# GEO Toolbox — complete glossary index

> Every AI-search / GEO term defined on https://geotoolbox.ai, in every locale. The navigation index is at https://geotoolbox.ai/llms.txt.

Any glossary URL also serves plain markdown: append `.md` to it, or send `Accept: text/markdown`.

## English (57 terms)

- [Glossary index](https://geotoolbox.ai/glossary)
- [AI Agent](https://geotoolbox.ai/glossary/ai-agent): An AI agent is an AI system that takes actions to reach a goal, not just answers a question. Where a chatbot replies, an agent plans steps, uses tools (search, code, APIs, a browser), and works through a multi-step task on your behalf. Coding assistants, deep-research modes, and computer-use tools are common examples.
- [AI Citation](https://geotoolbox.ai/glossary/ai-citation): An AI citation, in the AI-search sense, is when an AI search engine references and links your page as a source in its answer. Citations are the core unit of AI visibility: in a zero-click world, being one of the cited sources, not ranking a blue link, is what gets your brand seen.
- [AI Crawler](https://geotoolbox.ai/glossary/ai-crawler): An AI crawler is an automated bot that fetches web pages for an AI company, either to train models or to retrieve live content for answers. You control most of them through robots.txt, but reaching AI answers depends on the right crawlers being allowed.
- [AI Hallucination](https://geotoolbox.ai/glossary/ai-hallucination): An AI hallucination is when a model generates false or fabricated information and presents it as fact. In AI search it shows up as wrong claims, invented sources, or incorrect brand details. Grounding answers in retrieved, citable sources is the main defense, which is why clear, sourced content matters.
- [AI Inference](https://geotoolbox.ai/glossary/ai-inference): Inference is the act of running a trained AI model to produce an output, as opposed to training, which is building the model in the first place. Every time you send a prompt and get an answer, that is one inference. It is where the ongoing cost, speed, and energy use of AI mostly live.
- [AI Overviews](https://geotoolbox.ai/glossary/ai-overviews): AI Overviews are Google's AI-generated summaries that appear at the top of some search results. Powered by Gemini (Gemini 3 became the default model in January 2026), they synthesize an answer from multiple web sources and link to a few of them, so users often get the answer without clicking. Being one of the cited sources is the goal of optimizing for AI Overviews.
- [AI Search Engine Optimization](https://geotoolbox.ai/glossary/ai-search-engine-optimization): AI search engine optimization is the umbrella term for improving how your brand appears in AI-powered search, across ChatGPT, Perplexity, Gemini, and Google AI Overviews. It is the plain-English name for the same discipline also called generative engine optimization (GEO) and answer engine optimization (AEO).
- [AI Visibility](https://geotoolbox.ai/glossary/ai-visibility): AI visibility is how often, and how favorably, AI engines like ChatGPT, Google AI Overviews, Gemini, and Perplexity surface your brand in their answers. It is the AI-era equivalent of search rankings: instead of where you sit in a list of links, it measures whether the AI mentions, cites, or recommends you at all.
- [Amazonbot](https://geotoolbox.ai/glossary/amazonbot): Amazonbot is Amazon's web crawler, used to improve Amazon services such as letting Alexa answer more questions, with crawled data that may also help train Amazon's AI models. It identifies with the user agent Amazonbot and respects robots.txt.
- [Answer Engine](https://geotoolbox.ai/glossary/answer-engine): An answer engine is a search tool that returns a single synthesized answer to a question rather than a list of links to evaluate. Perplexity, Google's AI Overviews, and ChatGPT search are answer engines. Optimizing to be cited in them is the focus of answer engine optimization (AEO).
- [Answer Engine Optimization (AEO)](https://geotoolbox.ai/glossary/answer-engine-optimization): Answer engine optimization (AEO) is the practice of structuring content to win the direct answer in answer engines, including featured snippets, voice results, and AI answers. It overlaps almost entirely with generative engine optimization (GEO); the two are the same job described from different angles.
- [Applebot](https://geotoolbox.ai/glossary/applebot): Applebot is Apple's web crawler, powering Siri, Spotlight, and Safari search suggestions. A separate robots.txt token, Applebot-Extended, controls whether your content may be used to train Apple's foundation models, without affecting Applebot's search crawling.
