Engine Optimization) · RAG (Retrieval-Augmented Generation)

Précurseur du GEO, centré moteurs de réponse (Google SGE, Alexa).

aeo

Transversal Marketing positioning
Beginner
Claim
“AI-READY” and “AI-First” mean “ready for AI” — useful for explaining a vision to executives, but they are not official technical terms.
Example
Telling a client “your site must be AI-READY” is clear and compelling, but a developer will prefer concrete requirements such as “it needs JSON-LD and SSR rendering.”
Intermediate
Claim
AI-READY and AI-First are strategic and marketing positioning terms, useful for communicating a vision to non-technical executives, but they do not correspond to any standard recognized by the W3C, the IETF, or the AI ecosystem.
Evidence
Equivalent technical vocabulary already exists and is documented: GEO/Machine-First Design for “AI-READY,” and Agentic SEO/WebMCP-UCP for “AI-First.”
Warrant
Using terminology suited to the audience — marketing language with a decision-maker, technical language with a developer — avoids confusion and makes the message more credible.
Expert
Claim
“AI-READY” and “AI-First” are strategic positioning labels — explicitly used here, notably by Jean-Jacques Valognes — respectively describing the goal of visibility in generative engines, encompassing GEO/Machine-First Design, and an architecture that prioritizes agentic interaction, encompassing Agentic SEO/WebMCP/UCP; neither label is standardized.
Evidence
No organization — W3C, IETF, or an AI-specific consortium — defines or certifies “AI-READY” or “AI-First,” unlike JSON-LD, a W3C Recommendation, or even llms.txt, a convention documented by an identifiable author.
Warrant
The value of these terms is communicative — selling a vision — rather than technical — specifying an implementation; both uses are legitimate as long as they are not confused.
Backing
This finding aligns with the terminology clarification undertaken in this project — the marketing/technical mapping table and Vocabulaire.txt: explicitly identifying the term's marketing nature distinguishes this glossary from uncritical use of AI vocabulary.
Scope
Not being a standard does not make a term illegitimate: transparent, explicitly stated strategic positioning is sound practice, provided it is not falsely presented as a certification or technical standard.
Rebuttal
The term's structural weakness lies not only in its lack of recognition by a standards body, but also in the absence of measurable, auditable, and interoperable criteria attached to it: unlike JSON-LD (whose syntax can be validated) or llms.txt (whose presence can be verified), nothing makes it possible to determine objectively whether a website is 'AI-READY' — making the term structurally unsuitable for any form of certification, even informal.

Voir aussi : Machine-First Design · GEO (Generative Engine Optimization)

Explicit strategic positioning (Jean-Jacques Valognes), not a recognized W3C/IETF technical standard. A textbook case of the glossary's terminological transparency — no external source (proprietary term).

ai-ready

Transversal Emerging convention
Beginner
Claim
The Discovery Layer is everything that helps an AI agent quickly find and understand what a website is about without having to read it all.
Example
The llms.txt file at a website's root, which summarizes what the site does in a few lines, is part of the Discovery Layer.
Intermediate
Claim
The Discovery Layer groups the files and markup that let an agent find and understand a website's context unambiguously: llms.txt, robots.txt, sitemap.xml, and JSON-LD.
Evidence
An agent reads these elements before making a more costly decision — a full crawl or an action call — making them the entry point for every agent–website interaction.
Warrant
An agent that cannot quickly find this discovery layer incurs higher processing costs and may abandon the website for a source that is faster to interpret.
Expert
Claim
The Discovery Layer is the first of three agentic architecture layers — Discovery, Action, and Trust — and comprises the standardized entry points that expose a website's context, scope, and usage rules to an agent.
Evidence
In practice: llms.txt for an editorial summary, robots.txt for crawl permissions by user agent, sitemap.xml for the URL inventory, and JSON-LD for page-level semantics.
Warrant
Without this layer, an agent must infer context heuristically by parsing raw HTML — slower, more fragile, and more token-intensive for an LLM.
Backing
This three-layer taxonomy synthesizes 2026 industry practice without a single formal standard, yet it structures how recent protocols such as llms.txt, WebMCP, and UCP divide their functions.
Scope
The exact composition of the Discovery Layer evolves quickly, with new files regularly proposed, such as agents.md and ai-plugin.json; the 2026 list is not fixed.
Rebuttal
The three-layer taxonomy (Discovery/Action/Trust) is itself debatable: it is not a nomenclature validated by a standards body, its boundary with the Action Layer is porous (some discovery files are evolving toward actionable capabilities), and its composition changes too quickly (agents.md, ai-plugin.json, etc.) to claim exhaustiveness — it is a teaching tool used by this glossary, not a stabilized classification.

