GEO & AEO: the guide to showing up in ChatGPT, Claude, Gemini and Perplexity.

The Generative Engine Optimization (GEO) discipline is born academically with "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik R. Narasimhan and Ameet Deshpande (Princeton, Georgia Tech, Allen Institute for AI and IIT Delhi).

First version: arXiv 2311.09735 v1, November 16, 2023. Final v3 on June 28, 2024. Official publication at KDD '24 (ACM SIGKDD Conference on Knowledge Discovery and Data Mining), Barcelona, August 25 to 29, 2024. DOI 10.1145/3637528.3671900. Public resources: code on GitHub (GEO-optim/GEO), dataset on HuggingFace (GEO-optim/geo-bench) and the official site generative-engines.com/GEO/.

In plain terms, the paper defines a generative engine (GE) as a system with three moving parts. First, a set of generative models (query reformulation, summarization, final answer). Second, a search engine returning candidate sources. Third, a response written in sentences, each sentence optionally backed by a subset of the retrieved sources. This abstraction covers Bing Chat, Google SGE and AI Overviews, Perplexity, ChatGPT Search and Gemini under a single framework. The formal version is: f_GE: (q_u, P_U) to r, where r is the response and each sentence carries its own citation set.

Etymology matters here because it reveals different layers of the same problem:

• ▸SEO (Search Engine Optimization): coined in the mid-1990s, consolidated in the early 2000s with Google's dominance. Optimizes for link rankings.
• ▸AEO (Answer Engine Optimization): Jason Barnard (Kalicube) is credited with coining the term in 2018, in the context of featured snippets, Google's Answer Box, Alexa, Siri and Google Assistant. Optimizes for direct answers.
• ▸GEO (Generative Engine Optimization): academically formalized in November 2023 by Aggarwal et al. Optimizes for inclusion in LLM-synthesized responses.
• ▸LLMO, AIO, AIEO: practical terms, largely functional synonyms of GEO. Neil Patel Brazil puts it bluntly: "you can use GEO, AIO or LLMO, the important thing is to understand it's the same thing." Wikipedia (entry "Generative Engine Optimization") confirms there's no academic consensus yet to rigidly distinguish AEO and GEO in 2026.

• SEO: the set of technical and editorial practices that maximize visibility and ranking of web pages in SERPs (Search Engine Results Pages) based on classic retrieval plus ranking.
• AEO: content optimization to be returned as a direct answer by answer engines: featured snippets, Position Zero, knowledge panels, voice assistants, PAA, and non-generative SERP components.
• GEO: optimization to be cited, referenced or paraphrased inside synthesized responses from generative (RAG-powered) engines like ChatGPT, Perplexity, Claude, Gemini, Copilot, Google AI Overviews and AI Mode.

In practice, AEO and GEO converge. Microsoft Advertising even treats them as phases (AEO = discovery and indexing, GEO = citation assessment). This study treats AEO and GEO as complementary and overlapping disciplines, in line with the practice consolidated by HubSpot, Semrush, Frase and Search Engine Land.

No.

Lily Ray (Amsive) has observed that AI citations frequently mirror organic SEO strength: 92.36% of AI Overviews citations come from domains in Google's top 10. Seer Interactive studies show brands cited in AI Overviews gain 35% more organic clicks and 91% more paid clicks on the same queries. Rank well, get cited. Simple.

SEO is the base. AEO is the direct-answer layer. GEO is the synthesis-inclusion layer. Gartner reinforces: content quantity, quality, E-E-A-T and watermarking will be the differentiators in 2026.

The paper introduces metrics because classic SERP-ranking doesn't apply to synthesized answers (citation position, length and style vary).

• ▸Word Count (Imp_wc): fraction of response words attributable to citation cᵢ, split equitably when the sentence has multiple citations.
• ▸Position-Adjusted Word Count (Imp_pwc): adds exponential decay by position, motivated by power-law CTR studies (Goodwin 2011; Dean 2023).
• ▸Subjective Impression: LLM-as-judge score (G-Eval + GPT-3.5) across 7 sub-metrics, Relevance, Influence, Uniqueness, Subjective Position, Subjective Count, Click-likelihood, Diversity.
• ▸Relative improvement: (Imp(r′) − Imp(r)) / Imp(r) × 100%.

