More and more people ask AI instead of googling. If AI doesn't name you when someone asks about what you do, for that customer you don't exist. Here's how you get cited —it's called GEO— plus the dictionary of the 40 concepts you need to know.
Appearing in ChatGPT is called GEO (Generative Engine Optimization): working your content, your data and your authority so AI mentions and cites you when someone asks about what you do. SEO got you the click; GEO gets you the citation. It is not luck and it is not paid: they are concrete signals you can work on and, above all, measure.
There is no magic button. AI cites whoever makes it easy. These are the five levers that move the needle most, in order. Each one links to the dictionary concept that explains it in depth.
Make ChatGPT know who you are, what you do and why you can be trusted. A consistent name across your site, schema, and a real author page.
How it works →AI cites the concrete. Adding statistics, direct quotations and source references raised visibility by up to 30-40% in relative terms in the only academic study that has measured it (Aggarwal et al., KDD 2024 · arXiv:2311.09735).
How it works →Get talked about elsewhere: Reddit, YouTube, reviews, press. If you only exist on your own site, to the AI you're a single voice, and an interested party. When several independent sources agree, it stops being your claim and becomes a fact.
How it works →The engines reward what is recent. Updating and publishing every 7-14 days tells AI you are still alive and relevant.
How it works →What isn't measured doesn't improve. Track your Share of Answer, spot your citation gaps and act. Then check the citation actually landed. That loop is everything.
How it works →From the basics to the advanced, in four levels. Each definition is written to be understood at first read —and so AI itself can cite it.
Instead of handing you links, this AI writes the answer itself, and that means it decides who shows up and who doesn't.
Generative AI is a type of artificial intelligence that replies with written text rather than a list of links. In practice, it changes search: the answer names only a handful of brands, so appearing inside it, not just on Google, becomes the goal for any business.
An LLM doesn't know things: it predicts the next word from patterns. That's why it holds a conversation so well, and why it sometimes makes things up.
An LLM (Large Language Model) is an AI model trained on huge amounts of text that produces its answers by predicting the most likely next word. It doesn't "know" in the human sense: it recognises patterns in language. This is the technology behind ChatGPT, Gemini and Claude. Once you grasp that it works by probability, two things make sense: why it converses so naturally, and why it sometimes "hallucinates", filling gaps with plausible but false information when it doesn't have the fact.
There is no single "AI": there are four different engines that cite different sources. Winning on one doesn't win you the others.
There are four main AI search engines, and they work in different ways. ChatGPT (OpenAI) is the most widely used and, when it searches the web, leans on the Bing index; it has a training cutoff, so it doesn't know the very latest content. Gemini (Google) is built into Google's own search. Perplexity is an AI search engine that cites its sources in real time, which makes it the fastest route to being cited (sometimes within a day). Copilot (Microsoft) runs on Bing. Each engine cites different sources: ChatGPT and Perplexity overlap on only around 11% of the domains they cite. The practical takeaway is that visibility on one engine isn't enough; you have to work each one separately.
Google now answers you with AI before the usual links. And its answer barely overlaps with ChatGPT's, so you work it separately.
AI Overviews is the AI-generated answer box Google shows at the top of its results, above the traditional links. AI Mode is Google's conversational mode, where the user asks and follows up as if in a chat. Both take clicks away from the usual organic results, because many users stay with the answer and never scroll down to the links. On top of that, their overlap with what ChatGPT cites is low (around 13.7% between AI Overviews and AI Mode), so being in one engine doesn't guarantee being in the other: each surface is optimised on its own.
GEO is optimising so AI mentions and cites you in its answers. SEO gets you the click; GEO gets you the citation.
GEO (Generative Engine Optimization) is the discipline of optimising a brand's or professional's presence so they appear and get cited in the answers of AI search engines like ChatGPT, Perplexity, Gemini or Google AI Mode. It boils down to one idea: SEO aims to get people to click your link; GEO aims to get the AI to cite you inside its answer. Since each generated answer has room for only a few brands, being present in it becomes the goal. The first empirical study of GEO in Spain (76 companies, more than 20 sectors and over 2,280 prompts) analysed why AI cites some sources and not others.