Voir aussi : llms.txt · JSON-LD & Schema.org

Files and markup enabling an agent to discover and understand a website unambiguously (llms.txt, robots.txt, sitemap.xml). Synthesized taxonomy; no single external source is expected.

discovery-layer

GEO Emerging convention
Beginner
Claim
GEO is the art of making a website clear and well structured enough for an AI such as ChatGPT or Perplexity to understand it and cite it in its answers.
Example
If you search for “best bicycle repair shop Paris” on Perplexity and the answer cites a specific website with its name and address, that site has implemented GEO — not just traditional SEO.
Intermediate
Claim
GEO is the discipline of optimizing a website so that generative engines — ChatGPT, Perplexity, Gemini, and Claude — cite and recommend it, complementing traditional SEO focused on ranking in a list of links.
Evidence
Unlike SEO, where users click one of ten links, in GEO the AI synthesizes a single answer and may cite only a handful of sources — visibility depends on how citable the content is, not only on its ranking.
Warrant
Structured content — JSON-LD, semantic HTML, and direct answers at the top of the page — is mechanically easier for a retrieval-augmented generation system (RAG) to extract and is therefore more likely to be cited.
Expert
Claim
GEO encompasses the technical and editorial practices intended to increase the probability that content will be selected and cited by a generative engine answering a user's query — a visibility gain measured empirically on certain systems, not a causal 'retrieval → citation' chain demonstrated universally across all engines.
Evidence
Generative engines combine a retrieval index with an LLM-based synthesis mechanism that selects, summarizes, and cites a subset of sources — the citation window is narrow, often three to eight sources per answer.
Warrant
Citation probability increases with structured data (Schema.org/JSON-LD) that disambiguates the described entity, semantic HTML that facilitates chunking and extraction, and a direct, self-contained answer at the beginning of a section.
Backing
This link between structured content and increased citation is consistent with the RAG literature (Lewis et al., 2020): a retrieval system performs better on self-contained, unambiguous, properly segmented passages — yet the study demonstrates only a measured visibility gain, not a mechanical guarantee of citation.
Scope
GEO is not a registered technical standard — no W3C or IETF body defines it — but an industry-consensus term in 2026 that may evolve along with the retrieval and citation mechanisms of generative engines.
Rebuttal
Perfectly structured content may still never be cited if an engine favors other sources according to criteria specific to each engine that are unpublished and therefore unverifiable from the outside — GEO improves a probability, guarantees no citation, and the true magnitude of this effect remains sparsely documented in public.

Authoritative source: https://arxiv.org/abs/2311.09735

Voir aussi : AEO (Answer Engine Optimization) · LLMO (Large Language Model Optimization) · Machine-First Design

Terme consensuel industrie 2026, pas d'organisme normatif dédié. Remplace AI-READY en langage stratégique.

geo

Transversal Recognized standard
Beginner
Claim
JSON-LD is invisible code added to a page to tell a machine clearly “this is a recipe,” “this is an event,” and so on, without changing what a human visitor sees.
Example
On an event page, JSON-LD can specify the date, location, and price in a format that Google or an AI understands instantly, even when the same information also appears as normal page text.
Intermediate
Claim
JSON-LD is the recommended serialization format for embedding Schema.org structured data in a web page, enabling a machine to identify unambiguously the entity type described — product, article, person, or organization.
Evidence
JSON-LD is a formal W3C Recommendation; Schema.org, the vocabulary most commonly used with JSON-LD, is a community project supported by Google, Microsoft, Yahoo, and Yandex — not strictly a W3C standard.
Warrant
Separating structured data — in the head or at the end of the page — from visible content allows markup without risking layout breakage or relying on the machine's visual interpretation.
Expert
Claim
JSON-LD is a Linked Data serialization format and W3C Recommendation used to embed a structured-data graph — most often using Schema.org vocabulary — in an HTML page through a <script type="application/ld+json"> element, independently of visual rendering.
Evidence
A JSON-LD document can describe one entity or an @graph of multiple entities linked by @id — for example, a page linked to its author, who is linked to an organization — a pattern used notably on alchimiste-ia-home.
Warrant
W3C standardization gives JSON-LD stable syntax and durable tool support; although Schema.org is not a W3C standard, it has de facto support from the major engines that define its practical interpretation.
Backing
JSON-LD 1.1 is a W3C Recommendation published in 2020; Schema.org is publicly documented and versioned by its dedicated W3C Community Group, with several thousand specified types.
Scope
Syntactically valid JSON-LD guarantees neither use by a particular engine nor a rich result; each platform chooses which types and fields to process and how.
Rebuttal
JSON-LD that conflicts with visible page content — excessive markup or invented data — is a deceptive practice that search engines may penalize; structured data must remain faithful to the actual content.

