GEO glossary: 49 AI search terms, defined and sourced

Naming variants: AI SEO, LLMO (LLM optimization), GSO (generative search optimization) and AIO; no primary source we found defines them as separate disciplines. Spanish: optimización de buscadores generativos (Google’s Spanish guide).

The paper that introduced GEO built GEO-bench, 10,000 queries from 25 domains, and rewrote web sources in different ways to test which changes raised their visibility in generated answers. Answers came from gpt-3.5-turbo working on the top five Google results. In the main results (arXiv v3, Table 1), adding quotations raised Position-Adjusted Word Count (a source’s share of the answer’s words, weighted by position) by about 41% over unmodified sources, adding statistics by about 31% and citing sources by about 27%; keyword stuffing scored below the unmodified baseline. On Perplexity.ai the gains reached up to 37%, and results varied by domain. C-SEO Bench (NeurIPS 2025) retested eight of the paper’s methods with newer models and found most such methods “largely ineffective”, with traditional SEO working better.

Evidence

Peer-reviewed origin; a later peer-reviewed benchmark found most such optimization methods largely ineffective.

Related

Answer engine optimization (AEO)Generative enginePosition-Adjusted Word CountCitable content

Go deeper

GEO vs SEO: which to prioritize first

Spanish: optimización de buscadores de respuestas (Google’s Spanish guide). Google lists it next to GEO as a label for the same kind of work.

The label grew out of voice search. A Search Engine Watch article of February 7, 2018, argued that voice search was “transforming search engines into ‘answer engines’”, since most voice searches get a single answer read out by a digital assistant, and that optimizing for those answers needed a different strategy from classic SEO. The term now covers AI answers as well. Google’s guide says of AEO and GEO: “These are both terms you may see used to describe work specifically focused on improving visibility in AI search experiences.” It adds that, from Google Search’s perspective, this work is still SEO. Google’s page on third-party SEO advice names both labels and recommends checking such advice against its official guidance. In our engine test (September 29, 2026), every definition of AEO we collected was correct.

Evidence

A practitioner label with a disputed origin; Google names it but treats the work as SEO.

Related

Generative engine optimization (GEO)Generative engineStructured dataGoogle AI Overviews

Go deeper

Our GEO and AEO services, with deliverables per tier

Spanish: buscador generativo, buscador de respuestas, búsqueda con IA (Google’s Spanish guide says búsqueda con IA generativa).

The GEO paper introduced the term for search engines “that use generative models to gather and summarize information to answer user queries”, typically by “synthesizing information from multiple sources”. Answer engine is older: in 2018 Search Engine Watch used it for search engines that give one spoken answer to a voice query, and Perplexity, which presents its product as one, defines an answer engine as “a tool designed to give you direct, detailed answers to your questions”. AI search has no formal definition; Google applies it to “generative AI features on Google Search (such as AI Overviews and AI Mode)”. The three labels overlap: each describes a system that replies with a written answer built from sources, where a classic search engine returns a list of links.

Evidence

Only generative engine has a peer-reviewed definition; answer engine and AI search are industry labels.

Related

Generative engine optimization (GEO)Answer engine optimization (AEO)Google AI ModeLarge language model (LLM)

Go deeper

AI search optimization: six workstreams, in impact order

Spanish: búsqueda sin clic.

SparkToro measures it with clickstream panels from other companies, and the supplier has changed: Jumpshot in 2019, Datos (a Semrush company) in 2024 and Similarweb in 2026. The 2024 study found that 58.5% of US and 59.7% of EU Google searches ended without a click; the 2026 study found 68.01% for US searches from January to April 2026. Google had disputed the measure in 2021: Danny Sullivan wrote that it rests on “flawed methodology that misunderstands how people use Search”, since a search can end without a click because someone refines the query, gets a quick fact such as a sports score, or calls a business. Independent data from Pew Research Center (900 US adults, March 2025) found clicks on a traditional result in 8% of visits to a results page with an AI summary, against 15% without one, and on a link inside the summary in 1%.

Evidence

Measured by SparkToro, which sells audience research software, with third-party clickstream panels; Google disputes the framing.

Related

Google AI OverviewsGoogle AI ModeAI referral traffic

Spanish: modelo de lenguaje grande (Google’s Spanish glossary); Google’s Spanish documentation also uses modelo de lenguaje extenso.

Google’s Machine Learning Glossary defines an LLM as “At a minimum, a language model having a very high number of parameters” and, informally, “any Transformer-based language model, such as Gemini or GPT”. An LLM writes by predicting one token after another; Google’s Machine Learning Crash Course calls LLMs “essentially autocomplete mechanisms”. What the model learns in training is stored in its parameters, which can number in the hundreds of billions or trillions, so fluent output can still be wrong: the course says “LLMs hallucinate, meaning their predictions often contain mistakes.” In AI search, the model is paired with a retrieval step: the system first finds pages, then the model writes the answer from them. An answer can therefore reflect training, retrieved pages or both.

Related

HallucinationGroundingRetrieval-augmented generation (RAG)Training data vs live retrieval

Spanish: generación aumentada por recuperación (RAG). Google Search Central treats RAG and grounding as synonyms.

Lewis and colleagues described RAG models as ones “which combine pre-trained parametric and non-parametric memory for language generation”: a sequence-to-sequence model holds the parametric memory, and a dense vector index of Wikipedia, searched by a neural retriever, holds the non-parametric memory. The retrieved passages give the model information it did not store in training. Google’s guide defines RAG as “A technique (also known as grounding) used to improve the quality, accuracy, and freshness of AI responses”, and says it relies on Google’s core Search ranking systems to retrieve relevant, up-to-date pages from the Search index. So for Google’s AI features, a page has to be retrievable first: to appear as a supporting link in AI Overviews or AI Mode, it “must be indexed and eligible to be shown in Google Search with a snippet”.