SEO is getting the click on your link; GEO is getting the AI to cite you. They don't compete: they complement each other. And you don't need to be number 1 on Google.
SEO (Search Engine Optimization) aims to rank your links on Google so users click and visit your site. GEO (Generative Engine Optimization) aims to get AI search engines to mention or cite you inside their answer. They aren't mutually exclusive: they complement each other, and much of the content work serves both. One key difference: to get cited by AI you don't need to be the first result on Google. Retrieval systems (RAG) pick sources from mid-table positions, often between 4 and 20, based on how well the passage answers the question, not just on ranking. That opens the door to brands that don't yet lead traditional search.
AEO is optimising for answer engines (snippets, voice, direct answers). GEO is its evolution for generative AI. They overlap so much they're almost synonyms.
AEO (Answer Engine Optimization) means optimising content to appear in direct answers: featured snippets, voice assistants and other answers that skip the list of links. GEO is the evolution of that same idea applied to generative AI and large language models. The two overlap heavily and are often used almost interchangeably; the nuance is in their origin: AEO grew out of direct answers and snippets, GEO out of LLMs. More than the acronym, what matters is the shared goal: being the answer the engine gives.
The AI answers and the user visits no website. There are fewer clicks to go around, so appearing in the answer is worth more, and that's what you have to measure, not just visits.
Zero-click search happens when the user gets the answer straight from the search engine or the AI chat and visits no website. The consequence is twofold: there are fewer clicks to spread across pages, and at the same time appearing inside the answer, being mentioned or cited, gains value, because that's where the user actually sees the information. In practice, this forces a change in how you measure: tracking visits alone no longer cuts it; you have to measure your visibility inside the AI's own answers.
Nobody searches "lawyer Madrid" anymore: people ask in long, natural phrases. You optimise for your customer's real questions, not for isolated words.
A prompt is the question a user puts to an AI search engine, written in natural language, in long, conversational phrases. Traditional search optimised for keywords (short terms like "plumber Seville"); AI search optimises for the real questions customers ask ("I need a trustworthy plumber in Seville who can come today"). That's why people say the prompt is the new keyword. Measuring AI visibility means identifying those questions: in our system we measure around 90 prompts per client, the queries their ideal customer would genuinely ask, and check in how many the brand appears.
A mention is the AI naming you; a citation is the AI naming you and linking your URL. The citation is harder and more valuable, and they're two different metrics.
A mention is when the AI names a brand in the text of its answer. A citation goes one step further: on top of naming it, it links to it as a source with its URL, so the user can click through and visit. A linked citation is harder to earn, the AI has to trust the page enough to send traffic its way, and that's what makes it more valuable. They're two different indicators: in the dashboard they're measured separately as mentions (Share of Answer) and linked citation (Citation Rate), and each one is calculated and improved in Level 3 of the dictionary.
The date up to which a model "studied" the internet; anything after that it doesn't know from memory.
The training cutoff is the date up to which a language model was trained on internet data. Content published after that isn't part of its base knowledge, so new websites take a while to show up in engines like ChatGPT (3-6 months) unless the engine searches in real time, as Perplexity does.
Before answering, the model pulls up current documents and builds its reply on them.
RAG (Retrieval-Augmented Generation) is the technique by which an AI model, before answering, retrieves current documents from the internet and generates its response based on them. It lets the model get past its training cutoff and cite recent sources. Engines like Perplexity and Google AI Mode use it, which is why they can mention a page just days after it's published.
Anchoring the answer in real sources instead of just the model's memory.
Grounding is the practice of anchoring an AI engine's answer in real, verifiable sources rather than generating it from the model's memory alone. It cuts down on hallucinations (made-up facts) and makes the answer traceable. The specific pages an answer leans on are called grounding sources, and knowing which ones they are tells you where to publish or earn mentions so the AI cites you.
