For twenty years, the question in digital marketing was “how do we rank on Google”. The question now is different: when AI answers for your market, who does it cite? This guide brings together what the latest international studies show about how generative engines choose their sources, why the correlation between ranking and being cited collapsed in 2026, and the five-layer method, with a complete checklist, that we use at Flowup to turn brands into the answer. No magic promises: mechanics, data and work.
Direct answer
GEO (Generative Engine Optimization) is the discipline of preparing a brand to be understood, retrieved and cited by generative systems such as ChatGPT, Gemini, Perplexity and Google's AI Overviews. Getting cited depends on five layers this guide details with a checklist: a technical foundation (crawlability, structured data, performance), citable content (direct answers, statistics, sources, information gain), a consolidated entity (brand and author recognisable in the knowledge graph), external validation (mentions and cross-citations in the sources AIs rely on) and continuous measurement. The number that changes the game: the overlap between AI citations and the organic top 10 fell from 76% to 38% in seven months (Ahrefs, 2026). Ranking helps, but it is not enough. It's not enough to rank: be the answer.
What GEO is (and how it relates to SEO and AEO)
GEO stands for Generative Engine Optimization: the set of practices that make a brand understood, retrieved and cited by generative engines. The term left the jargon phase and gained academic grounding with the Princeton University study presented at KDD 2024, which built a 10,000-query benchmark, tested nine optimisation tactics and measured visibility gains of up to 40% in generative answers for content with statistics, cited sources and clear language.
GEO does not replace SEO or AEO: the three disciplines form a system. SEO keeps content retrievable and competitive in the indexes that feed the systems; AEO structures information for the direct-answer format; GEO maximises the chance of the brand being the source used and cited in the synthesis. This guide integrates all three, because that is how the systems actually work. If you are new to the shift, our analysis of the end of the click maps the behavioural change behind it.
How AIs choose whom to cite: the mechanics
To optimise with judgement, you need to understand the process. Broadly, web-connected generative systems operate in three stages. First, retrieval: given the question, the system searches indexes and connected sources for candidate documents. Second, synthesis with grounding: the model generates the answer anchoring claims in the retrieved documents, and citations point to the sources that supported each passage. Third, and increasingly decisive, query fan-out: the original question is decomposed into sub-queries, each with its own retrieval, and the final answer aggregates everything.
The practical rules of the game follow from that mechanics. If your content is not crawlable and indexable, it never enters the retrieved set: the game ends before it starts. If it is retrieved but has no clear, verifiable, extractable passages, it loses the citation slot to content that does. If it repeats what the corpus already knows, it falls into the redundancy bucket the systems discard, a documented mechanism ranging from training-data deduplication to Google's information gain patent. And if the brand is not a recognisable entity, with consistent attributes across sources, the system has less statistical confidence to attribute anything to it.
A citation is not a literary quality award. It is a statistical trust decision: the system anchors the answer in the source offering the best balance of relevance, verifiability and consistency.
The numbers defining the game in 2026
No serious GEO guide can be written without the data, and the data tells a clear story in four acts.
Act 1: the AI answer became the default search experience. AI Overview presence jumped from roughly 30% to 48% of tracked searches in a year per BrightEdge (58% growth), and when they appear the summaries exceed 1,200 pixels in height, pushing organic results below the fold. Prompts got longer too: around 23 words on average, against 3 or 4 in classic search. Regional markets follow: in Brazil, ChatGPT alone concentrated 99% of generative AI traffic in 2025 (Cadastra/Similarweb).
Act 2: the click shrank. The Pew Research Center measured users clicking traditional results in 8% of searches with an AI Overview versus 15% without, and clicking the cited sources inside the summary only 1% of the time. Seer Interactive recorded organic CTR falling from 1.76% to 0.61% on queries with an AI summary. The strategic detail: the same Seer research measured that cited brands saw relatively better click performance (an effect in the order of 35%) than absent ones. Being in the answer became the new being on page one.