- [Bingbot](https://geotoolbox.ai/glossary/bingbot): Bingbot is Microsoft's web crawler that builds the Bing search index. It matters for AI visibility because Bing's index also backs Microsoft Copilot and ChatGPT search, so a page Bingbot cannot reach can be missing from those AI answers too.
- [Brand Mention](https://geotoolbox.ai/glossary/brand-mention): A brand mention is any reference to your brand across the web, linked or unlinked. For AI search, mentions act as corroboration: the more consistently your brand and its facts appear on trusted sources, the more confidently engines recognize and cite you.
- [Bytespider](https://geotoolbox.ai/glossary/bytespider): Bytespider is ByteDance's web crawler, used to collect data to train its large language models. It is known for crawling aggressively and reportedly does not always respect robots.txt, so blocking it often takes a server or WAF rule rather than robots.txt alone.
- [CCBot](https://geotoolbox.ai/glossary/ccbot): CCBot is the crawler operated by Common Crawl, a nonprofit that publishes a free, open dataset of web pages. Because that dataset is widely used to train large language models, CCBot is one of the most common indirect routes your content takes into AI systems. It respects robots.txt.
- [ChatGPT Search](https://geotoolbox.ai/glossary/chatgpt-search): ChatGPT Search is ChatGPT's ability to retrieve live web results and answer with citations, instead of relying on training data alone. Pages are gathered by OAI-SearchBot, and appearing in its answers is a separate goal from using ChatGPT to write content.
- [ChatGPT-User](https://geotoolbox.ai/glossary/chatgpt-user): ChatGPT-User is the OpenAI agent that fetches a specific web page in real time when a user's ChatGPT prompt requires it, such as following a link or answering a question about a page. It is distinct from GPTBot (training) and OAI-SearchBot (search indexing).
- [ClaudeBot](https://geotoolbox.ai/glossary/claudebot): ClaudeBot is Anthropic's training crawler, which gathers publicly available content that may train future Claude models. It is one of three Anthropic bots: Claude-SearchBot indexes pages for Claude's web search, and Claude-User fetches pages live during a user's session. Each respects robots.txt and is controlled separately.
- [Content Chunking](https://geotoolbox.ai/glossary/content-chunking): Content chunking is structuring a page into self-contained units, each making sense on its own, so AI systems can retrieve and cite a single passage cleanly. Engines pull chunks, not whole pages, so a section that stands alone without 'as mentioned above' is far easier to quote.
- [Context Window](https://geotoolbox.ai/glossary/context-window): A context window is how much text, measured in tokens, a language model can hold at once — the prompt, any retrieved sources, and the model's own generated answer, including the hidden reasoning a thinking model produces before it replies. In AI search, retrieved pages are loaded into the context window so the model can ground its answer in them.
- [E-E-A-T](https://geotoolbox.ai/glossary/e-e-a-t): E-E-A-T stands for Experience, Expertise, Authoritativeness, and Trustworthiness, the qualities Google's Search Quality Rater Guidelines use to judge content. It is not a direct ranking factor, but the signals behind it (first-hand experience and corroborated credibility) are what make content safer to cite, including in AI answers.
- [Entity SEO](https://geotoolbox.ai/glossary/entity-seo): Entity SEO is optimizing around clearly-defined entities (your brand, people, products, and concepts) and the relationships between them, rather than around keyword strings. It helps search and AI engines recognize who you are and connect facts to you, which makes you safer to cite.
- [Featured Snippet](https://geotoolbox.ai/glossary/featured-snippet): A featured snippet is the boxed answer Google pulls from a ranking page and shows at the very top of results, above the blue links. It predates AI search but rewards the same thing AI engines do: a concise, directly-stated answer that can be lifted out of the page.
- [Fine-Tuning](https://geotoolbox.ai/glossary/fine-tuning): Fine-tuning is further training a pretrained model on a smaller, focused dataset so it specializes in a task, tone, or domain. The model keeps its general knowledge but adjusts its weights toward your examples, which is why a fine-tuned model can sound on-brand or follow a niche format that prompting alone struggles to enforce.