The specific domains and URLs the AI uses as sources when it answers.
Cited sources (Grounding Source Prominence, GSP) are the specific domains and URLs an AI engine uses as sources when answering a query. Analyzing them shows which third-party sites shape the answers in a given sector, so you can pinpoint exactly where to publish content or earn mentions to raise your odds of being cited.
Something identifiable, a person or a company, with attributes, not loose text.
An entity is a uniquely identifiable object, a person, company, product or place, with attributes and relationships attached, as opposed to a plain string of text. When an AI engine recognizes a brand or professional as an entity, it can recommend them with confidence. Pages with 15 or more recognized entities are 4.8x more likely to be selected. You consolidate an entity by linking its identity with `sameAs` to verified profiles (LinkedIn, Wikidata, etc.).
Experience, Expertise, Authoritativeness and Trust: why the AI trusts a source.
E-E-A-T (Experience, Expertise, Authoritativeness and Trust) is the framework AI engines use to judge how trustworthy a source is. You prove it with an author who has verifiable credentials, your own data and studies, and consistency between what a site claims and what third parties say about it. 96% of citations in Google's AI Overviews come from sources with strong E-E-A-T, which makes it a core factor in citability.
Structured data that tells the AI what each thing on your site actually is.
Schema (structured data, usually in JSON-LD format) is code you add to a page to spell out exactly what each element represents: an Organization, a Person, a Service or a FAQPage. It helps AI engines read the content without ambiguity. The FAQPage case is the clearest: it hands the model the exact unit it already needs, a question and its answer, already separated, instead of forcing it to carve one out of running prose. It's worth being precise about what is known and what isn't: no public study quantifies how much marking up a FAQPage increases citation. The '3x' that circulates in the industry has no published backing, and the people repeating it rarely produce the source. The mechanism (an extractable format, entities linked through `@graph` to give the site semantic coherence) stands on its own; the multiplier doesn't.
A file at the root of your domain with a curated index of your content for the AI.
llms.txt is a text file you place at the root of a domain (for example, `domain.com/llms.txt`) to give AI models a curated index of the site's most relevant content and how to interpret it. It follows the same logic as `robots.txt`, but it's meant to guide rather than block. It's an emerging standard, still only partly adopted, simple to create and free, which makes it a low-risk improvement.
The AI favors what's recent; anything you don't update slowly loses citations.
Content freshness is AI engines' preference for recent information. The mechanism is direct: engines with live search, Perplexity and Google AI Mode, resolve the query on the spot and pick from what they find, so a two-year-old page competes against one updated last week. A site that gets published and then abandoned steadily loses ground to competitors who do maintain their content. A 7 to 14 day review cycle is a recommended practice, not a measured figure. The modification date (`dateModified`) has to match real content changes for the signal to be legitimate.
If you only exist on your own website, to the AI you're a single voice, and an interested party.
Consensus signals are the mentions and confirmations of a brand on third-party platforms, Reddit, YouTube, G2, Trustpilot, Wikipedia, forums, outside its own domain. AI engines use them to cross-check what a brand says about itself against what everyone else says. The logic is the same as any person about to hire someone: they don't just take the company's website at face value, they look at what's said elsewhere. If a brand only exists on its own domain, to the model it's a single voice, and an interested party. When several independent sources agree, the claim stops belonging to the brand and becomes a checked fact about the world, and that is what an engine can cite without exposing itself. Getting cited therefore requires existing verifiably beyond your own site. No public study quantifies how much citation this earns you; the mechanism, on the other hand, is direct and observable.
The share of AI answers your brand shows up in. Your real visibility.
Share of Answer (SOA) is the share of AI answers —across a fixed set of measured prompts— that mention a brand. It measures real visibility in generative engines: how often the AI includes you when answering a potential customer. Example: showing up in 9 out of 90 prompts is a 10% SOA.