Related

GroundingTraining data vs live retrievalHallucinationQuery fan-out

Spanish: fundamentación (Google’s Spanish documentation). Google Search Central uses it as a synonym of retrieval-augmented generation (RAG).

Google uses the word in two ways. In the Gemini API, “Grounding with Google Search connects the Gemini model to real-time web content”, and the response carries annotations linking parts of the answer to their sources; Google says this lets Gemini “cite verifiable sources beyond its knowledge cutoff”. Google Search Central, by contrast, defines retrieval-augmented generation as “also known as grounding”. For teaching, it helps to keep them apart: grounding is the goal, an answer tied to sources that can be checked, and RAG is one way to reach it. Google’s AI features page also separates grounding from training: to limit either in some of Google’s other systems, it points site owners to Google-Extended.

Related

Retrieval-augmented generation (RAG)HallucinationCitationGoogle-Extended

Also called fan-out queries. Spanish: ramificación de búsquedas (Google also uses distribución múltiple de consultas).

Google named the technique when it launched AI Mode, and its documentation says both AI Overviews and AI Mode may use it. Google’s own example: for “how to fix a lawn that’s full of weeds”, fan-out queries might include “best herbicides for lawns”, “remove weeds without chemicals” and “how to prevent weeds in lawn”. Each sub-query brings back its own results, so a page can be cited for one part of a question without ranking for the question as a whole. Google does not publish how many sub-queries a question produces or which ones it runs, so tools that list “your fan-out queries” are estimating them.

Related

Google AI ModeGoogle AI OverviewsRetrieval-augmented generation (RAG)Scaled content abuse

Go deeper

How prompt and query research works in AI search optimization

Also framed as parametric vs non-parametric memory. Spanish: datos de entrenamiento frente a búsqueda en tiempo real; knowledge cutoff is fecha límite de conocimiento in OpenAI’s Spanish help center.

Research calls the two sources parametric and non-parametric memory. Petroni et al. (2019) analyzed the relational knowledge that pre-trained language models absorb from their training data, and Lewis et al. (2020), starting from the finding that such models “store factual knowledge in their parameters”, added a searchable index as a second memory. Parametric knowledge ends at the model’s knowledge cutoff, the date of its most recent training data, and stays fixed until the model is retrained. Retrieved knowledge can be as recent as the index being searched. For GEO, the consequence is direct: a page published after a model’s cutoff can reach that model’s answers only through retrieval, while material the model absorbed in training can shape answers even when nothing is searched.

Related

Retrieval-augmented generation (RAG)Large language model (LLM)ChatGPT searchFreshness

Also called source attribution or a source link. Google’s documentation calls the links in AI Overviews and AI Mode “supporting links”. Spanish: cita.

AI search products attach citations so readers can check where a statement came from: ChatGPT responses that use web search may include citations, Perplexity adds numbered citations to each answer, and Google’s AI Overviews and AI Mode show supporting links to web pages. Research treats this as a question of attribution: the Attributable to Identified Sources (AIS) framework of Rashkin et al. holds that generated statements about the world should be verifiable against an identified, independent source. A link alone does not guarantee this. In a human evaluation of four generative search engines (Bing Chat, NeevaAI, perplexity.ai and YouChat), Liu, Zhang and Liang found that, on average, only 51.5% of generated sentences were fully supported by their citations. For a website, being cited means one of its pages is linked inside the answer, which differs from the brand being named without a link.

Related

Citation rateMisattributionHallucinationGrounding

Tested from 2023 as SGE (Search Generative Experience) in Search Labs. Spanish: vistas creadas con IA, Google’s official name. In our engine test (September 29, 2026), ChatGPT and Claude, answering in Spanish, used other names for it, and Google’s own AI Overview used the English name.

Google first tested the feature in SGE (Search Generative Experience), a Search Labs experiment announced on May 10, 2023, and on May 14, 2024, began rolling out AI Overviews to everyone in the US, built on a Gemini model customized for Google Search. They arrived in Spain, in Spanish and English, on March 25, 2025, for signed-in users aged 18 or over. On May 20, 2025, Google said they were available in more than 200 countries and territories and in more than 40 languages. According to Google’s documentation, AI Overviews and AI Mode may use query fan-out, issuing several related searches across subtopics and data sources, and they show a wider and more varied set of links than a classic web search.

Related

Google AI ModeQuery fan-outSearch generative AI controlZero-click search

Go deeper

Gemini SEO: visibility in Google’s AI answers

Spanish: modo IA, the name Google uses in its Spanish documentation.

Google introduced AI Mode on March 5, 2025, as an early experiment in Labs, inviting Google One AI Premium subscribers first, and began rolling it out to everyone in the US on May 20, 2025. It started rolling out globally in Spanish on September 23, 2025. Google describes it as “particularly helpful for queries where further exploration, reasoning, or complex comparisons are needed”. It relies on query fan-out: it breaks a question into subtopics, issues many searches at once and combines the results into one response with links. SparkToro, using Similarweb’s US desktop and mobile web panel (which excludes the Google app), found that 0.34% of Google searches reached AI Mode from January to April 2026.

Evidence

The usage figure comes from an industry study by SparkToro, which sells audience research software, using data from Similarweb, which sells web traffic data.

Related

Google AI OverviewsQuery fan-outSnippet controlsZero-click search

Tested in 2024 as the SearchGPT prototype. OpenAI’s help article is now titled “Searching the web with ChatGPT”. Spanish: búsqueda de ChatGPT.