These concepts form the core of modern thinking about "AI visibility".

SparkToro research (Rand Fishkin, 2025): LLM responses are highly inconsistent when recommending brands. Fishkin proposes measurement must be statistical at scale, not on individual responses. A Genezio test with six different tools applied to the same brand (Honda) produced radically divergent results, "each tool told a different story".

Practical implication: monitor aggregated trends (averages across hundreds of prompts over time), not snapshots.

In practice, a mature AEO/GEO dashboard combines four layers:

1. 01GSC + GA4: traditional organic traffic + queries with high impressions and low CTR (classic signal of AIO or featured snippet capturing the click).
2. 02AI visibility platform (Profound, Peec, Otterly): SoV, citation rate, sentiment by LLM.
3. 03AI crawler log analytics (Cloudflare / Profound Agent Analytics): who is crawling and at what volume.
4. 04Manual QA: monthly testing of the top 30–50 prompts across ChatGPT, Perplexity, Gemini, Claude, essential to calibrate automated platforms.

Aggarwal et al. built GEO-Bench: 10,000 queries (8K train, 1K validation, 1K test), 9 datasets (MS MARCO, ORCAS-I, Natural Questions, AllSouls, LIMA, Davinci-Debate, Perplexity Discover, ELI5, GPT-4 generated), 25 domains, realistic distribution (~80% informational, 10% transactional, 10% navigational). Two-step GE pipeline mimicking Liu et al. 2023: Google top-5 + GPT-3.5-turbo with citations; 5 samples temperature 0.7; 5 seeds.

1. Authoritative, authoritative tone
2. Statistics Addition, add statistics
3. Keyword Stuffing, classic SEO control
4. Cite Sources, add citations
5. Quotation Addition, add quotes
6. Easy-to-Understand, simplify language
7. Fluency Optimization, improve fluency
8. Unique Words, unique words
9. Technical Terms, technical terminology

Baseline: ~19.3–19.5 on both metrics.

Reading: the three winners all rely on external evidence: Quotation Addition, Statistics Addition, Cite Sources. Fluency Optimization confirms LLMs value presentation, not just content. Keyword stuffing is counterproductive. The most cartoonish classic SEO tactic fails in GE.

Perhaps the most important finding for small businesses:

• ▸Rank 5 sources gain up to +115.1% visibility with Cite Sources; +99.7% with Quotation Addition; +97.9% with Statistics Addition.
• ▸Rank 1 sources can lose up to −30% when all sources optimize.

What the authors say: GEs don't lean on backlinks or domain authority like classic search engines. That levels the field for smaller creators. This is the philosophical divide between SEO and GEO. SEO rewards established domains. GEO rewards well-structured content, regardless of the source.

• ▸Authoritative works best in Debate, History, Science.
• ▸Cite Sources shines in Factual, Law & Government.
• ▸Quotation Addition dominates People & Society, Explanation, History.
• ▸Statistics Addition wins in Law & Gov., Debate, Opinion.

Technique pairs beat single applications by >5.5%. The best pair: Fluency Optimization + Statistics Addition. Cite Sources synergizes well even when weak individually.

Paper's Table 5, "GEO in the wild": Quotation Addition +22% in Pos-Adj Word Count; Statistics Addition +37% in Subjective Impression. Keyword Stuffing again −10%. The techniques transfer to real production systems, with different magnitudes but consistent direction.

• AutoGEO (Wu et al., arXiv 2510.11438, accepted at ICLR 2026): RL trained to learn GE preferences. Reports +50.99% over Fluency Optimization (Aggarwal's strongest baseline). Presents AutoGEO_API (prompt-based) and AutoGEO_Mini (compact model).
• Chen et al., "GEO: How to Dominate AI Search" (arXiv 2509.08919, 2025): AI Search has a systematic bias in favor of "Earned Media" over brand-owned/social content.
• AgenticGEO: self-evolving system, state-of-the-art on GEO-Bench.
• E-GEO (arXiv 2511.20867): extends the testbed to e-commerce (Columbia/MIT).
• Kumar & Lakkaraju (2024), "Manipulating LLMs to Increase Product Visibility" (arXiv 2404.07981): adversarial counterpoint to GEO, demonstrates that "strategic text sequences" can inject bias into LLM recommendations, raising ethical alarms.
• Liu, Zhang, Liang (Stanford, EMNLP Findings 2023), "Evaluating Verifiability in Generative Search Engines" (arXiv 2304.09848): found that only 51.5% of generated sentences are fully supported by citations and 74.5% of citations actually support their claims.