Act 3: ranking stopped guaranteeing citation. This is the data point that reorganises strategy. In July 2025, Ahrefs measured 76% of AI Overview citations coming from top-10 pages. In the February 2026 update (863,000 keywords, 4 million cited URLs), the number fell to 38%, with the remainder split almost evenly between positions 11 to 100 (31.2%) and beyond the top 100 (31%). BrightEdge, with a different methodology, measures about 17% overlap with the top 10, meaning roughly 5 out of 6 citations come from outside page one. seoClarity adds nuance: 94% of AI Overviews cite at least one URL from the top 20, position 1 gets cited in 43% of cases versus 7% for position 20, and still 44% of citations come from beyond the top 20. On ChatGPT the independence is even greater: market analyses point to only 6.8% overlap with Google's top 10, with 28% of its most-cited pages having no organic visibility at all. Methodologies differ and exact figures vary, but the direction is one: position helps, and it is not enough.
Act 4: the traffic that remains is worth more. Analyses based on Semrush data point to AI-sourced traffic converting about 4.4 times better than traditional search, and Ahrefs reported, on its own site, AI visitors generating 12.1% of signups from just 0.5% of traffic. Fewer clicks, more intent. It is the same dynamic behind why generic volume content became a liability.
With the mechanics and the numbers on the table, the method. The five layers below are the spine of GEO work at Flowup, and the complete checklist operationalises each one.
Layer 1: technical foundation
Nothing happens if machines cannot access, understand and trust the infrastructure. The foundation has four fronts. Access: robots.txt and firewall allowing the crawlers of the systems where you want presence (GPTBot, ClaudeBot, PerplexityBot, Google-Extended), confirmed in server logs. Indexing and health: strategic pages indexed, clean sitemap, correct canonicals, performance within Core Web Vitals. Graph-based structured data: JSON-LD connecting organisation, site, pages, authors and services with stable identifiers reused across the ecosystem, always coherent with the visible content; our JSON-LD in WordPress guide covers the implementation in depth. Agent readability: an official AI knowledge base with the brand's canonical facts in a clean format, complemented or not by llms.txt, whose isolated effectiveness is limited, as we showed in our llms.txt analysis.
Layer 2: citable content
Citable content is what an AI can extract and anchor with confidence. The attributes validated by the Princeton study and by practice: a direct answer at the top (the central question answered in the first paragraphs, in extractable language); statistics with named, dated sources; a clean structure of headings, lists and tables that parsers navigate effortlessly; real FAQs, mirrored in structured data; and, above all, information gain: primary data, signed positions and syntheses that exist nowhere else. Volume without novelty is a liability. And fan-out changes the editorial design: instead of one heroic page per keyword, a territory covered in depth, with pillar pages and connected satellites.
Layer 3: entity and authority (E-E-A-T)
Generative systems attribute information to entities, not loose URLs. Consolidating the entity means: a brand with identical name, description, attributes and relationships across every source (site, profiles, directories, press); real, qualified authorship, with complete author pages, verifiable credentials and consistent external presence; and the explicit association between the brand and the topics it wants to own, built through co-occurrence, recurring content and Person and Organization structured data. It is the practical translation of E-E-A-T for the generative era: experience and expertise must be machine-readable. This guide, signed, with an identified author and a public methodology, is an applied example of the standard.
Layer 4: external validation
Citation studies are unanimous: AIs lean heavily on distributed validation. Semrush and Peec AI place encyclopaedias, communities (Reddit first), YouTube and LinkedIn among the most-cited domains, and Evertune's analysis of 200 million answers shows no single domain exceeding 5% of citations: authority is distributed and earned topic by topic. There is also the brand effect: brands in the top quartile of web mentions receive an order of magnitude more citations. The operational consequence: correct presence on the platforms systems use as sources, and continuous data-driven PR generating mentions and cross-citations in authoritative outlets. The entity consolidates outside your own site.
Layer 5: measurement and iteration
What is not measured becomes opinion. GEO measurement combines: direct monitoring of strategic prompts across the main systems, recording who gets cited and with what framing; Google Search Console's AI features filter, which aggregates performance in AI experiences; traffic and lead analysis by origin, including customer-declared origin; and volatility reading, because citation patterns move fast, as the Semrush study of 230,000 prompts documented. In the Dra. Ana Vega project, publicly documented by Flowup, it was this measurement that recorded 71.3% of June 2026 leads coming from the organic base and the first lead attributed to ChatGPT, the kind of early signal that guides the next iteration.