OpenAI tested SearchGPT, a prototype shown to a small group of users and publishers, from July 25, 2024, and launched ChatGPT search on October 31, 2024, for Plus and Team subscribers and SearchGPT waitlist users. It reached all logged-in users on December 16, 2024, and everyone in regions where ChatGPT is available on February 5, 2025. ChatGPT may search automatically when a question would benefit from current information, or the user can choose Search from the tools menu. OpenAI says it uses third-party search providers and content from its partners, and that after reviewing the first results it may send more specific queries to other providers. For search, OpenAI runs a dedicated crawler, OAI-SearchBot, separate from GPTBot, which crawls content that may be used for training.

Related

AI crawlers and user agentsrobots.txtCitationAI referral traffic

Go deeper

How to rank in ChatGPT: what earns a citation

Also: hallucinated content. Research distinguishes intrinsic from extrinsic hallucination. Spanish: alucinación.

Two definitions are in use. Google’s machine learning glossary treats it as a factual error: plausible-seeming output that is factually incorrect, from a model presenting it as a claim about the real world. Ji et al.’s survey (ACM Computing Surveys, 2023) measures it against the source content instead and, following earlier work, splits it in two: intrinsic hallucination contradicts the source, and extrinsic hallucination can be neither supported nor contradicted by it. In AI search, where the source is the set of retrieved pages, an intrinsic hallucination misstates what a cited page says and an extrinsic one adds a claim that no retrieved page makes. Research published by OpenAI (vendor research, September 5, 2025) argues that “language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty.”

Related

GroundingRetrieval-augmented generation (RAG)CitationMisattribution

Also: browser agents. Google’s guide speaks of “agentic experiences”. “Agentic search” is also used, with no formal definition. Spanish: agentes de IA.

Google’s guide to its generative AI features defines AI agents as autonomous systems that can perform tasks on behalf of people, and says browser agents may gather what they need from a site by analyzing screenshots, inspecting the DOM and interpreting the accessibility tree. The web.dev guide (April 1, 2026) adds that modern agents combine these inputs: they take a structured list of interactive elements from the DOM and the accessibility tree and cross-check it against a visual rendering. Its advice follows from that: semantic HTML (real buttons and links rather than styled div elements), labels linked to their inputs, stable layouts and no transparent overlays over controls. OpenAI’s ChatGPT agent (July 17, 2025) works on its own virtual computer with a visual browser, a text-based browser and a terminal; ChatGPT Atlas (October 21, 2025) is a browser built around ChatGPT, with an agent mode.

Evidence

“Agentic search” has no formal definition; Google documents AI agents and browser agents.

Related

AI crawlers and user agentsServer-side renderingllms.txtChatGPT search

Tools use the same label for different metrics: Semrush’s AI Visibility score and Peec AI’s Visibility. Spanish: visibilidad en IA.

It has no single formula. Semrush’s AI Visibility is a benchmark score from 0 to 100 combining topic coverage (how many topics include the brand in AI answers, compared with all other domains) and mention consistency (how often the brand is mentioned across responses within those topics). Peec AI’s Visibility is a mention rate: responses mentioning the brand ÷ total responses × 100. The GEO paper defined a website’s visibility in a generative engine’s answer through impression metrics, such as the share of the answer’s words in sentences citing it. Microsoft’s AI Performance report in Bing Webmaster Tools (public preview, February 10, 2026) tracks citations instead, with Total Citations, Average Cited Pages and Grounding Queries (the phrases the AI used when retrieving cited content). Two visibility figures are comparable only if calculated the same way, on the same prompts, engines and dates.

Evidence

Practitioners agree on the broad meaning, but each tool calculates it with its own formula.

Related

Mention rateCitation rateAI share of voicePrompt set and prompt tracking

Go deeper

The free AI Visibility Diagnosis: where your brand stands on five engines

Also called AI SOV. Bing Webmaster Tools reports a citations-only version called Citation Share. Spanish: cuota de voz en IA.

The term comes from advertising, where share of voice meant a brand’s share of the media exposure in its category. AI tools count different things under the same name: Semrush counts the brand’s mentions against competitors’, Ahrefs weights mentions by the search volume of the prompts (a share of “impressions”), and Microsoft’s Bing Webmaster Tools reports Citation Share, a site’s share of all citations shown for one grounding query. A figure means little without its formula, prompt set, engines, dates and competitor list. Answers also vary from run to run: in a SparkToro study of 2,961 runs, there was less than a 1 in 100 chance that ChatGPT or Google’s AI returned the same list of brands twice, so the share should be averaged over many prompts and repeated runs.

Evidence

No standard formula: vendors define it differently.

Related

Mention rateCitation ratePrompt set and prompt trackingAI visibility

Go deeper

How we measure share of voice in the Sourcelift Loop

Peec AI calls it Visibility (Visibility Score). Spanish: tasa de menciones.

Formula: mention rate = answers that mention the brand ÷ total answers × 100. Vendors calculate it the same way under different names: Peec AI’s Visibility Score uses exactly this formula, and ZipTie calls it mention rate (“AI responses naming your brand ÷ Total relevant AI responses monitored”). Semrush instead reports Mentions as a count: “The total number of prompts in which a brand is included in AI responses.” A mention only requires the brand to appear in the answer text; no link or citation is needed. The rate depends on the prompt set and on the run: in a SparkToro study, the same prompt rarely returned the same list of brands twice. Average it over many prompts and repeated runs, and report it with its prompts, engines and dates.

Evidence

Vendors agree on the formula but not on the name.

Related

Citation rateAI share of voicePrompt set and prompt trackingBrand mentions

Spanish: tasa de citas. Bing Webmaster Tools reports Total Citations and Citation Share instead.