- [Generative Engine Optimization (GEO)](https://geotoolbox.ai/glossary/generative-engine-optimization): Generative engine optimization (GEO) is the practice of structuring your content and brand presence so AI engines like ChatGPT, Perplexity, and Google AI Overviews cite, quote, and recommend you in their answers. Where SEO competes for a ranked link, GEO competes to be part of the synthesized answer itself.
- [Google AI Mode](https://geotoolbox.ai/glossary/google-ai-mode): Google AI Mode is Google's conversational, AI-first search experience, powered by Gemini, that returns a generated answer you can follow up on instead of a traditional list of ten links. It runs each question through query fan-out (many background searches at once) and sits alongside AI Overviews as a surface where being a cited source, not a ranked link, is the goal.
- [Google SGE](https://geotoolbox.ai/glossary/google-sge): Google SGE (Search Generative Experience) was the experimental name for Google's AI-generated answers in Search, launched in Labs in 2023. It graduated and was rebranded as AI Overviews, which rolled out broadly in May 2024, so SGE is now the retired label for the same surface.
- [Google-Extended](https://geotoolbox.ai/glossary/google-extended): Google-Extended is a robots.txt token that controls whether your content can be used to train Google's AI models, including Gemini, and to ground Gemini's answers. It does not affect your ranking in Google Search: blocking it affects Gemini training and grounding, not search.
- [GPTBot](https://geotoolbox.ai/glossary/gptbot): GPTBot is OpenAI's web crawler that gathers publicly available content which may be used to train its models. You control it through robots.txt. It is separate from OAI-SearchBot, the crawler that surfaces pages in ChatGPT's search answers, so blocking GPTBot opts you out of training without removing you from ChatGPT search.
- [Grounding](https://geotoolbox.ai/glossary/grounding): Grounding is the practice of tying an AI model's answer to verifiable external sources retrieved at query time, rather than relying on the model's internal memory alone. Grounded answers cite where each claim came from, which reduces hallucination and makes your content quotable.
- [Knowledge Graph](https://geotoolbox.ai/glossary/knowledge-graph): A knowledge graph is a structured network of entities (people, places, brands, concepts) and the relationships between them, often modeled as triples (subject, relation, object). Google's Knowledge Graph powers knowledge panels and helps engines understand who you are. A clear, consistent entity presence makes you easier to recognize and cite.
- [Knowledge Panel](https://geotoolbox.ai/glossary/knowledge-panel): A Google Knowledge Panel is the box of facts about an entity (a brand, person, or place) that appears in search results, generated automatically from the Knowledge Graph. It is the clearest sign Google recognizes you as an entity, and it cannot be requested, only earned and then claimed.
- [Large Language Model (LLM)](https://geotoolbox.ai/glossary/large-language-model): A large language model (LLM) is an AI model trained on vast amounts of text to understand and generate language. LLMs like GPT, Gemini, and Claude power AI search engines, producing answers from patterns learned in training plus sources retrieved at query time.
- [LLM SEO (LLMO)](https://geotoolbox.ai/glossary/llm-seo): LLM SEO, sometimes called LLMO, is optimizing content to be surfaced and cited by large language model tools like ChatGPT and Claude. It frames the work around the models specifically, but in practice it is the same discipline as generative engine optimization (GEO) and answer engine optimization (AEO).
- [llms.txt](https://geotoolbox.ai/glossary/llms-txt): llms.txt is a markdown file at a site's root that gives AI systems a curated map of its most important content. As of 2026 it is not a Google Search or AI Overviews ranking signal, but Google's Chrome Lighthouse now checks for it in an agentic-browsing audit (a missing file is marked Not Applicable, not a failure), so it is becoming low-cost infrastructure for helping AI agents navigate your site.
- [meta-externalagent](https://geotoolbox.ai/glossary/meta-externalagent): meta-externalagent is Meta's web crawler, used to gather public content to train its AI models and index the web. It identifies with the user agent meta-externalagent and respects robots.txt. It is distinct from Meta-ExternalFetcher, which fetches links on a user's behalf.
- [Mixture of Experts (MoE)](https://geotoolbox.ai/glossary/mixture-of-experts): A mixture of experts is a model architecture split into many smaller 'expert' sub-networks, where a routing layer activates only a few experts for each token instead of running the whole model every time. This lets a model hold a very large number of parameters while keeping the compute (and cost) per answer low.