The complete citability checklist
The checklist below operationalises the five layers. Use it as an audit: every unmet item is a hypothesis for why AI does not cite your brand yet.
Block 1: technical foundation
- robots.txt allowing GPTBot, ClaudeBot, PerplexityBot and Google-Extended (or a documented, deliberate blocking decision)
- Firewall and CDN not silently blocking AI crawlers, confirmed in server logs
- Strategic pages indexed with no accidental noindex; sitemap.xml up to date
- Correct canonicals and internal duplicate content resolved
- Healthy Core Web Vitals on answer pages
- Graph JSON-LD: Organization, WebSite, WebPage, author (Person) and services with stable, reused @id values
- Structured data 100% coherent with visible content, validated in the Rich Results Test
- Official AI knowledge base published, with the brand's canonical facts in a clean format
Block 2: citable content
- Extractable direct answer in the first paragraphs of every strategic piece
- Statistics with named, dated sources for every central argument
- At least one proprietary primary data point per territory (research, operational number, documented case)
- Clean heading hierarchy (one H1, descriptive H2/H3) and tables or lists where information is comparative
- Real FAQ per strategic page, mirrored in FAQPage schema
- Territory covered in depth: pillar page and satellites connected by descriptive internal links
- Information gain test: what does this page add to what AI already answers today?
- Visible, truthful publication and update dates; scheduled review of core content
Block 3: entity and authority
- Brand name, description and attributes identical across site, profiles and relevant directories
- Complete institutional page with verifiable facts (founding, registration, address, specialisations)
- Real authors with author pages, credentials and Person schema connected to the content
- Topic-brand association reinforced by recurring content in the chosen territory
- External author and brand profiles consistent and active (the entity lives beyond the site)
- sameAs in schema pointing to the correct official profiles
Block 4: external validation
- Mentions in authoritative industry outlets (earned media built on data, not empty exchanges)
- Correct technical presence on the platforms AIs cite most in your topic (communities, video, professional)
- Cross-citations: other sources confirming the same facts your base declares
- Proprietary data distributed to press and creators (citable data generates citable mentions)
- Testimonials and cases published with consent and a methodology note
Block 5: measurement and iteration
- Strategic prompt list monitored monthly across the main systems
- Citation log: who is cited, with what framing, from which source
- AI features filter tracked in Google Search Console
- Leads and conversions tracked by origin, including self-declared origin
- Quarterly volatility review and editorial priority adjustment
What existing guides miss
Before writing this guide, we analysed the patterns of available content on GEO and AI citations. Without dismissing anyone's work, three gaps repeat. First, the absence of data: most texts define the concepts and recommend best practices without citing a single measurable study, which is ironic, since the academic evidence says statistics and sources are precisely what earns citations. Second, the outdated thesis: much of the material still treats GEO as an automatic extension of SEO ("rank well and you will be cited"), a claim the 2026 numbers contradict, with top-10-to-citation overlap falling from 76% to 38% in seven months. Third, the lack of operationalisation: concept without checklist, recommendation without audit criteria, promise without a measurement method. This guide exists to close those three gaps, and that is also why it declares methods, limits and sources: we apply to this very content the standard we recommend.
Eight mistakes that cancel the work
1. Blocking AI crawlers in robots.txt or the firewall and expecting citations. 2. Publishing volumes of generic AI-generated content, which systems deduplicate as redundancy. 3. Schema contradicting the visible content, which erodes trust instead of building it. 4. Ghost authorship: content with no identifiable author in a world that measures E-E-A-T. 5. Optimising a single heroic page per keyword and ignoring fan-out, which evaluates territories. 6. Trusting llms.txt or any single artefact as a silver bullet. 7. Ignoring external validation: an entity that only exists on its own site is a weak entity. 8. Not measuring: without a historical series of prompts and citations, strategy becomes guesswork.
Frequently asked questions
What is Generative Engine Optimization (GEO)?