AirOps research: pages with clear structure plus schema markup get 2.8× more AI citations. JSON-LD is Google's preferred format. The schemas that matter:

• ▸FAQPage: essential for PAA and AI citations.
• ▸HowTo: step-by-step with HowToStep, tool, supply.
• ▸Article + Person (author): demonstrable E-E-A-T.
• ▸Organization + sameAs: Knowledge Graph and Wikidata connection.
• ▸BreadcrumbList: hierarchy.
• ▸Product + AggregateRating: e-commerce with reviews.
• ▸QAPage: prioritized by ChatGPT and Perplexity. It has upvoteCount, answerCount and dateCreated, which FAQPage lacks.
• ▸SpeakableSpecification: voice (beta, US news publishers).
• ▸ClaimReview: fact-checking.

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "What is Answer Engine Optimization?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "AEO is the practice of structuring content so that AI systems like ChatGPT, Perplexity, Claude and Google AI Overviews can extract, understand and cite it as an authoritative answer."
    }
  }]
}

Cross-platform connection via sameAs (critical for Knowledge Graph): LinkedIn, Wikidata QID, Wikipedia, Crunchbase, Twitter/X, GitHub.

Typology and tactics:

• ▸Paragraph snippet (~82% of FS): H2 with literal question + 40–60-word answer immediately below.
• ▸List snippet (numbered/unordered): `ol`/`ul` with 5–8 items; for how-to, each step as H3 + short paragraph.
• ▸Table snippet: actual HTML `table` (not image), minimum 3 columns × 4 rows.
• ▸Video snippet: VideoObject schema + transcript + timestamps (SeekToAction/Clip).

Canonical formula: [question as H2] + [direct answer 40–60 words] + [elaboration] + [list/table].

• ▸Mining with AlsoAsked, AnswerThePublic, Semrush Topic Research.
• ▸Each PAA question becomes an H2/H3.
• ▸Answer immediately below in 2–3 sentences.
• ▸Stack related PAA clusters.

• ▸Conversational language; 7+ word queries.
• ▸Average response ~29 words (Backlinko).
• ▸Reading level 9th grade or below (Flesch 60–70+).
• ▸LocalBusiness schema, consistent NAP, up-to-date GBP.
• ▸SpeakableSpecification on key snippets.
• ▸FAQ starting with "how", "what", "when", "where", "why".
• ▸HTTPS (~70% of voice results).

• HubSpot, the most cited AI Overviews decline case: monthly organic traffic dropped from ~13.5M in Nov/2024 to <7M in Dec/2024, reaching 6M in subsequent months (Ahrefs data). CEO Yamini Rangan admitted on the earnings call that "organic search traffic is declining globally" and that "AI overviews are giving answers, and fewer people are clicking through to websites". A LinkedIn analysis sums it up: "If HubSpot, with one of the best SEO teams in the world, can experience this, none of us are safe".
• DMG Media (Daily Mail): CTR dropped from 25.2% on regular desktop to 2.8% when AI Overviews appear, an 89% decline. However, AI Overviews rarely trigger for breaking news, limiting aggregate impact.
• Bank of America (Profound, Jun/2025): 32.2% visibility across AI platforms for banking queries. Navy Federal Credit Union gained disproportionate representation vs. traditional advertising investment.
• AthenaHQ: publishes the strongest case studies in the category, reports 10× citation growth and 50% more demos for customers.
• Peec AI: published a 1M-citation benchmark study and an analysis of 232k listicle citations, original research at significant scale.
• Semrush study (89k LinkedIn URLs): 11% of AI responses cite LinkedIn, and individual authors are being cited, not company pages.