Authoritative source: https://www.w3.org/TR/json-ld11/

Voir aussi : Discovery Layer · Trust Layer

JSON-LD is a W3C Recommendation (see the url_ref_externe link). Schema.org is a community vocabulary (Google/Microsoft/Yahoo/Yandex), not formally ratified by the W3C — a nuance to retain at the expert level. See also https://schema.org/.

json-ld

GEO Emerging convention
Beginner
Claim
llms.txt is a small text file at a website's root that summarizes what an AI should know about the site — similar to robots.txt, but providing context rather than restrictions.
Example
A website with a 20-line /llms.txt file listing its main pages and area of expertise greatly helps an agent understand it quickly.
Intermediate
Claim
llms.txt is a convention for a Markdown file at the domain root that provides a summary and editorial context directly readable by an LLM without crawling the entire website.
Evidence
Initiated by Jeremy Howard of Answer.AI, the format is not adopted by any standards body, but major technical publishers including Vercel, Stripe, and Anthropic implemented it in 2025–2026.
Warrant
Simple Markdown with no rendering dependency can be consumed directly by an LLM without HTML parsing, explaining its rapid adoption despite the lack of a formal standard.
Expert
Claim
llms.txt is a community convention — neither W3C nor IETF — for a /llms.txt file at a domain root that uses Markdown to structure an editorial summary and priority resource list for LLMs, complementing robots.txt and sitemap.xml.
Evidence
The typical format includes a website title and summary, links to primary resources, areas of expertise, the responsible entity, content policy, and last-updated date.
Warrant
The absence of HTML and CSS markup in llms.txt removes all layout noise; an LLM can ingest it directly as system context at minimal processing cost.
Backing
It is documented by Howard's initial 2024 proposal and growing adoption among leading technical publishers, although no binding formal specification exists to date.
Scope
llms.txt has no binding force: nothing requires an LLM or crawler to read it or follow its content — it is a goodwill convention, not a technical enforcement mechanism.
Rebuttal
A poorly maintained llms.txt file (outdated or inconsistent with the website's actual content) can mislead an agent rather than help it, and nothing in the format enables an agent to detect that drift automatically. An important nuance: syntactically valid JSON-LD does not guarantee that the data it contains is truthful either — both formats share this limitation, as a valid structure never certifies accurate content.

Authoritative source: https://llmstxt.org/

Voir aussi : Discovery Layer

A community convention (initiated by Jeremy Howard / Answer.AI), not a W3C/IETF standard, with growing real-world adoption (Vercel, Stripe, Anthropic).

llms-txt

SEO Recognized standard
Beginner
Claim
A SERP is the results page a search engine displays after a query — the ranked list of links that traditional SEO seeks to optimize, as opposed to the single answer targeted by GEO.
Example
Typing “best Italian restaurant Paris” into Google and seeing a ranked list of ten websites is a SERP — unlike an AI that would answer directly with a single restaurant name.