GEO is the discipline of preparing a brand's digital presence to be understood, retrieved and cited by generative AI systems such as ChatGPT, Gemini, Perplexity and Google's AI Overviews. The term was formalised in the Princeton academic study presented at KDD 2024, which tested nine optimisation tactics and measured visibility gains of up to 40% for content with statistics, cited sources and clear language. While SEO competes for positions in lists of links, GEO competes for a place inside the answer itself.
What is the difference between GEO, AEO and SEO?
SEO optimises for traditional search engines, competing for rank and clicks. AEO (Answer Engine Optimization) structures information so a brand is eligible as the direct answer to specific questions, in featured snippets and answer engines. GEO prepares content to be used and cited by generative systems, which synthesise answers from multiple sources. The three disciplines overlap and reinforce each other: without SEO, a source often never enters the retrieved set; without GEO and AEO, a retrieved source never becomes a citation.
Does ranking well on Google guarantee AI citations?
No, and that is the most important shift measured in 2026. In July 2025, an Ahrefs study showed 76% of AI Overview citations coming from top-10 pages. In the February 2026 update, covering 863,000 keywords and 4 million cited URLs, that number fell to 38%, with the remainder split almost evenly between positions 11 to 100 and beyond the top 100. BrightEdge, using its own methodology, measures roughly 17% overlap with the top 10. Ranking still helps, but it stopped being sufficient: systems evaluate sub-queries (fan-out) and seek source diversity.
Which AI systems should a brand prioritise?
Google's AI Overviews and AI Mode, ChatGPT, Perplexity and Gemini cover most of the demand in Western markets, with regional variations: in Brazil, for instance, ChatGPT concentrated 99% of generative AI traffic in 2025 according to Cadastra and Similarweb data. The good news is that the fundamentals of citability (verifiable content, a consistent entity, structured data, external validation) serve every system at once, so the base work is shared and the monitoring is per platform.
What are retrieval and grounding, and why do they matter for GEO?
Retrieval is the stage where the system searches indexes and the web for candidate documents to feed the answer; grounding is the process of anchoring the generated answer in those documents, with citations. If your content is not retrievable (indexing, crawlability, topical clarity), it never enters the race; if it is retrieved but is not clear, verifiable and extractable, it does not become the citation. GEO works on both stages: getting into the retrieved set and maximising the chance of being the anchored source.
What is query fan-out?
It is the behaviour where the AI system decomposes the user's question into several related sub-queries, retrieves sources for each one, and synthesises everything into a single answer. In practice, your content can be cited for a sub-query you do not even track. That is why deep coverage of a topical territory, with connected pages answering different angles, tends to beat the strategy of one strong page for one keyword.
Is content with statistics and sources cited more often?
Yes, with academic evidence. The Princeton study (KDD 2024) measured visibility gains of up to 40% in generative answers for content that incorporates statistics, cites sources and uses clear, fluent language. The explanation is simple: systems that need to anchor claims prefer verifiable passages. Vague claims are not anchorable; sourced numbers are.
How much traffic does AI actually take from websites?
Clicks drop when an AI answer appears: the Pew Research Center measured users clicking traditional results in 8% of searches with an AI Overview versus 15% without, and clicking the cited sources inside the summary only 1% of the time. Seer Interactive measured organic CTR falling from 1.76% to 0.61% on queries with an AI Overview. On the other hand, the traffic that does arrive from AI is more qualified: analyses based on Semrush data point to roughly 4.4 times higher conversion, and Ahrefs reported AI visitors on its own site converting far above traditional organic.
Is being cited without getting the click worth anything?
Yes, increasingly so. A citation is brand exposure at the exact moment of doubt, with authority lent by the AI interface itself. Seer Interactive measured that brands cited in AI Overviews saw meaningfully better click performance than absent ones (a relative effect of about 35%), and behavioural studies show users relying on the answer to decide whom to trust. Beyond that, a share of high-value journeys still ends in a visit, a contact or a direct search for the cited brand.
What are entities and why do they decide who gets cited?