- [Multimodal AI](https://geotoolbox.ai/glossary/multimodal-ai): Multimodal AI is a model that can work with more than one type of data: text, images, audio, and video, in a single system rather than handling text alone. A multimodal model can read a screenshot, describe a chart, transcribe speech, and answer questions about a video, often in the same conversation.
- [Named Entity Recognition (NER)](https://geotoolbox.ai/glossary/named-entity-recognition): Named entity recognition (NER) is the natural language processing step that detects and classifies the entities (people, brands, places, products) mentioned in text. It is how search and AI engines turn a string of words into known things they can connect to facts in a knowledge graph.
- [OAI-SearchBot](https://geotoolbox.ai/glossary/oai-searchbot): OAI-SearchBot is OpenAI's crawler that surfaces and links websites in ChatGPT's search answers. It respects robots.txt and is separate from GPTBot (training): if you block OAI-SearchBot you can disappear from ChatGPT search results, even though you stay eligible for training.
- [Open Source AI](https://geotoolbox.ai/glossary/open-source-ai): Open source AI refers to AI models released under licenses that let anyone use, study, modify, and share them. In the strict Open Source Initiative sense that requires open code, weights, and enough training detail to recreate the model. In practice the label is used loosely: many models called 'open source' are really open weights, with the data and recipe kept private.
- [Open Weights](https://geotoolbox.ai/glossary/open-weights): An open-weights model is one whose trained parameters (the weights) are published for download, so you can run and usually fine-tune them, subject to the license. It is not the same as open source: the weights are released, but the training code, data, and recipe usually are not, so you cannot fully reproduce or audit how it was built.
- [Perplexity-User](https://geotoolbox.ai/glossary/perplexity-user): Perplexity-User is the agent Perplexity uses to fetch a specific page in real time when a user's question requires it. Because the request is user-initiated, it generally ignores robots.txt, so a robots.txt block will not stop a direct user-driven fetch.
- [PerplexityBot](https://geotoolbox.ai/glossary/perplexitybot): PerplexityBot is Perplexity's crawler, designed to surface and link websites in Perplexity's search results. Perplexity states it is not used to train foundation models and recommends allowing it in robots.txt. Blocking it removes you from the index Perplexity builds answers from.
- [Query Fan-Out](https://geotoolbox.ai/glossary/query-fan-out): Query fan-out is the technique where an AI search engine expands a single user question into many parallel sub-queries, retrieves results for each, and synthesizes one answer. Ask for a 5-day trip to Japan and it quietly searches hotels, weather, train passes, and more at once.
- [Reasoning Model](https://geotoolbox.ai/glossary/reasoning-model): A reasoning model is a large language model trained to work through a problem step by step before giving its final answer, rather than responding in one pass. Examples include OpenAI's GPT-5.6 reasoning modes (the successor to its retired o-series), DeepSeek V4's thinking mode (the successor to DeepSeek-R1), Gemini Deep Think, and Claude's adaptive thinking. The extra 'thinking' improves accuracy on hard math, coding, and multi-step logic, at the cost of more time and tokens.
- [Retrieval-Augmented Generation (RAG)](https://geotoolbox.ai/glossary/retrieval-augmented-generation): Retrieval-augmented generation (RAG) is the technique behind most AI search: instead of answering only from memory, the model retrieves relevant documents at query time and grounds its generated answer in them, then cites the sources it used. It is why fresh, reachable, extractable pages can be quoted right away.
- [robots.txt](https://geotoolbox.ai/glossary/robots-txt): robots.txt is a plain-text file at the root of a site that tells crawlers which paths they may or may not fetch, by user agent. For AI search it is the primary control for allowing or blocking crawlers like GPTBot, OAI-SearchBot, ClaudeBot, and PerplexityBot. Google is the exception: Google-Extended is a control token, not a crawler, and it only opts you out of Gemini model training and grounding — appearance in AI Overviews and AI Mode is governed by Googlebot's robots.txt access plus the nosnippet and max-snippet preview controls. Most well-behaved AI crawlers respect it.