Semrush Study (10M+ keywords, 2025):

• ▸AIO prevalence: 6.49% in Jan/2025 → peak ~25% in Jul/2025 → 15.69% in Nov/2025.
• ▸Jan/2025: 91.3% of AIO queries were informational; Oct/2025: 57.1% (with growth in commercial and transactional).
• ▸Navigational with AIO: 0.74% → 10.33% (Jan–Oct 2025).
• ▸Google Ads on SERPs with AIO: 25.56% in Oct/2025 (+394% vs Mar/2025).

Most saturated industries (% of keywords with AIO): Science, Computers & Electronics, People & Society (>17%). Least saturated: Real Estate, Shopping, Arts & Entertainment (<3%). Fastest growth Jan–Mar 2025: Science (+22.27%), Health (+20.33%), People & Society (+18.83%), Law & Gov (+15.18%).

Critical insight (Seer): brands cited in the AIO gain +35% organic clicks and +91% paid clicks on the same query. Being cited in the AIO is worth more than ranking #1 below it.

• ▸800M+ WAU in Dec/2025; ~900M in Feb/2026 (OpenAI / Superlines / Arvow).
• ▸2.5 billion prompts/day (July 2025); 18 billion messages/week.
• ▸Annualized revenue: US$ 6B (2024) → US$ 20B (2025) → US$ 25B+ (Feb/2026). 2026 target: US$ 29.4B.
• ▸Enterprise seats: >7M, +9× YoY, 88% retention after 12 months.
• ▸34% of US adults have used ChatGPT (Pew, 2025); 58% among those under 30.
• ▸Chatbot market share (StatCounter Jan/2026): ChatGPT 79.98%; Perplexity 7.89%; Gemini 7.18%; Copilot 3.5%.
• ▸ChatGPT accounts for ~82% of AI platform referrals to the open web (Statcounter May/2025: 79.8%).
• ▸AI-referred traffic converts ~4.4× better than organic search (Superlines/Playwire).

• ▸~30M MAU (Apr/2025); 45M (H2 2025).
• ▸780M queries in May/2025 (30M/day); +20% MoM.
• ▸239.97M visits in Nov/2025; 80.5M mobile downloads.
• ▸Valuation: ~US$ 18–20B. ARR: ~US$ 148M (2025).

• ▸AI tools grow from 0.24% → 0.64% of US desktop usage (SparkToro/Datos 2025). Europe: 0.26% → 0.78%.
• ▸Fishkin: Google still runs ~210× more searches/day than ChatGPT; traditional search is NOT in absolute decline, it's losing relative share.
• ▸Zero-click: 58% in 2024 → 60% in 2025 (Semrush/Datos).
• ▸Semrush/Datos: when the same keyword gets an AIO, zero-click rate dropped slightly (33.75% → 31.53%), AIO is not a direct cause of zero-click, it's a symptom of already-hard-to-convert intent.

• ▸Googlebot: 5:1 (the classic model).
• ▸OpenAI: 785:1 (Oct/2025) → peak 1,851:1 (Dec/2025) → 692:1 (Apr/2026), 56% improvement with OAI-SearchBot active.
• ▸Anthropic: 10,347:1 (up to 73,000:1 at some points; peaks of 500,000:1).
• ▸Perplexity: 118:1.

Origin: Jeremy Howard (Answer.AI) on Sep 3, 2024 (llmstxt.org). Solves the problem of small context windows + HTML noise (ads, JS, navigation).

• ▸/llms.txt, Markdown summary of the site structure (curated navigation).
• ▸/llms-full.txt, full documentation in concatenated Markdown.

Required structure: H1 with project name, blockquote with summary, optional Markdown sections, H2 sections with link lists [Title](URL): description, ## Optional section for less essential resources.