Voir aussi : AEO (Answer Engine Optimization) · GEO (Generative Engine Optimization)

Terme fondateur du SEO classique, permet le contraste direct avec le GEO (réponse synthétisée unique vs liste de résultats classés).

serp

Transversal Emerging convention
Beginner
Claim
The Action Layer consists of the doors a website opens so an AI agent can act on it — not merely read, but book, buy, or complete a form.
Example
An agent that adds an item to a cart and places an order through a dedicated interface, rather than clicking like a human, uses the website's Action Layer.
Intermediate
Claim
The Action Layer groups the programmatic interfaces — APIs and endpoints — that let an agent interact with a website rather than merely read it.
Evidence
Unlike the Discovery Layer, which is read-only, the Action Layer triggers a transaction or changes website state, such as a cart, booking, or contact request.
Warrant
Reliable action by an agent requires an explicit interface contract with a defined input/output schema; an agent cannot infer a web form as reliably as it can call a typed function.
Expert
Claim
The Action Layer is the agentic architecture layer that exposes structured programmatic interfaces — REST APIs, tool/function calling, and protocols such as UCP or WebMCP — through which an agent can perform a transactional action on a website.
Evidence
The key 2026 protocols are UCP — Universal Commerce Protocol for e-commerce — and WebMCP — a W3C Community Group draft led by Google and Microsoft for general agent/tool/data connectivity — both built on tool/function calling.
Warrant
A structured interface reduces an agent's error rate during an action compared with automation that must interpret an HTML DOM designed for humans.
Backing
Tool calling has been documented since the first LLM agent architectures, including OpenAI function calling in 2023; WebMCP and UCP are web-native and sector-specific extensions of it.
Scope
In 2026, the Action Layer remains uncommon outside e-commerce, where UCP is progressing; most websites expose no interface on which an agent can act.
Rebuttal
Exposing an Action Layer without a robust Trust Layer is risky: an agent or its operator will not authorize an automated transaction on a website whose reliability cannot be verified.

Voir aussi : WebMCP · UCP (Universal Commerce Protocol)

Programmatic interfaces enabling an agent to act on the website (book, purchase, or make contact). Synthesized taxonomy; no single external source is expected.

action-layer

GEO Emerging convention
Beginner
Claim
Agentic SEO means preparing a website so that an autonomous AI agent — not just a human — can navigate it and take action: search for a product, book, or pay.
Example
An AI agent asked to “book me a table for two tonight” that directly completes a restaurant's booking form is an example of Agentic SEO on the visited website.
Intermediate
Claim
Agentic SEO is a term whose usage remains unstable: depending on the author, it denotes either the adaptive automation of SEO workflows themselves by AI agents (the more prevalent meaning in current industry usage) or — by extension — a website's ability to be operated by an external AI agent (search, booking, or purchase), in which case it largely overlaps with the Action Layer.
Evidence
Unlike GEO, which targets citation in an answer, the second meaning targets the agent's effective action on the website (filling a cart or initiating a payment); the first targets iterative optimization of the SEO work itself (audits and automated content adjustments).
Warrant
No terminological consensus has yet been established in 2026 — both usages coexist in industry literature, with no organization or prevailing convention deciding between them.
Expert
Claim
Agentic SEO is a term whose usage remains unstable, with two distinct meanings: (1) the adaptive automation of SEO workflows themselves by AI agents (audits and iterative content optimization) — the dominant meaning in current industry usage; (2) by extension among some authors, a website's ability to be operated by an external AI agent — a usage that then largely overlaps with the Action Layer. In the absence of an established consensus, this glossary does not choose between them.
Evidence
Emerging 2026 protocols (UCP for e-commerce and WebMCP for tool/data connectivity) define structured interaction contracts between agents and websites — relevant to meaning (2), with no direct relationship to meaning (1).
Warrant
A website exposing these interfaces (meaning 2) drastically reduces an agent's failure and hallucination rates during a transactional task compared with an agent forced to infer the action through the DOM.
Backing
Meaning (2) extends tool/function calling (an LLM's ability to invoke an external function according to a defined schema), the basic technical building block underlying WebMCP/UCP; meaning (1) instead belongs to industry SEO literature on agent-driven automation (e.g. Ahrefs, 2026).
Scope
An emerging discipline in 2026 in both senses: few websites implement actionable interfaces (meaning 2), while agentic SEO workflows (meaning 1) remain largely experimental — not to be confused with the Action Layer, which is the name of the technical layer, not the discipline.
Rebuttal
Presenting Agentic SEO (meaning 2) as a discipline autonomous and distinct from the Action Layer is debatable: without a strong Trust Layer, an agent may refuse to act even when the technical interface is present, and the definition itself largely overlaps with that of the Action Layer — the distinction remains more editorial than technical.