Entities are the knowledge units systems recognise: people, brands, methods, products and places, with attributes and relationships. AIs prefer attributing information to consolidated entities with consistent data across multiple sources. A brand with a well-structured entity (correct structured data, coherent presence, cross-citations) is easier to recognise and to cite. A brand with scattered, contradictory signals is statistically risky for the system, and tends to be passed over.
Does structured data (JSON-LD) influence AI citations?
It influences understanding, which is the gateway to citation. JSON-LD is not a magic citation switch, but it declares in machine-readable form who the organisation is, who the author is, what the page is about and how entities relate, reducing ambiguity. Google uses structured data to understand content, and AI systems benefit from the same clarity. The essential rule is coherence: the schema must reflect exactly what the visible content says.
Does llms.txt work for getting cited?
On its own, no. llms.txt is a proposed standard for offering AIs a clean map of a site's content, but the major providers have not confirmed official support, and there is no evidence that the file alone generates citations. It can be part of a machine-readability strategy, together with an official AI knowledge base and structured data, but the heavy lifting is done by citable content and a consolidated entity. We published a dedicated analysis of this topic on the blog.
Should I allow AI crawlers in robots.txt?
If the goal is to be cited, yes: blocking GPTBot, ClaudeBot, PerplexityBot, Google-Extended and similar crawlers takes your brand out of the game in the corresponding systems. The decision is strategic (some publishers block over copyright positions), but it requires coherence: you cannot block the robots and demand presence in the answers. The technical checklist in this guide includes auditing robots.txt, firewall rules and server logs to confirm the relevant crawlers reach your content.
How long does it take for a brand to start being cited?
It depends on the entity's starting point and the competitiveness of the territory. Early signals (appearances in niche answers, first mentions) can show up within a few months when the technical base is correct and the content has real information gain; consolidation as a recurring source is longer-term work, because it involves external validation and history. Citation patterns are also volatile: the Semrush study of 230,000 prompts showed dominant sources changing drastically within months. That is why continuous measurement is part of the method.
How do I measure whether AIs are citing my brand?
Combine four fronts: direct monitoring of relevant prompts across the main systems (manually or with answer-tracking tools), Google Search Console's AI features filter (which aggregates impressions and clicks from AI experiences), referral and mention analysis in your analytics, and lead source checks including self-declared origin. The important shift is treating citation as a marketing metric with a historical series, not a curiosity.
Can AI-generated content get cited by AI?
It can, but it starts at a structural disadvantage: text generated without new data tends to have high similarity with what already exists and near-zero information gain, exactly the profile systems deduplicate and demote. The intelligent use of AI in production is as a structuring and scaling tool over proprietary raw material (data, experience, positions), with expert review and real authorship. What defines citability is not who typed, it is the verifiable informational value.
Can small brands compete for citations against major publishers?
They can, within defined territories. Citation studies show distributed authority: Evertune's analysis of 200 million AI answers found no single domain holding more than 5% of citations. AIs cite by topic and by sub-query, not by company size. A specialised brand with primary data, a consistent entity and real depth in a niche can become that niche's default answer even against giant generalists. That is the logic of authority territories we apply with verticalised domains.
Reddit, YouTube and Wikipedia dominate citations. How do I leverage that?
International studies (Semrush, Peec AI) place encyclopaedias, communities such as Reddit, and platforms such as YouTube and LinkedIn among the sources AI systems cite most. The strategic reading is not to abandon your own site, but to distribute validation: correct technical presence on the platforms systems use as sources, your own citable content as the base, and data-driven PR generating mentions in authoritative outlets. Citation is born from the ecosystem, not from a single URL.
Does GEO require changing the site architecture?
It almost always requires adjustments: direct-answer blocks at the top of strategic content, real FAQs with matching schema, a clean heading hierarchy, graph-based structured data with consistent identifiers, pillar pages connected to supporting content, and healthy technical performance. You do not need to rebuild everything at once: the path is to audit, prioritise the highest-value territories and evolve the architecture in layers, as in this guide's checklist.
Does being cited by AI generate real business, or is it a vanity metric?