No standard formula exists. Peec AI defines it as “how often a URL is cited per AI answer relative to how often it is retrieved”: citations ÷ retrievals, where a retrieval means the page made it into the list of candidates the model might use for its answer. Because one answer can cite a URL several times, this rate can exceed 1; in Peec’s benchmarks, 2.0 means a URL “was cited twice per every AI answer” on average. ZipTie divides “AI responses citing your URL” by “Total relevant AI responses monitored”, a share that cannot exceed 100%. Microsoft’s first-party metrics in Bing Webmaster Tools include Total Citations, the number of citations shown as sources in AI answers over a period, and Citation Share, a site’s percentage of all citations shown for one grounding query. Check which base a figure uses before comparing it.

Evidence

No standard formula: vendors divide by different bases.

Related

Mention rateCitationAI share of voiceCitation gap

Related vendor terms: AI visibility gap and Missing Sources (Semrush), brand gap analysis (Ahrefs). Spanish: brecha de citas.

No standard definition exists, and every tool builds the gap from its own prompts and competitor list. Semrush’s Missing Sources report lists “External websites cited in AI answers that mention competitors but not your brand.” Semrush’s guide to AI visibility gaps separates five kinds (mention, prompt, source, citation and narrative) and describes the citation gap as: “The AI doesn’t cite content from your website in its responses.” Ahrefs calls its version brand gap analysis, which “measures the difference between your brand’s potential visibility and its actual presence across Google, AI results, and the wider web”. A gap is only meaningful with its prompt set, engines, dates and competitors stated; change any of them and the gap can change. Answers also vary between runs: in a SparkToro study, the same prompt rarely returned the same list of brands twice, so a gap seen once may not hold.

Evidence

No standard definition: each tool computes the gap from its own prompts and competitor list.

Related

Citation rateThird-party sources in AI answersPrompt set and prompt trackingAI share of voice

Semrush’s feature is called Prompt Tracking. Search Console classifies Google Search queries, not AI prompts, as branded or non-branded. Spanish: set de prompts, seguimiento de prompts.

Because AI answers change from run to run, visibility is measured over many prompts and runs. In a SparkToro study with Gumshoe.ai (January 28, 2026), 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI Overviews 2,961 times; there was less than a 1 in 100 chance that ChatGPT or Google’s AI gave the same list of brands in any two responses. Rand Fishkin concluded that visibility “across dozens to hundreds of prompts run multiple times is a reasonable metric”. Semrush tracks “daily visibility for a custom set of prompts”; Peec AI reruns prompts every 24 hours. Branded prompts name the brand; non-branded prompts don’t. Peec AI’s default mix is 20% branded. Search Console’s branded queries filter (announced November 2025) makes the same split for Google Search queries only, and Google warns that “Some queries might be incorrectly identified as branded or non-branded.”

Evidence

The data on answer variability come from SparkToro (sells audience research software) with Gumshoe.ai (sells AI visibility monitoring).

Related

AI visibilityMention rateAI share of voiceCitation gap

Go deeper

The best GEO tools in 2026, and what their prompt tracking measures

Both are impression metrics from the GEO paper. Spanish: recuento de palabras ajustado por posición; impresión subjetiva.

Aggarwal et al. proposed it in the GEO paper to measure a source’s visibility inside a generative answer. The simpler Word Count metric is the words in sentences citing the source ÷ all words in the answer, with a sentence’s words split equally when several sources cite it. The position-adjusted version reduces each sentence’s weight “by an exponentially decaying function of the citation position”, so material cited early counts more. Scores are normalized so the impressions of all citations in an answer sum to 1. Subjective Impression covers what word counts miss, including relevance to the query, influence, uniqueness and the probability of clicking the citation; an LLM, GPT-3.5, scored each facet with a method similar to G-Eval. The paper applied both to GEO-bench, 10,000 queries across 25 domains. They are research metrics; the vendor metrics in this glossary are simpler counts of answers or citations.

Evidence

Accepted to KDD 2024, per the paper’s arXiv record.

Related

Generative engine optimization (GEO)Citable contentAI visibilityGenerative engine

Google Analytics channel: AI Assistant (Spanish interface: Asistente de IA). Spanish: tráfico de referencia desde IA.

Google Analytics added an AI Assistant channel to its default channel group on May 13, 2026; the Spanish interface calls it Asistente de IA. Google defines it as “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok”. Classification depends on the referrer: GA sets the medium to ai-assistant and the campaign to (ai-assistant) when the referrer matches its list of AI assistants. Visits from Google’s AI Overviews and AI Mode are not in this channel; they count as Organic Search. Search Console’s generative AI performance report covers AI Overviews and AI Mode but defines only impressions, so neither of Google’s tools documents a separate count of visits from its own AI features.

Related

Generative AI performance reportZero-click searchGoogle AI OverviewsGoogle AI Mode

Spanish interface: informe de rendimiento de IA generativa. Discover has its own generative AI performance report.

Google announced the report on June 3, 2026, and first rolled it out to a subset of site owners in the UK for testing; the Help page says it reached all websites worldwide as of August 31, 2026. It includes impressions for AI Overviews and AI Mode, leaves out Search Labs experiments, and draws on the Web search type of the standard Performance report. Data can be grouped by page, country, date and device. Impressions count “how many times links to your site were shown to a user in a generative AI feature on Google Search”. Google Discover has a separate report. Google announced the report together with the Search generative AI control, an opt-out: sites that opt out “will not receive traffic or impressions from our generative AI features”.

Related

AI referral trafficSearch generative AI controlGoogle AI OverviewsGoogle AI Mode

Also called AI bots. Spanish: rastreadores de IA.