- [Schema Markup (Structured Data)](https://geotoolbox.ai/glossary/schema-markup): Schema markup is code, usually JSON-LD using the schema.org vocabulary, that labels what the content on a page means so machines can parse it. It does not force AI citations, and Google says no special schema is required for AI features, but it helps engines understand your entities and content.
- [Semantic Search](https://geotoolbox.ai/glossary/semantic-search): Semantic search matches content to a query by meaning and intent rather than exact keywords, using vector embeddings to compare concepts. In practice retrieval is hybrid — semantic matching runs alongside keyword matching — which is why exact strings like product names, SKUs, and error codes still need to appear on the page verbatim. It is why AI engines can pull your page for a question that does not contain your exact phrasing.
- [Share of Voice (AI)](https://geotoolbox.ai/glossary/share-of-voice): AI share of voice is the percentage of AI answers, across a defined set of prompts, in which your brand is mentioned or cited, compared with competitors. It is the closest thing to a ranking metric in AI search: not a position, but how often you are part of the answer.
- [Topical Authority](https://geotoolbox.ai/glossary/topical-authority): Topical authority is the depth and breadth of credible content a site has on a subject, signaling to search and AI engines that it is a reliable source on that topic. It is built with connected content clusters and consistent entity signals, not a single page.
- [Transformer Model](https://geotoolbox.ai/glossary/transformer-model): The transformer is the neural-network architecture behind nearly every modern large language model, introduced by Google researchers in the 2017 paper 'Attention Is All You Need.' Its key idea, self-attention, lets the model weigh how much every word in the input relates to every other word, which is what makes it so good at language.
- [Vector Embeddings](https://geotoolbox.ai/glossary/vector-embeddings): A vector embedding is a numerical representation of text (or images, audio) that captures its meaning as a point in high-dimensional space. Pieces of content with similar meaning sit close together, which lets AI systems retrieve relevant passages by similarity. Embeddings power semantic search and RAG.
- [Wikidata](https://geotoolbox.ai/glossary/wikidata): Wikidata is a free, collaborative knowledge base of machine-readable facts about entities, structured as items with unique IDs (QIDs) and referenced statements. It feeds Google's Knowledge Graph and is one of the sources AI engines lean on to recognize and describe a brand.
- [Zero-Click Search](https://geotoolbox.ai/glossary/zero-click-search): A zero-click search is a query that ends without the user clicking through to any website, because the answer is shown directly on the results page, in a featured snippet, knowledge panel, or AI Overview. It is the core reason ranking can hold steady while traffic falls.

## Français (20 terms)

- [Glossary index](https://geotoolbox.ai/fr/glossary)
- [Agent IA](https://geotoolbox.ai/fr/glossary/agent-ia): Un agent IA est un système d'IA qui agit pour atteindre un objectif, au lieu de simplement répondre à une question. Là où un chatbot répond, un agent planifie des étapes, utilise des outils (recherche, code, API, navigateur) et mène une tâche en plusieurs étapes à votre place. Les assistants de code, les modes de recherche approfondie et les outils de pilotage d'ordinateur en sont des exemples courants.
- [AI Overviews (aperçus IA)](https://geotoolbox.ai/fr/glossary/ai-overviews): Les AI Overviews (aperçus IA) sont les résumés générés par l'IA de Google en haut de certaines pages de résultats. Produits par Gemini (Gemini 3 par défaut depuis janvier 2026), ils synthétisent une réponse à partir de plusieurs sources web et en citent quelques-unes ; l'internaute obtient donc souvent sa réponse sans cliquer. Déployés en France depuis le 22 juillet 2026, ils font de la citation l'objectif à viser.
- [Answer Engine Optimization (AEO)](https://geotoolbox.ai/fr/glossary/answer-engine-optimization): L'Answer Engine Optimization (AEO) consiste à structurer un contenu pour décrocher la réponse directe dans les moteurs de réponses : featured snippets, résultats de recherche vocale et réponses générées par l'IA. Il recouvre presque entièrement le Generative Engine Optimization (GEO) ; les deux désignent le même travail, décrit sous un angle différent.