It generates measurable business when the work targets high-value intent. In the Dra. Ana Vega project, publicly documented by Flowup, the organic base accounted for 71.3% of June 2026 leads, and the month registered the first lead attributed to ChatGPT, an early GEO signal in a high-value market. Add to that the superior conversion data for AI-sourced traffic. Citation is the means: the end is entering decision journeys at the moment they happen.
What is the first practical step to start with GEO?
An honest diagnosis built on three questions: what do AIs answer today when asked about your topic and your brand, which sources do they use in those answers, and what exists in your ecosystem that would deserve to be cited. From that snapshot, prioritise in the method's order: technical base and crawlability, citable content with information gain, entity consolidation, external validation and measurement. Flowup runs exactly that initial mapping for clients worldwide.
When AI answers for your market, who does it cite today?
Flowup Agency has been engineering digital authority since 2011, for clients in Brazil and abroad. We run the mapping that opens this guide: what the systems answer about your topic, which sources they use, where your entity is fragile and which checklist layers to prioritise first. It's not enough to rank. Be the answer.
Talk to FlowupAbout the author
Methodology note
Sources verified on 3 August 2026. The cited studies use different methodologies, samples and time windows, and some measure different phenomena (AI Overviews, ChatGPT, multi-system panels); the figures are not directly comparable, and the divergence between Ahrefs (38%) and BrightEdge (17%) on top-10 overlap is declared in the text. The up-to-40% gain refers to the Princeton academic benchmark metrics, not to projections for individual cases. Conversion multipliers (4.4x; and Ahrefs' report on its own site) are third-party measurements in specific contexts. The Dra. Ana Vega project indicators are public, with approximate channel attribution and no guarantee of results. The analysis of gaps in existing guides describes market patterns without evaluating specific publications. No figure here should be read as a performance promise.
References
- Aggarwal et al. (KDD 2024). GEO: Generative Engine Optimization, the Princeton academic benchmark with gains of up to 40% for content with statistics, sources and clarity. arxiv.org/abs/2311.09735
- Search Engine Journal / Ahrefs (2026). AI Overview citations from top-10 pages drop from 76% to 38% (863,000 keywords, 4 million URLs), with fan-out as context. searchenginejournal.com/google-ai-overview-citations-from-top-ranking-pages-drop-sharply
- BrightEdge (2026). AI Overviews at the one-year mark: presence from 30% to 48% of tracked searches, summaries over 1,200 pixels, and roughly 17% citation overlap with the organic top 10. brightedge.com/resources/weekly-ai-search-insights/ai-overviews-one-year-presence-size-citing
- seoClarity (2025). The overlap between AI Overview citations and rankings: 94% cite at least one top-20 URL; position 1 cited in 43% of cases versus 7% for position 20. seoclarity.net/research/aio-rankings-overlap
- Search Engine Journal / Pew Research (2026). Consolidated CTR studies: 8% clicks with an AI Overview versus 15% without, 1% on cited sources, and Seer Interactive data (1.76% to 0.61%; cited brands performing ~35% better). searchenginejournal.com/ai-overview-ctr-fell-61-but-clicks-didnt-collapse
- Semrush (2025). Study of 230,000 prompts on the domains AI systems cite most and the volatility of citation patterns. semrush.com/blog/most-cited-domains-ai
- Search Engine Land / Peec AI (2025). Analysis of 30 million citations: Reddit, YouTube and LinkedIn among the sources AI systems use most. searchengineland.com/ai-search-engines-cite-reddit-youtube-and-linkedin-most-study
- Contently / Evertune (2026). 200 million AI answers analysed: no single domain concentrates more than 5% of citations. contently.com/2026/04/29/top-sources-llms-cite
- PikaSEO / Semrush (2026). Compiled conversion data for AI-sourced traffic (about 4.4 times traditional search). pikaseo.com/articles/zero-click-search-ai-overviews-2026
- Blog Tropk / Cadastra + Similarweb (2025). The Brazilian AI search market: ChatGPT with about 310 million monthly visits and 99% of national generative AI traffic. blog.tropk.ai/o-mercado-brasileiro-de-ai-search-em-2026
- Google Search Central. Spam policies and structured data documentation: mandatory coherence between schema and visible content. developers.google.com/search/docs/essentials/spam-policies