AI companies now run separate bots for separate jobs, and robots.txt rules apply to each name on its own. Training: OpenAI’s GPTBot and Anthropic’s ClaudeBot collect content that may be used to train models. Search: OAI-SearchBot, Claude-SearchBot and PerplexityBot surface pages in the assistants’ search answers, and Applebot feeds search in Spotlight, Siri and Safari. User requests: ChatGPT-User, Claude-User and Perplexity-User may visit a page when someone asks a question. OpenAI says robots.txt rules “may not apply” to these visits and Perplexity says its fetcher “generally ignores robots.txt rules”; Anthropic says disabling Claude-User prevents them. Applebot-Extended “does not crawl webpages”: it is how publishers opt out of training Apple’s models. For AI Overviews and AI Mode, Google points to Googlebot, because AI “is built into Search”.

Related

robots.txtGoogle-ExtendedChatGPT searchServer-side rendering

Go deeper

What the GEO audit checks in crawler access

Formally, the Robots Exclusion Protocol (RFC 9309). Spanish: archivo robots.txt.

The file must sit at the top level of a site as /robots.txt, in lowercase, and is read in groups: one or more user-agent lines naming crawlers, followed by allow and disallow rules for URL paths. Google says the file is used mainly to avoid overloading a site with requests. It governs crawling, not indexing: a URL blocked in robots.txt can still appear in Google’s results without a description, and can be indexed if other sites link to it. For AI, it is where you address each company’s bots by name. OpenAI, for example, lets a site disallow its training crawler, GPTBot, while allowing OAI-SearchBot, the bot that surfaces sites in ChatGPT search.

Related

AI crawlers and user agentsGoogle-ExtendedSnippet controlsllms.txt

Spanish: archivo llms.txt.

Jeremy Howard proposed the format in September 2024 so that language models and agents could read a site’s key content without wading through navigation, scripts and ads. The only required element is an H1 with the site’s name; a short summary in a blockquote and H2 sections listing links are optional. Unlike robots.txt, it allows or blocks nothing, and the proposal has no fields for training permissions, usage rights or attribution. Google’s guide says Google Search ignores these files, including for its generative AI features. Google’s Chrome team takes a different angle: Lighthouse checks for the file because, without it, agents “may spend more time crawling the site to understand its high-level structure”. OpenAI’s and Perplexity’s crawler documentation does not say their bots read it.

Evidence

A proposal with no documented effect on Google Search.

Related

robots.txtAI crawlers and user agentsStructured dataEntity consistency (NAP)

Also called SSR. Related: pre-rendering. The alternative is client-side rendering (CSR). Spanish: renderizado en el servidor.

With client-side rendering, the browser builds the page by running JavaScript that modifies the DOM, so a bot that does not execute scripts may get little text. Googlebot processes such pages in three phases (crawling, rendering and indexing) and queues every page that returns HTTP 200 for rendering, unless told not to index it. Google nonetheless says “server-side or pre-rendering is still a great idea”, because it makes sites faster “for users and crawlers, and not all bots can run JavaScript.” For AI assistants the evidence is one industry study from 2024: monitoring Vercel’s network, Vercel and MERJ found that the major AI crawlers they measured, from OpenAI, Anthropic, Meta, ByteDance and Perplexity, did not render JavaScript, while Gemini (through Googlebot’s infrastructure) and Applebot did.

Evidence

Google’s advice is official documentation; the AI-crawler finding rests on one 2024 study by Vercel, which sells frontend cloud hosting, with the consultancy MERJ.

Related

AI crawlers and user agentsrobots.txtAI agents

Also called schema markup. Formats: JSON-LD, Microdata and RDFa. Spanish: datos estructurados.

Google recommends JSON-LD and also reads Microdata and RDFa. It says the markup gives it “explicit clues about the meaning of a page” and can make the page eligible for rich results; most markup uses schema.org vocabulary, but Google treats its own documentation as definitive for Search. For its generative AI features, Google’s guide says structured data is not required and no special schema.org markup is needed. Rich result types also change: the FAQ rich result stopped appearing in Google Search on May 7, 2026. In our engine test (September 29, 2026), Gemini, in both languages, still recommended FAQPage and HowTo schema for answer engines.

Related

Answer engine optimization (AEO)Entity and Knowledge Graphllms.txtGoogle AI Overviews

Go deeper

How the GEO audit reviews your structured data

Google also calls them preview controls. Spanish: controles de fragmentos (Google’s Spanish pages say controles de vista previa).

Google defines them in its robots meta tag specification. nosnippet removes the text snippet and video preview and also stops the content from being used as a direct input for AI Overviews and AI Mode. max-snippet:[number] caps the snippet at that many characters (0 is equivalent to nosnippet) and also limits how much content can serve as direct input for those features. The data-nosnippet attribute, on span, div or section elements, excludes only the marked text. Google’s AI features page lists these rules, plus noindex, as the way to limit what Search shows from your pages. The trade-off: to be shown as a supporting link in AI Overviews or AI Mode, a page “must be indexed and eligible to be shown in Google Search with a snippet”.

Related

Search generative AI controlGoogle-ExtendedGoogle AI OverviewsGoogle AI Mode

A robots.txt product token, not a crawler. Spanish: Google-Extended (not translated).

Disallowing Google-Extended in robots.txt covers two uses of content Google has crawled: training “future generations of Gemini models that power Gemini Apps and Vertex AI API for Gemini”, and grounding (supplying content from the Search index to the model at prompt time) in Gemini Apps and in Grounding with Google Search on Vertex AI. Google’s AI features page sends site owners to Google-Extended “to limit AI training and grounding in some of Google’s other systems”, while robots.txt rules for Googlebot govern Search itself. The Search Console Help page for the Search generative AI control adds a use inside Search: “To limit training of the models used to generate responses in Search generative AI features, use Google-Extended.” Read together, Google-Extended governs training and Gemini grounding; appearing in AI Overviews and AI Mode depends on Googlebot, snippet controls and that Search Console setting.