A 0-to-100 index that measures not just whether the AI mentions you, but where and how strongly.
LLM Visibility Index (LVI) is a 0-to-100 index that measures how well a brand is recognized in AI engines, based on where it gets mentioned (100 = first mention). Unlike Share of Answer, which only counts whether you appear, LVI weighs your position and how strongly the AI names you. A high LVI means the model treats you as a relevant source; a low LVI with a high SOA means you show up, but always last.
Your share of mentions against the whole niche. How much of the AI's conversation is you versus your competitors.
Share of Mentions is a brand's share of mentions out of the whole niche (the brand plus its competitors) inside AI answers. It measures how much of the conversation each player owns: share of voice applied to generative engines. Unlike Share of Answer, it's a relative metric: raising your visibility doesn't guarantee raising your share if a competitor grows faster.
The share of prompts where the AI actively recommends you, not just mentions you. The jump from showing up to being the pick.
Share of Recommendation is the share of prompts where the AI actively recommends a brand —using explicit recommendation language— rather than just mentioning it. It's the final step of visibility: going from showing up to being the option the model puts first. A wide gap between mentions and recommendations flags the so-called Library Effect: the AI informs people about you, but doesn't put you forward.
The share of answers that include your URL as a cited source. The hardest KPI to move, and the most valuable.
Citation Rate is the share of AI answers that include a domain's URL as a cited source. Don't confuse it with a mention, which just names the brand without linking to it. It's the hardest KPI to move, because models cite sources very selectively. Perplexity, searching in real time, is the fastest route to earning a citation with a URL; ChatGPT takes months of building authority.
The domains the AI cites where your competitors appear and you don't. Your direct action list. Money on the table.
Citation Gap is the set of domains the AI cites as a source where a brand's competitors appear but the brand itself doesn't. It works as a direct action list: it points to exactly where to create content or earn a presence to win those citations. In the dashboard it shows up as the Opportunities view, with the domains, the URLs, and the prompts that trigger them.
Your presence on the platforms where the AI cross-checks opinions: the places where others talk about you, not where you talk about yourself.
Consensus Score measures a brand's presence on the platforms where the AI cross-checks third-party opinions, Reddit, YouTube, G2, Trustpilot, Wikipedia and the like. The model treats these as neutral sources, unlike a brand's own website, where everybody claims to be the best. The engine reasons the way anyone picking a restaurant does: it doesn't believe the sign on the door, it reads the reviews. This deserves to be said plainly: no public study quantifies the effect of being present on those platforms, and the '3x' doing the rounds in the industry has no backing. What is verifiable, client by client, is which sources the AI actually cites in a given sector, and which of them the brand is missing from.
How the AI describes you when it mentions you —positive, neutral or negative— and in what words. Showing up isn't enough: what it says about you matters.
Brand Sentiment in AI analyzes how the AI describes a brand when it mentions it: whether the tone is positive, neutral or negative, and in exactly what words. Showing up in the answers isn't enough; what matters is whether the model presents you as a "leader" or as "expensive". It helps you spot reputational risks before they do damage and fix the narrative at its source. It's a metric available with our own measurement engine.
Your site's technical grade for getting cited by AI: robots, llms.txt, schema, content and authority. Is your house ready to receive the AI?
AI Readiness is the technical grade that assesses how prepared a website is to be cited by AI, across five factors: robots.txt, llms.txt, schema (structured data), content quality and authority. If the site isn't technically ready, no content strategy performs at its best: it's the foundation the rest of the generative-engine visibility metrics rest on.
The estimate of leads and revenue your AI visibility can generate — starting from your real conversion rate, not from an invented multiplier.