Voir aussi : Action Layer · UCP (Universal Commerce Protocol) · WebMCP

Terme d'usage instable : recouvre à la fois l'automatisation des workflows SEO par des agents IA (sens dominant en usage sectoriel) et, par extension, la capacité d'un site à être actionné par un agent externe (proche de l'Action Layer).

agentic-seo

Transversal Emerging convention
Beginner
Claim
Machine-First Design means building a website by first considering what a machine must understand, not only what looks attractive to a human.
Example
Adding invisible code (JSON-LD) that explicitly states “this is a recipe with these ingredients,” alongside the visible text, is Machine-First Design.
Intermediate
Claim
Machine-First Design is a web architecture approach in which machine readability — JSON-LD and a semantic DOM — takes precedence over or equals human readability, replacing the marketing positioning “AI-First.”
Evidence
A Machine-First website systematically exposes an explicit semantic layer such as Schema.org alongside its visual rendering instead of relying on AI to infer meaning from visual HTML and CSS.
Warrant
A machine perceives neither layout nor visual intent; only explicit structure — markup and semantic hierarchy — removes ambiguity that layout alone cannot resolve for a parser.
Expert
Claim
Machine-First Design is a software architecture principle under which every meaningful piece of page information must be accessible and unambiguous to an automated parser — crawler, agent, or LLM — independently of its visual rendering.
Evidence
The principle translates into SSR or static content with no JavaScript dependency for primary content, systematic JSON-LD, semantic HTML, and a consistent heading hierarchy.
Warrant
Client-side-only JavaScript rendering is invisible to a significant share of AI crawlers that do not execute JavaScript; Machine-First Design eliminates this failure point by construction.
Backing
An empirical finding documented in this project illustrates it: a dashboard rendered only with JavaScript before its SSR refactor delivered an empty DOM on initial load, invisible to every non-JavaScript crawler.
Scope
Machine-First Design does not conflict with a good human experience — it adds a requirement rather than replacing it.
Rebuttal
Taking Machine-First Design too far — excessive markup or JSON-LD that poorly duplicates visible content — can create inconsistencies between what a machine reads and what a human sees, risking a penalty for deceptive content.

Voir aussi : JSON-LD & Schema.org

An architectural approach in which machine readability takes precedence over, or is placed on an equal footing with, human readability. Replaces AI-First in technical language. A concept synthesized by JJV; no single external source is expected.

machine-first-design

Transversal Recognized standard
Beginner
Claim
RAG is when an AI looks up fresh information on the internet or in a database before answering instead of relying only on what it learned during training.
Example
When ChatGPT searches the web to answer a question about recent news and cites its sources, it is using RAG.
Intermediate
Claim
RAG — Retrieval-Augmented Generation — combines an information-retrieval system with a generative model (LLM), grounding an answer in external sources rather than solely in the model's parametric memory.
Evidence
RAG reduces hallucination risk by giving the LLM factual context retrieved at query time that the model may not have memorized during training.
Warrant
Properly structured, indexable content is more likely to be selected during retrieval and therefore to influence the generated answer — the direct link between RAG and GEO.
Expert
Claim
RAG is a two-stage architecture: retrieval of relevant passages from an external corpus, often through vector search or dense retrieval, followed by answer generation by an LLM conditioned on those passages; it was notably formalized by Lewis et al. in 2020.
Evidence
Retrieval depends on indexing quality — embeddings and chunking — and content availability — crawlable, unblocked, and not JavaScript-only; generation depends on the LLM's ability to synthesize the supplied passages faithfully.
Warrant
Well-segmented, structured, accessible content — see LLMO — directly improves retrieval; GEO specifically seeks to optimize the probability of being retrieved and then cited in generated output.
Backing
The foundational academic reference is Lewis, P. et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” NeurIPS 2020; the technique has since been widely industrialized.
Scope
Not all generative engines use traditional RAG; some combine real-time web search, caches, and internal RAG — the term denotes a family of techniques, not one implementation.
Rebuttal
RAG reduces but does not eliminate hallucinations: an LLM may misinterpret or misattribute a correctly retrieved passage — retrieval quality is necessary but insufficient for a reliable answer.