Evidence

Two Google pages describe its scope differently; the entry quotes both.

Related

Search generative AI controlAI crawlers and user agentsrobots.txtSnippet controls

In Search Console: Settings > Search generative AI. Spanish: control de la IA generativa de la Búsqueda.

It offers three options: include the site’s links and content (the default for every property), exclude them, or inherit the value of a parent property. Included content can appear as links and help ground AI responses in those features. Excluded content is neither shown nor linked there, so the site gets no traffic or impressions from them; Google says exclusion generally takes a few days. The control does not cover model training: for that, the Help page points to Google-Extended. Until it arrived, Google’s documented ways to limit these features were robots.txt rules for Googlebot and page-level controls (nosnippet, data-nosnippet, max-snippet, noindex), all of which also change how a page appears in regular results.

Related

Google-ExtendedSnippet controlsGenerative AI performance reportGoogle AI Overviews

Stands for experience, expertise, authoritativeness and trustworthiness (the rater guidelines say Trust). Formerly E-A-T. Spanish: experiencia, conocimiento, autoridad y fiabilidad.

E-E-A-T comes from the guidelines for Google’s search quality raters, people who tell Google whether its algorithms “seem to be providing good results”. Google says its ranking systems use several factors to identify content with good E-E-A-T and give it even more weight on “Your Money or Your Life” (YMYL) topics, which could significantly affect people’s health, financial stability or safety, or the welfare of society. The guidelines define Experience as the creator’s “first-hand or life experience for the topic” and call Trust “the most important member of the E-E-A-T family”: an untrustworthy page has low E-E-A-T however expert it seems. Content does not have to show all four. Google’s AI features retrieve pages through the same core ranking systems. In our engine test (September 29, 2026), all ten answers we read said E-E-A-T is not a direct ranking factor, in line with Google.

Related

Non-commodity contentCore update vs spam updateScaled content abuseGoogle AI Overviews

Opposite: commodity content. Spanish: contenido no genérico (Google’s Spanish name), as opposed to contenido genérico.

Google uses the term in its guide to optimizing for generative AI features, added on May 15, 2026. Commodity content, the guide says, “is often based on common knowledge, which could originate from anyone, and typically adds little unique insight for readers”, while non-commodity content “provides unique expert or experienced takes that go beyond common knowledge and the ordinary”. Its examples: a generic “7 Tips for First-Time Homebuyers” versus a first-hand account of why a buyer waived an inspection and saved money. The idea predates AI search: Google’s self-assessment asks “Does the content provide original information, reporting, research, or analysis?”, and its original content systems aim to show original reporting “ahead of those who merely cite it”. Google recommends non-commodity content for its own AI features; its effect on citations in other AI engines is not documented.

Related

E-E-A-TCitable contentScaled content abuseGoogle AI Overviews

Related: knowledge panel. Spanish: entidad; Google’s Spanish documentation calls the Knowledge Graph “Gráfico de conocimiento” and knowledge panels “paneles de información”.

Google introduced the Knowledge Graph on May 16, 2012, to move search from matching keywords to understanding “things, not strings”. Its launch example was the query Taj Mahal, which could mean the monument or the musician: the graph lets Google tell them apart. At launch it held “more than 500 million objects, as well as more than 3.5 billion facts about and relationships between these different objects”. Knowledge panels are the information boxes Google shows when someone searches for an entity in the graph. Google’s public Knowledge Graph Search API returns entries from the graph, but Google says it “is not suitable for use as a production-critical service”. Google’s guide to its generative AI features does not mention entities or the Knowledge Graph, so their role in AI Overviews is not documented there.

Related

Entity consistency (NAP)Brand mentionsStructured dataThird-party sources in AI answers

NAP stands for name, address and phone number. Also called NAP consistency. Spanish: coherencia de entidad, coherencia NAP.

The idea comes from local SEO. Google’s Business Profile guidelines say a profile’s name should “reflect your business’s real-world name, as used consistently on your storefront, website, stationery, and as known to customers”, and Google says businesses “with complete and accurate info are more likely to show up in local search results”. Its generative AI guide adds that Business Profiles “can help your products and services to be visible in both AI responses and other Google Search results”. Beyond Google’s own profiles, the case rests on practitioner reasoning: AI engines take business facts from third-party sources (in a small BrightLocal test, “All LLMs are using directories and citations for business information across every industry”), so mismatched names, addresses or phone numbers hand them conflicting facts. We found no engine documentation or controlled study that measures this effect on AI answers.

Evidence

Google documents name rules for Business Profiles; the effect of consistency across other sites on AI answers is undocumented.

Related

Entity and Knowledge GraphThird-party sources in AI answersStructured dataMisattribution

Includes linked and unlinked mentions; Ahrefs calls them branded web mentions. Not the same as mentions inside AI answers (see mention rate). Spanish: menciones de marca.

GEO looks at web mentions because AI features draw on what other sites say: Google’s guide says its AI features “can show what’s being said about products and services across the web, including in blogs, videos, and forum discussions”. An industry study by Ahrefs, which sells SEO and AI visibility software, looked at 75,000 brands. “Branded web mentions” (the brand’s name anywhere on the web) had the strongest correlation with how often Google’s AI Overviews mentioned a brand (Spearman 0.664), ahead of branded anchor text (0.527) and the number of backlinks (0.218). The study covers AI Overviews only and shows correlation, not cause (“correlation ≠ causation”, Ahrefs notes). It counts mentions with or without a link, so it cannot tell linked from unlinked mentions apart.