ROI of AI Visibility estimates the leads and revenue a brand presence in generative engines can generate. A widespread myth needs dismantling first: the claim that ChatGPT traffic converts about nine times better than organic comes from a single-site case study and does not hold up. The published evidence points to AI traffic converting roughly the same as organic search: the most rigorous study, peer-reviewed in Marketing Science across 973 e-commerce sites ([Kaiser & Schulze, 2026](https://pubsonline.informs.org/doi/10.1287/mksc.2025.0489)), finds it converts 13% worse, while other analyses put it up to 31% higher. The real investment case is not the conversion rate but the growth of the channel: generative search is a tiny fraction of traffic today and is growing at triple-digit rates, so visibility gets built before the channel matures. An honest estimate starts from the site actual conversion rate and gives a range, not a single figure.
When AI mentions you often as information but almost never recommends you as an option.
The Library Effect describes a brand that AI mentions often as information but rarely recommends as a buying option. You measure it as the gap between visibility (mentions) and recommendation; a wide gap (above ~20 points) signals authority without conversion, and you fix it with decision-oriented content: case studies, comparisons, and recommendation language.
Measure your AI visibility without kidding yourself: drop the brand prompts, show the sample size, and report a range, not a single number.
Statistical honesty in AI visibility measurement comes down to three corrections: exclude brand prompts (questions about your own name, which artificially inflate presence), always show the sample size (n), and report a confidence interval —the 95% Wilson confidence interval— instead of a bare percentage. A real example: joseredondo.es scored an honest Share of Answer of 2.1% (interval from 1% to 4.6%) across 281 neutral prompts, below the naive figure you get without cleaning the data. The principle: an uncomfortable but true number beats a high, misleading one.
AI doesn't answer the same way twice, so I run each question 2-3 times and average it, turning an estimate into a metric I can defend.
Multi-run is the practice of firing each measurement prompt two or three times and averaging the results, rather than measuring once. Language models bake in variability by design: the same question can produce different answers, so a single measurement captures one moment, not real behaviour. Averaging several passes cuts that noise and turns an estimated percentage into a defensible metric, at the cost of slightly more measurement spend.
Winning on one engine doesn't mean winning on all of them; the sources each AI cites barely coincide, so you have to optimise engine by engine.
Overlap between engines measures how much the sources one AI engine cites coincide with those another one cites. It's low: only 11% of the domains ChatGPT cites also show up in Perplexity, and the overlap between Google's AI Overviews and AI Mode is 13.7%. The strategic implication is that each engine is an independent surface with its own citation logic: a win on one doesn't automatically transfer to the rest, so optimisation has to be done engine by engine.
Aggarwal et al. (KDD 2024, arXiv:2311.09735) measured what makes AI cite you more: hard numbers, authoritative sources and expert quotes, up to +30-40% relative.
Statistics, Cite and Quote are the three writing methods that most increase citation in AI engines according to the only academic study that has measured it: Aggarwal et al., 'GEO: Generative Engine Optimization', KDD 2024 ([arXiv:2311.09735](https://arxiv.org/abs/2311.09735)). Adding statistics with concrete numbers, citations of authoritative sources and direct expert quotes raised visibility by up to 30-40% in relative terms on their benchmark. Keyword stuffing, by contrast, did not improve visibility. It's the only GEO effectiveness figure I use, and it always travels with its source: the other multipliers doing the rounds in the industry (the schema '3x', the consensus one, the freshness one) have no study behind them.
There are ways of writing AI extracts far better: definition, question→answer, table, process, standout figure.
Citable structures are content formats AI engines extract more easily. The five main ones: the definition block (25-50 words starting with the term), the question→answer pair (the question as the heading, the direct answer right below), the comparison table with a 'best for' row, the numbered process and the standout figure. The mechanism is simple: a generative engine cites self-contained fragments, the ones it can copy as-is without dragging context along. An answer buried in a long paragraph forces the model to reconstruct it, and often it just won't. The related effect is measured: adding statistics, direct quotations and source references raised visibility by up to 30-40% in relative terms in Aggarwal et al. (KDD 2024, [arXiv:2311.09735](https://arxiv.org/abs/2311.09735)). The effect of formatting on its own has not been publicly quantified.