Authoritative source: https://arxiv.org/abs/2005.11401

Voir aussi : GEO (Generative Engine Optimization) · AEO (Answer Engine Optimization)

An established AI research technique (Lewis et al., 2020, NeurIPS — foundational paper linked in url_ref_externe). GEO aims to become the grounding source for RAG.

rag

GEO Emerging convention
Beginner
Claim
LLMO refers to techniques that make content easier for an LLM to ingest and reuse — a more technical term, used less often in strategy than GEO.
Example
Splitting an article into short, self-contained paragraphs instead of one 2,000-word block is an LLMO practice.
Intermediate
Claim
LLMO is the technical, developer-oriented counterpart to GEO: it optimizes how an LLM ingests content — segmentation, formatting, and removal of visual noise — rather than overall visibility strategy.
Evidence
Developers use LLMO for specific implementation choices such as chunk size and clean Markdown or HTML structure, whereas GEO more often describes the business goal of being cited.
Warrant
Poorly segmented or JavaScript-heavy content forces an LLM, or its RAG pipeline, into degraded processing — LLMO directly targets this mechanical stage.
Expert
Claim
LLMO encompasses content-engineering practices that optimize the segmentation, cleaning, and formatting of a page for the chunking and embedding stages of a RAG pipeline.
Evidence
A RAG pipeline splits text into chunks, often 200–800 tokens, before vectorizing them; noisy HTML — menus, ads, or unrendered JavaScript — contaminates that segmentation.
Warrant
Reducing structural noise — clean SSR, semantic HTML, and self-contained text — mechanically improves the quality of generated chunks and therefore retrieval relevance.
Backing
This follows documented corpus-preparation practices for vectorization in RAG, common in ML engineering, without constituting a standard formally defined under the name LLMO.
Scope
LLMO is a usage-based term with narrower adoption than GEO or AEO; its exact scope varies depending on whether a developer or a marketer uses it.
Rebuttal
Content that is technically perfect for LLMO may still go uncited if it lacks authority from the Trust Layer — LLMO optimizes ingestion, not the trust granted to the source.

Voir aussi : GEO (Generative Engine Optimization) · RAG (Retrieval-Augmented Generation)

Terme technique développeur, moins répandu que GEO côté stratégie.

llmo

Transversal Emerging convention
Beginner
Claim
The Trust Layer consists of the evidence a website shows an AI to say “you can trust me” — who I am, who wrote this, and whether the information can be verified.
Example
Clearly naming an article's author and linking to their expertise profile, rather than publishing anonymously, strengthens the page's Trust Layer.
Intermediate
Claim
The Trust Layer groups the elements — Entity Schema, author markup, and cited primary sources — that let an agent validate information reliability before citing it or acting on it.
Evidence
An AI agent generally cross-checks several signals — the website, third-party profiles, and reviews — rather than trusting one isolated page.
Warrant
Information that cannot be attributed to a verifiable entity — an author or organization — is statistically less likely to be selected as a source by a system seeking to limit hallucinations.
Expert
Claim
The Trust Layer is the agentic architecture layer that provides verifiability and authority signals (Entity Schema, authorship, primary sources, and cross-platform consistency), enabling an agent to assess the editorial reliability of content before citation or action. DNSSEC belongs at a different level: it authenticates the origin and integrity of DNS data, not the reliability of the content published there.
Evidence
In practice: Person/Organization Schema.org markup, consistent NAP data (name, address, and phone number) across the website and third-party directories, aggregated customer reviews, content age and freshness, transport security (HTTPS), and domain-resolution security (DNSSEC, with no bearing on editorial reliability).
Warrant
These signals reduce the risk perceived by the answer-generating system: citing a verifiable source limits its own exposure to producing a falsely attributed hallucination.
Backing
This principle extends Google's E-E-A-T criteria — Expertise, Experience, Authoritativeness, and Trustworthiness — already documented in traditional SEO to the agentic context.
Scope
The Trust Layer has no single objective measure, unlike JSON-LD, which can be validated syntactically; assessment remains largely heuristic and specific to each engine or agent.
Rebuttal
A strong Trust Layer cannot compensate for structurally invisible content — no SSR or Discovery Layer: trust matters only for content the agent has already reached and understood.

