AI Search Glossary for Marketers

AI search basics

Answer engine. Any system that answers a question directly by synthesizing sources into one written response. ChatGPT, Perplexity, Gemini and Google's AI features all work this way.

AI Overview. The AI-written summary that appears at the top of some Google results, with links to the pages it cites.

AI Mode. Google's conversational AI search tab. Its activity has counted in the Web totals of Google Search Console since June 2025.

Large language model (LLM). The kind of AI model behind answer engines, trained on large amounts of text to predict and generate language.

Prompt. The question or instruction a person types into an AI assistant. In AI search measurement, a prompt also means a buyer-style question used to test which brands an engine names.

Generative engine optimization (GEO). Work that aims to make a brand appear, and appear accurately, in answers written by AI systems.

Answer engine optimization (AEO). A near synonym for GEO, used for the same practice of earning mentions and citations in AI answers.

Search engine results page (SERP). The page a search engine returns for a query, including links, ads and features such as AI Overviews.

Organic result. A listing on the results page that a site earns by ranking. Paid ads are counted separately.

Featured snippet. A short excerpt from one page that Google shows above the regular results to answer a query.

Zero-click search. A search that ends without a click to any website.

How engines find and use sources

Query fan-out. Google's own term for splitting one question into many related sub-searches that run at the same time before the answer is written. Credit it to Google, which introduced it when it announced AI Mode. Source

Retrieval. The step in which an engine searches an index and pulls in pages to read before it answers.

Retrieval-augmented generation (RAG). A design in which a language model retrieves documents first and then writes its answer from them.

Search index. The store of pages an engine actually reaches when it searches. An assistant may rely on an index run by a different company from its own, so being findable in one search engine does not guarantee being reachable by every assistant.

Training data. The text a model learned from before release. Anything the model knows without searching comes from this data.

Model memory. What a model recalls from its training without fetching any page. Engines often draw on memory to decide which brands belong in an answer before any search runs.

Grounding. Tying an AI answer to specific retrieved sources so that its claims can be checked.

Hallucination. A statement an AI model presents as fact that is false or unsupported by any source.

Crawler. A program that fetches web pages for a search engine or AI system.

Client-side rendering. Building page content in the browser with JavaScript after the page loads. A crawler that reads only the HTML the server sends may see an empty page.

Citation. A link to a source page shown inside an AI answer.

Mention. A brand named in the text of an AI answer, with or without a link. A brand can be mentioned often and cited rarely, and the reverse also happens.

Entity. A distinct thing a search system recognizes, such as a company, product or person, along with the facts it associates with it.

Multi-turn session. A real conversation in which a person adds details over several messages, so each answer narrows what the next one can include. Few measurement tools reproduce this.

Single-turn prompt. One question asked of an engine with no earlier context. It is the standard measurement method in the field, and it captures a conversation that real buyers rarely have.

Search Console and click measurement

Google Search Console. Google's free tool that reports the clicks, impressions and positions a site earns in Google Search.

Impression. One appearance of a URL on a results page a user loaded, counted once per page. Google has stated that impressions were over-reported from 13 May 2025 until a fix in late April 2026, while clicks were unaffected. Source

Click. A visit to a site from a search result.

Click-through rate (CTR). Clicks divided by impressions.

Average position. The mean ranking of a site's pages for a query in Search Console.

Branded query. A search that contains a company, product or brand name. Branded queries tend to hold up, and sometimes improve, when an AI Overview appears.

Non-branded query. A search with no brand name in it, such as a category or how-to question. This is where AI answers take the most clicks.

Great Decoupling. The split, first named in May 2025, in which Search Console impressions rose while clicks fell.

num=100. A URL parameter that loaded 100 results on one page. Google stopped honoring it around 10 September 2025, and impressions fell on most sites. Source

Core update. A broad change Google makes to its ranking systems several times a year. Rankings move during the rollout weeks, which can distort before and after comparisons. Source

Search Generative AI performance reports. Search Console reports, in limited rollout since May 2026, that show impressions in Google's AI features by page, country and device. They include no clicks and no queries. Source

Clickstream. Records of real browsing collected from panels, browser extensions and apps.

AI referral. A visit that arrives from an AI assistant, which analytics tools show as the referring site.

Second search. The search a buyer runs for a specific brand or product after an AI answer has narrowed the choice.

Measuring AI visibility

Visibility score. The share of tracked AI answers that mention a brand. Vendors count in different ways, so scores from two tools are not comparable; track the trend inside one tool.

Citation share. The share of all sources cited in a set of answers that belong to one domain.

Prompt types. The kinds of questions used to test engines, each with an expected kind of answer. Common types ask for brand names, a verdict between competitors, a yes or no, a checkable fact, pros and cons, or the criteria buyers use to choose.

Trigger word. A category noun such as software, tool, platform or provider. Without one, a question tends to return general advice in place of brand names.

Coverage. How many differently worded prompts a topic needs before the result stops depending on phrasing. The field agrees coverage matters, and no published source sets a number for it.

Prompt volume. A vendor estimate of how often people ask AI assistants about a topic. No engine publishes this data and vendors do not disclose their models, so treat it as the least verifiable number in the field.

Attribute. A specific quality a brand wants to be known for. Prompts are grouped under attributes, and results are read at that level.

Category frame. The category a brand chooses to be judged within, a concept from April Dunford's work on positioning. Changing the frame changes which competitors an engine compares a brand against.

Evidence and method

Modeled estimate. A figure calculated from public data and published averages, such as keyword volumes and click penalties. It is useful for planning and should always be labeled as modeled.

Observed measurement. A figure taken from real data about a site or its users, such as its own Search Console. It is stronger evidence than a modeled estimate.

Baseline. A measurement kept unchanged so later results can be compared with it. Without a fixed baseline, change cannot be shown.

Control group. A matched set of pages or queries left untouched, so the change in the treated set can be judged against it. Without one, a before and after result shows movement and leaves the cause open.

Primary source. The earliest traceable origin of a figure, such as the study, the dataset or the press release, as distinct from the articles that repeat it.

Corroboration. Agreement among separate original sources. Fifty articles that trace back to one press release corroborate a claim once, and agreement between sources with opposing incentives counts for more.

Factual accuracy. Whether a claim an AI answer makes about a brand, such as a price, certification or integration, matches reality.

Files and technical signals

llms.txt. A proposed file meant to guide AI systems through a site's content. Independent studies found most published files receive no requests, and Google has confirmed it does not use them, so treat it as harmless and unproven. Source

robots.txt. A file at the root of a site that tells crawlers which pages they may fetch. Many AI companies publish the names of their crawlers so sites can allow or block them.

Structured data. Code added to a page, often in the schema.org vocabulary, that labels facts such as product names, prices and reviews for machines to read.

Sources

• Expanding AI Overviews and introducing AI Mode, Google, Robby Stein, 5 March 2025
• Data anomalies in Search Console, Google Search Console Help, entry dated 3 April 2026
• Google modifies search results parameter, affecting SEO tools, Search Engine Journal, September 2025
• Google Search Status Dashboard, ranking incident history, Google
• Search Generative AI performance reports, Google Search Central Blog, June 2026
• The llms.txt proposal, llmstxt.org