Evidence

Industry study by Ahrefs, which sells SEO and AI visibility software; correlation only, and Google AI Overviews only.

Related

Mention rateBacklinksThird-party sources in AI answersAI answer manipulation

Go deeper

Brand mentions in ChatGPT: how to earn yours

Also called inbound links. Spanish: enlaces entrantes, backlinks.

Links are among Google’s oldest signals. Its ranking systems guide says link analysis systems “understand how pages link to each other” to work out what pages are about, and that PageRank, used since Google launched, “continues to be part of our core ranking systems”. Google’s AI features retrieve pages through those core systems. For AI answers generally, the evidence is correlational. Semrush, which sells SEO and AI visibility software, studied 1,000 randomly selected domains with Kevin Indig (published October 16, 2025) across ChatGPT (with and without search), Gemini, Perplexity and Google AI Overviews. Authority Score, Semrush’s proxy for link quality, correlated with how often each domain was mentioned in AI answers (Pearson 0.65, Spearman 0.57), and nofollow links showed nearly the same association as followed links. A correlation cannot show that links cause those mentions.

Evidence

Industry study by Semrush, which sells SEO and AI visibility software, showing correlation only; Google documents link analysis for its own ranking.

Related

Link spamBrand mentionsAI visibilityThird-party sources in AI answers

Includes review platforms, business directories and forums such as Reddit. Spanish: fuentes de terceros en las respuestas de IA.

Vendor studies show AI answers citing these sites, but each measures a different slice at a different time. Local directories: BrightLocal, which sells listings and citation-building software, ran 20 local searches across 10 niches on ChatGPT search, Gemini, Perplexity and Google’s AI Mode, and found Yelp used as a source in 33% of its searches (published July 22, 2025; search dates not given). Software review sites: SE Ranking, which sells SEO software including an AI Overviews tracker, checked 30,000 US commercial keywords on December 1, 2025; 34.5% of the AI Overview responses cited at least one of 23 review platforms, such as G2 and Capterra. Forums: Semrush, which sells SEO and AI visibility software, found ChatGPT cited Reddit in close to 60% of responses in early August 2025 and around 10% by mid-September, in a study covering July 14 to October 12, 2025.

Evidence

Studies by BrightLocal, SE Ranking and Semrush, which sell local SEO, SEO and AI visibility software; samples and dates differ.

Related

CitationBrand mentionsEntity consistency (NAP)AI visibility

Covers three methods from the GEO paper: Cite Sources, Quotation Addition and Statistics Addition. Spanish: contenido citable.

The idea comes from the paper that introduced GEO (Aggarwal et al., first posted November 16, 2023, accepted to KDD 2024), which tested nine ways of rewriting source pages. Cite Sources adds citations from credible sources, Quotation Addition adds quotations from them, and Statistics Addition replaces qualitative statements with quantitative data where possible. In the paper’s main results table (arXiv version 3), these raised the Position-Adjusted Word Count score from 19.3 for unmodified pages to 27.2 (quotations), 25.2 (statistics) and 24.6 (sources); Fluency Optimization scored 24.7 and Keyword Stuffing 17.7. The answers were generated by gpt-3.5-turbo from the top five Google results, and a test on Perplexity.ai showed gains of up to 37%. C-SEO Bench (NeurIPS 2025) retested these three methods with two 2024 models across six domains and found no significant gain from any of them.

Evidence

Peer-reviewed studies disagree: the 2023 GEO paper found gains, and a 2025 benchmark found no significant gain from these methods.

Related

Generative engine optimization (GEO)Position-Adjusted Word CountCitationNon-commodity content

Go deeper

How to optimize for ChatGPT: a page checklist

Google’s name for its freshness systems: “query deserves freshness” systems. Spanish: actualidad del contenido; Google’s Spanish documentation says “sistemas de actualización”.

Google’s ranking systems guide says: “We have various ‘query deserves freshness’ systems designed to show fresher content for queries where it would be expected.” Its example: someone searching for a movie that has just been released probably wants recent reviews rather than articles from when production began. Freshness also shapes AI answers. Google describes retrieval-augmented generation, which it also calls grounding, as a technique to improve “the quality, accuracy, and freshness of AI responses” by retrieving “up-to-date web pages” from its Search index.

Related

GroundingTraining data vs live retrievalCore update vs spam updateNon-commodity content

Google calls a spam update that deals specifically with link spam a link spam update. Spanish: actualización principal frente a actualización de spam.

Core updates are “significant, broad changes” to Google’s search algorithms and systems that “don’t target specific sites or individual web pages”, so a drop during one is not a sanction aimed at that site; Google points owners to its content self-assessment. Spam updates improve automated systems that run all the time. A site that fixes violations can improve as those systems learn “over a period of months” that it complies, but after a link spam update any ranking benefit from spammy links is lost and “cannot be regained”. Google announces both on its Search Status Dashboard. As of September 29, 2026, it listed 2026 core updates starting March 27 and May 21, and spam updates starting March 24, June 24, August 18 and September 24; the September update, which Google said applies globally and to all languages, was still rolling out.

Related

Manual actionLink spamScaled content abuseE-E-A-T

Informally called a Google penalty. Spanish: acción manual.