- [Balisage Schema.org (données structurées)](https://geotoolbox.ai/fr/glossary/balisage-schema): Le balisage Schema.org — les données structurées — est du code, généralement en JSON-LD, qui décrit le sens du contenu d'une page pour que les machines puissent l'interpréter. Il ne force aucune citation IA, et Google indique qu'aucun balisage particulier n'est requis pour ses fonctionnalités IA, mais il aide les moteurs à comprendre vos entités et vos contenus.
- [Citation IA](https://geotoolbox.ai/fr/glossary/citation-ia): Une citation IA, au sens de la recherche IA, désigne le fait qu'un moteur de recherche IA reprenne votre page comme source dans sa réponse et y renvoie par un lien. Les citations sont l'unité de base de la visibilité IA : dans un monde sans clic, c'est le fait de figurer parmi les sources citées, et non le positionnement d'un lien bleu, qui rend votre marque visible.
- [E-E-A-T](https://geotoolbox.ai/fr/glossary/e-e-a-t): E-E-A-T désigne l'expérience, l'expertise, l'autorité et la fiabilité, les quatre qualités que les Search Quality Rater Guidelines de Google utilisent pour juger un contenu. Ce n'est pas un facteur de classement direct, mais les signaux qui le sous-tendent — l'expérience vécue et une crédibilité corroborée ailleurs — sont ceux qui font d'un contenu une source que les moteurs peuvent citer en confiance, y compris dans les réponses IA.
- [Fenêtre de contexte](https://geotoolbox.ai/fr/glossary/fenetre-de-contexte): La fenêtre de contexte désigne la quantité de texte, mesurée en tokens, qu'un modèle de langage peut contenir d'un seul coup : le prompt, les sources récupérées et la réponse qu'il génère lui-même, y compris la chaîne de pensée masquée qu'un modèle de raisonnement produit avant de répondre. Dans la recherche IA, les pages récupérées sont chargées dans cette fenêtre pour que le modèle puisse y ancrer sa réponse.
- [Fine-tuning](https://geotoolbox.ai/fr/glossary/fine-tuning): Le fine-tuning consiste à poursuivre l'entraînement d'un modèle pré-entraîné sur un jeu de données plus petit et ciblé, afin qu'il se spécialise sur une tâche, un ton ou un domaine. Le modèle conserve ses connaissances générales mais ajuste ses poids vers vos exemples : c'est ainsi qu'un modèle affiné respecte une voix de marque ou un format de niche que le prompt seul peine à imposer.
- [Generative Engine Optimization (GEO)](https://geotoolbox.ai/fr/glossary/generative-engine-optimization): Le Generative Engine Optimization (GEO) consiste à structurer votre contenu et votre présence de marque pour que les moteurs IA comme ChatGPT, Perplexity ou les AI Overviews (aperçus IA) de Google vous citent, vous reprennent et vous recommandent dans leurs réponses. Là où le SEO se bat pour une place dans les résultats, le GEO vise à faire partie de la réponse générée elle-même.
- [Grand modèle de langage (LLM)](https://geotoolbox.ai/fr/glossary/grand-modele-de-langage): Un grand modèle de langage (LLM) est un modèle d'IA entraîné sur d'immenses volumes de texte pour comprendre et générer du langage. Les LLM comme GPT, Gemini ou Claude alimentent les moteurs de recherche IA : ils produisent des réponses à partir des régularités apprises pendant l'entraînement, complétées par les sources récupérées au moment de la requête.
- [Hallucination IA](https://geotoolbox.ai/fr/glossary/hallucination-ia): Une hallucination IA est une information fausse ou inventée qu'un modèle génère et présente comme un fait. Dans la recherche IA, elle prend la forme d'affirmations erronées, de sources fabriquées ou de détails inexacts sur une marque. La principale parade consiste à ancrer les réponses dans des sources récupérées et citables, d'où l'importance d'un contenu clair et sourcé.
- [llms.txt](https://geotoolbox.ai/fr/glossary/llms-txt): llms.txt est un fichier markdown placé à la racine d'un site qui donne aux systèmes d'IA une carte sélective de ses contenus les plus importants. En 2026, ce n'est pas un signal de classement pour Google Search ni pour les AI Overviews (aperçus IA), mais Chrome Lighthouse le vérifie désormais dans un audit de navigation agentique (un fichier absent est signalé « Non applicable », il n'est pas compté en échec) — une infrastructure peu coûteuse pour aider les agents IA à parcourir votre site.