Voir aussi : JSON-LD & Schema.org

Elements (Entity Schema, authorship, and primary sources) enabling an agent to validate the reliability of information. Synthesized taxonomy; no single external source is expected.

trust-layer

GEO Recognized standard
Beginner
Claim
UCP is a common language that lets AI shop in any online store in the same way, without every website developer having to reinvent the process.
Example
With UCP, an AI agent could find a product, add it to the cart, and pay in any compatible store using the same steps each time.
Intermediate
Claim
UCP — Universal Commerce Protocol — is an open standard co-developed by Shopify and Google in 2026 that lets AI agents interact uniformly with e-commerce stores for catalog search, cart management, and checkout.
Evidence
Before UCP, each e-commerce website exposed — or did not expose — its own agent-integration logic; UCP aims to provide one contract regardless of the website.
Warrant
A unified transactional interface lowers integration costs for agent developers — one protocol to support — and increases the coverage of e-commerce that agents can actually operate.
Expert
Claim
UCP is an open, standardized protocol co-developed by Shopify and Google in 2026 that defines e-commerce operations actionable by an AI agent — catalog search, cart management, and checkout — independently of the underlying commerce platform.
Evidence
It is the leading Action Layer building block for e-commerce in 2026, positioned as more mature and widely adopted than generic alternatives still in draft form, such as WebMCP.
Warrant
The weight of its co-developers — Shopify as a leading commerce infrastructure provider and Google as a leading discovery engine — creates network effects favorable to UCP becoming a de facto standard.
Backing
UCP belongs to the same family as ACP — Agentic Commerce Protocol — a broader term covering UCP and private agent-to-website transaction APIs.
Scope
It is a sector-specific e-commerce standard, not a general-purpose one: it does not cover non-commercial use cases such as service bookings or contact forms.
Rebuttal
UCP adoption depends on effective integration by third-party e-commerce platforms beyond Shopify; a website on an incompatible platform cannot benefit despite UCP's open-standard status.

Authoritative source: https://ucp.dev/

Voir aussi : WebMCP · Action Layer

An open standard co-developed by Shopify and Google (2026) for agent/e-commerce interactions (catalog, cart, and checkout).

ucp

GEO Emerging convention
Beginner
Claim
WebMCP is a project under discussion intended to let AI connect easily to a website's tools and data, rather like a universal power socket between AI agents and websites — it is not yet an official standard.
Example
With WebMCP, an AI agent could query an online store's inventory directly instead of inferring the information by reading a product page.
Intermediate
Claim
WebMCP is a draft (Community Group Draft, explicitly outside the Standards Track) discussed in a W3C Community Group (with participation from Google and Microsoft, initiated in early 2026), intended to define a JavaScript API that standardizes connections between LLMs/agents and external tools or data — the web-native evolution of Anthropic's MCP protocol (late 2024).
Evidence
MCP — Model Context Protocol — was initially designed to connect an LLM to local tools and servers; WebMCP transposes that principle to the web context, from website to website and agent to website.
Warrant
Standardizing this connection prevents every website from defining its own ad hoc protocol and every agent from developing a website-specific integration.
Expert
Claim
WebMCP is a Community Group Draft (explicitly outside the W3C Standards Track and therefore not a standards proposal in the strict sense) discussed within a W3C Community Group (with participation from Google and Microsoft, initiated in early 2026). It defines a JavaScript API — not a generic network protocol — that lets an agent or LLM discover and invoke tools exposed by a web page, directly following in the lineage of Anthropic's Model Context Protocol (2024).
Evidence
The protocol aims to define discovery of available website tools, their invocation schemas with typed inputs and outputs, and the call mechanism — at the same abstraction level as tool/function calling, but standardized across websites.
Warrant
A protocol developed in the W3C ecosystem can gain network effects and longevity that proprietary website-by-website integrations cannot offer.
Backing
Its lineage from MCP — a specification published by Anthropic in late 2024 and widely adopted for AI development tools — gives this web extension a proven technical foundation.
Scope
It remains a draft in 2026: the specification is not final and real implementation by browsers and websites is marginal — it should be treated as a direction, not an operational standard.
Rebuttal
A draft may be abandoned or substantially changed before standardization; using WebMCP today as a production foundation risks rapid obsolescence, unlike JSON-LD, a stable W3C Recommendation.

Authoritative source: https://webmachinelearning.github.io/webmcp/

Voir aussi : Action Layer · UCP (Universal Commerce Protocol)

A Community Group Draft (explicitly outside the W3C Standards Track), led by Google and Microsoft and initiated in early 2026, for connecting LLMs/agents to external tools and data through a JavaScript API.

webmcp