Manual actions are the exception. Google says its algorithms are “extremely good at detecting spam” and in most cases find and remove it automatically; people step in to protect the quality of its index. When a reviewer issues one, Google notifies the owner in Search Console’s Manual actions report and message center, and the report names the issue, such as “Unnatural links to your site” or “Cloaking and/or sneaky redirects”. Once every listed issue is fixed on every page, the owner can select Request Review; Google says most reconsideration reviews “can take several days or weeks”, and link-related ones may take longer. If the report is empty, a ranking drop has another cause, such as automated spam systems or a core update. Google’s documentation now says the impact of site reputation manual actions doesn’t apply to results shown to users in the European Economic Area.

Related

Core update vs spam updateSite reputation policyLink spamCloaking

Spanish: abuso de contenido a gran escala.

Google’s spam policies list examples: using generative AI or similar tools to generate many pages without adding value for users; scraping feeds, search results or other content, including through automated transformations such as synonymizing or translating, where little value is provided; stitching together content from different pages without adding value; and creating multiple sites to hide the scaled nature of the content. Google’s guide to AI search extends the policy to AI answers: separate content for every variation of a query, fan-out queries included, made primarily to manipulate rankings or generative AI responses violates it. A related policy, doorway abuse, covers sites or pages created to rank for specific, similar queries. In our engine test (September 29, 2026), all ten answers we read, in English and Spanish, described the policy correctly.

Related

Query fan-outNon-commodity contentSite reputation policyCore update vs spam update

Site reputation abuse was Google’s name for the policy when it was announced in 2024, and the name the European Commission used in 2025. Spanish: política de reputación de sitios (antes, abuso de reputación del sitio).

Third-party content comes from anyone separate from the host site, such as its users, freelancers or white-label services; the policy applies only when it is published mainly to benefit from the host’s ranking signals. In an update published on August 28, 2026, Google changed enforcement in the European Economic Area (EEA). In results shown to EEA users, affected pages may be categorized as separate from the main domain, so each part of the site ranks on its own merits, but manual actions no longer apply; Google said it would lift previous manual actions under the policy for those results, and it offers EEA sites a new reconsideration process and alternative dispute resolution. Outside the EEA, manual actions still apply. The European Commission opened a Digital Markets Act investigation into the policy on November 13, 2025.

Related

Manual actionScaled content abuseThird-party sources in AI answersLink spam

Google also speaks of link schemes; its matching manual action is “Unnatural links to your site”. Spanish: spam de enlaces.

Google’s examples include buying or selling links for ranking purposes (exchanging money, goods or services for links, or sending someone a product in exchange for a post with a link), excessive link exchanges, links created with automated programs, text ads that don’t block ranking credit, and links with optimized anchor text in articles, guest posts or press releases distributed on other sites. Paid links as such are allowed: Google calls buying and selling links “a normal part of the economy of the web” for advertising and sponsorship, provided they are marked rel="sponsored" or rel="nofollow". Google fights link spam with automated systems, including link spam updates, and with manual actions for “a pattern of unnatural, artificial, deceptive, or manipulative links”. Since May 15, 2026, its documentation states that the spam policies also apply to generative AI responses in Google Search.

Related

BacklinksManual actionCore update vs spam updateBrand mentions

Google’s matching manual action is “Cloaking and/or sneaky redirects”. Spanish: encubrimiento.

Google’s spam policies give the example of inserting text or keywords into a page only when the user agent requesting it is a search engine, not a human visitor. Since May 15, 2026, Google’s documentation says the spam policies also apply to generative AI responses in Google Search. The same technique can target other AI systems, whose crawlers and fetchers identify themselves by user agent, such as OpenAI’s OAI-SearchBot and ChatGPT-User. In controlled experiments published in October 2025 by SPLX, which sells AI security software, a test site served AI crawlers an altered page, and AI tools including OpenAI’s Atlas browser repeated the altered version.

Evidence

Google’s definition is official; the AI-crawler test is a vendor experiment.

Related

AI crawlers and user agentsAI answer manipulationManual action

Includes indirect prompt injection. Spanish: manipulación de respuestas de IA (inyección de prompts, texto oculto).

Google’s spam policies define spam as techniques used to deceive users or manipulate its systems into featuring content prominently, including “attempting to manipulate generative AI responses in Google Search”. Its hidden text and link abuse policy covers content placed “solely to manipulate search engines and not to be easily viewable by human visitors”, such as white text on a white background. Google’s policies do not use the term prompt injection; researchers do. Greshake et al. (2023) showed how to compromise applications built on language models, including Bing’s GPT-4 powered chat, “by strategically injecting prompts into data likely to be retrieved”. Kumar and Lakkaraju (2024; an arXiv preprint, not peer-reviewed as far as we know) showed, with a catalog of fictitious coffee machines, that adding a “strategic text sequence” to a product’s information page “can significantly increase its likelihood of being listed as the LLM’s top recommendation”.

Evidence

The policy is Google’s; the attack research includes a peer-reviewed paper and an arXiv preprint.

Related

CloakingRetrieval-augmented generation (RAG)Large language model (LLM)Scaled content abuse

Also discussed as citation accuracy. Spanish: atribución errónea.

The Tow Center for Digital Journalism measured it in a study published in Columbia Journalism Review on March 6, 2025. Researchers took excerpts from 10 articles by each of 20 news publishers and asked eight AI search tools, including ChatGPT search, Gemini and Perplexity, to identify each article’s headline, original publisher, publication date and URL: 1,600 queries in all. Collectively, the tools gave incorrect answers to more than 60% of queries. More than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs that led to error pages. On some occasions, the chatbots linked to syndicated copies on other platforms instead of the original, often even when the publisher was known to have a licensing deal with the AI company. The tool versions tested date from early 2025, so current results may differ.

Evidence

Study by the Tow Center for Digital Journalism at Columbia’s Graduate School of Journalism, an academic research center, not a vendor; it was not peer-reviewed.

Related

CitationHallucinationCitation rateGrounding