- [Mixture of Experts (MoE)](https://geotoolbox.ai/fr/glossary/mixture-of-experts): Le mixture of experts (MoE) est une architecture de modèle découpée en de nombreux sous-réseaux « experts » plus petits, où une couche de routage n'active que quelques experts par token au lieu de faire tourner le modèle entier à chaque fois. Un modèle peut ainsi contenir un très grand nombre de paramètres tout en gardant un coût de calcul faible par réponse.
- [Modèle de raisonnement](https://geotoolbox.ai/fr/glossary/modele-de-raisonnement): Un modèle de raisonnement est un grand modèle de langage (LLM) entraîné à dérouler un problème étape par étape avant de donner sa réponse finale. Exemples : les modes de raisonnement de GPT-5.6 (successeurs de la série o, retirée), le mode réflexion de DeepSeek V4 (successeur de R1), Gemini Deep Think et la réflexion adaptative de Claude. Cette réflexion supplémentaire améliore la justesse en mathématiques, en code et en logique à plusieurs étapes, au prix de plus de temps et de tokens.
- [Part de voix (IA)](https://geotoolbox.ai/fr/glossary/part-de-voix): La part de voix (share of voice) IA est le pourcentage de réponses IA, sur un ensemble défini de prompts, dans lesquelles votre marque est mentionnée ou citée, face à vos concurrents. C'est ce qui se rapproche le plus d'un indicateur de positionnement dans la recherche IA : non pas une place, mais la fréquence à laquelle vous faites partie de la réponse.
- [Poids ouverts (open weights)](https://geotoolbox.ai/fr/glossary/poids-ouverts): Un modèle à poids ouverts est un modèle dont les paramètres entraînés (les poids) sont publiés en téléchargement : vous pouvez l'exécuter et, en général, le fine-tuner, selon sa licence. Ce n'est pas la même chose que l'open source : les poids sont diffusés, mais le code d'entraînement, les données et la recette ne le sont généralement pas, si bien qu'on ne peut ni reproduire ni auditer complètement sa fabrication.
- [RAG (génération augmentée par récupération)](https://geotoolbox.ai/fr/glossary/rag): La RAG (génération augmentée par récupération) est la technique qui sous-tend la plupart des recherches IA : au lieu de répondre uniquement de mémoire, le modèle récupère des documents pertinents au moment de la requête, y ancre la réponse qu'il génère, puis cite les sources utilisées. C'est pourquoi une page à jour, accessible et facile à extraire peut être citée immédiatement.
- [Recherche sémantique](https://geotoolbox.ai/fr/glossary/recherche-semantique): La recherche sémantique associe un contenu à une requête selon le sens et l'intention plutôt que selon les mots-clés exacts, en comparant les concepts au moyen d'embeddings vectoriels. En pratique, la récupération est hybride : l'appariement sémantique fonctionne en parallèle de l'appariement par mots-clés, d'où la nécessité que les chaînes exactes — noms de produits, références, codes d'erreur — figurent telles quelles sur la page. C'est ce qui permet à un moteur IA de retenir votre page pour une question qui ne reprend pas votre formulation.
- [robots.txt](https://geotoolbox.ai/fr/glossary/robots-txt): Le fichier robots.txt est un fichier texte placé à la racine d'un site qui indique aux robots d'exploration, agent par agent, les chemins qu'ils peuvent ou non récupérer. Pour la recherche IA, c'est le principal levier pour autoriser ou bloquer des robots comme GPTBot, OAI-SearchBot, ClaudeBot et PerplexityBot — Google faisant exception, via l'identifiant de contrôle Google-Extended. La plupart des robots IA légitimes le respectent.
- [Visibilité IA](https://geotoolbox.ai/fr/glossary/visibilite-ia): La visibilité IA mesure la fréquence et la tonalité avec lesquelles les moteurs IA comme ChatGPT, les AI Overviews (aperçus IA) de Google, Gemini et Perplexity font apparaître votre marque dans leurs réponses. C'est l'équivalent du positionnement à l'ère de l'IA : au lieu de votre place dans une liste de liens, elle indique si l'IA vous mentionne, vous cite ou vous recommande.