AI doesn't search your question as-is: it splits it into several sub-questions, searches each one separately, and assembles the answer; that's why you have to cover the spin-offs too.
Query fan-out is the mechanism by which an AI engine breaks a query into several sub-questions, searches each one separately, and synthesises the final answer by combining the results. That's why answering the main question isn't enough: you have to cover its spin-offs too (semantic completeness). The recommended tactic is self-contained content blocks of 134 to 167 words, each with a complete answer to a specific sub-question, extractable without depending on the rest of the page.
When several of your URLs compete for the same citation, they split the strength instead of adding it up, and AI can't tell which one to prioritise.
Content cannibalisation happens when several URLs on the same site compete for the same citation in AI and split the strength instead of adding it up, leaving the engine without a clear signal of which one to prioritise. A real example: on joseredondo.es, the query "what do you know about joseredondo.es?" spread the citation across six different in-house URLs. The fix is to consolidate the content into a single strong reference page and point the others to it, rather than keeping several weak pages competing against each other.
AI visibility by city or area ("best X in [city]"), where local competitors usually rule and you measure against your real service area.
Local GEO is a brand's visibility in AI engines for queries with a geographic component ("best X in [city]"). On this turf local competitors usually dominate over the big brands, which opens an opportunity for small businesses with a defined service radius. You measure it against the client's real service area —not generic locations like Madrid or Barcelona—: you expand the prompts across cities, measure each one, and spot where the brand is weak so you can act with priority.
The full system —measure where you don't show up, act, prove the citation landed, and translate it into euros— that turns a data dashboard into a results machine.
The closed loop or Autopilot is the full AI visibility system in four steps: (1) MEASURE where the brand doesn't show up and on which prompts it loses; (2) ACT by creating the content or earning the mention that's missing; (3) PROVE the citation landed, by querying the AI again to verify that it now cites the brand; and (4) translate it into euros via Google Analytics, measuring the real traffic arriving from AI. That's what turns a data dashboard into a results machine: the citation-verification step is what separates the system from a simple "create content" recommendation.
The questions I get most about appearing in ChatGPT and the other AI search engines.
Appearing in ChatGPT isn't luck and isn't paid. You work at it: make AI know who you are (a clear entity, with schema and a solid author page), give it data it can cite (statistics and direct definitions), get others to talk about you (Reddit, YouTube, reviews) and keep your content fresh. That's GEO. And above all: measure it, because what isn't measured doesn't improve.
GEO stands for Generative Engine Optimization. It is the discipline of working your content, your data and your authority so that ChatGPT, Perplexity, Gemini or Google AI Mode mention and cite you when someone asks about what you do. SEO got you the click; GEO gets you the citation.
No, though they lean on each other. SEO fights for a spot in a list of blue links. GEO fights to be inside the answer the AI gives, which is where the game is decided now. An answer only holds two or three names; if yours isn't there, for that customer you don't exist. Good SEO helps GEO, but it isn't enough.
It depends on the engine. Perplexity searches in real time and can cite you within days if your page is well built. ChatGPT, on the other hand, carries its training cut-off and usually takes months to include you reliably. That is why the strategy mixes quick wins (Perplexity, Google AI Mode) with the groundwork that eventually lands in ChatGPT.
Yes, and you should. The base metric is Share of Answer: the percentage of AI answers —across a fixed set of your ideal customer's questions— in which you appear. If you show up in 9 of 90, your SOA is 10%. You see it week by week. You can start free with the audit: it tells you whether AI already mentions you and who shows up in your place.
Because of concrete signals: being a recognisable entity, offering citeable data, having external consensus (people mention you on forums, videos, reviews) and keeping your content fresh. I'm not making it up: I got it from Spain's first empirical GEO study —76 companies, more than 20 sectors and over 2,280 prompts— analysing why AI cites some sources and not others.
Theory is fine, but what counts is your case. Analyse your site for free and I'll tell you whether ChatGPT, Perplexity or Google AI Mode already cite you —and who shows up in your place.