• ▸Ahrefs: ~10% of domains have implemented it (sample biased toward tech/dev).
• ▸SE Ranking (39k domains): only 0.13%.
• ▸Notable adopters (tech docs): Anthropic, Cursor, Bolt.new, Windsurf, Mintlify.
• ▸John Mueller (Google Search Advocate): "FWIW no AI system currently uses llms.txt", compared to the old meta keywords tag.
• ▸Gary Illyes (Google, Search Central Live Jul/2025): "Google doesn't support LLMs.txt and isn't planning to".
• ▸Carolyn Shelby (Yoast) rebuts the meta keywords analogy: "LLMs tend to drop into specific pieces of content, Mission Impossible-style… This context-first approach is exactly what llms.txt enables".
• ▸December 2025: Google added llms.txt to its own docs then removed it, Mueller confirmed "not an endorsement".

Practical 2026 consensus: implement as low risk / low effort, but don't prioritize over robots.txt, schema, E-E-A-T.

1. Training crawlers (block = protect IP): GPTBot, ClaudeBot, anthropic-ai, CCBot, Google-Extended, Applebot-Extended, Meta-ExternalAgent, Amazonbot, cohere-ai, Bytespider, Diffbot.
2. Search/Retrieval crawlers (block = disappear from AI search): OAI-SearchBot, Claude-SearchBot, PerplexityBot, YouBot, DuckAssistBot, Applebot, Bingbot, Googlebot.
3. User-action fetchers (high referral value): ChatGPT-User, Claude-User, Perplexity-User, MistralAI-User, Google-NotebookLM, Meta-ExternalFetcher.

Balanced strategy: allow retrieval/user-action; block training bulk.

# Allow retrieval and user-action (AI visibility)
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /

# Block training bulk (IP protection)
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
User-agent: Google-Extended
Disallow: /
User-agent: CCBot
Disallow: /
User-agent: Applebot-Extended
Disallow: /

# Traditional search
User-agent: Googlebot
Allow: /
User-agent: Bingbot
Allow: /

• ▸Cloudflare top 10,000 Jan/2025: GPTBot 7.8% disallow; Google-Extended 5.6%.
• ▸GPTBot blocked on ~5.6M sites (Oct/2025, +70% vs Jul/2025).
• ▸ClaudeBot: 3.2M → 5.8M sites (+81%) Jul–Dec 2025.
• ▸Tollbit Q2/2025: +336% YoY in blocking; 13.26% of AI requests ignore robots.txt (vs 3.3% in Q4/2024).
• ▸arXiv study: blocking on reputable sites: 23% (Sep/2023) → ~60% (May/2025); average of 15.5 forbidden user agents.
• ▸ALM Corp (publishers): 69% block ClaudeBot; 62% GPTBot; 49% OAI-SearchBot; 40% ChatGPT-User, 71% inadvertently block at least one retrieval bot, destroying their AI search presence unintentionally.
• ▸Cloudflare (Jul–Dec 2025): 416 billion AI bot requests; 2.5M sites use managed robots.txt; 88% of top news outlets block AI crawlers (Wired).
• ▸Cloudflare since Jul/2025: blocks AI crawlers by default on new domains (protecting ~20% of the web).

• ▸JSON IP ranges published by OpenAI, Google, Perplexity, Bing.
• ▸Reverse DNS (dig -x IP → PTR should match the vendor domain).
• ▸Web Bot Auth (RFC 9421 / HTTP Message Signatures), emerging cryptographic authentication standard.

Still relevant indirectly: LCP ≤2.5s, INP ≤200ms (replaced FID in Mar/2024), CLS ≤0.1. More critical: SSR/SSG required, AI crawlers are typically lightweight and don't render heavy JS. Aleyda Solis and Lily Ray report cases of sites "invisible" to AI because content depends on client-side JS.

• 2014: E-A-T introduced.
• Dec/2022: second "E" (Experience) added → E-E-A-T.
• Jan/2025 QRG: new section 4.7, AI content without human review qualifies as Lowest quality if "created to benefit owner…with very little or no attempt to benefit website visitors". Targets "scaled content abuse". Inflated filler content is penalized. "As a language model, I don't have…" → automatically Lowest.
• Sep/2025 QRG: 181 → 182 pages; first time with rating examples for AI Overviews; YMYL renamed to "YMYL Government, Civics & Society".

Critical quote from QRG 2025: "Trust is the most important member of the E-E-A-T family because untrustworthy pages have low E-E-A-T no matter how Experienced, Expert, or Authoritative they may seem."

AI Overviews recycle organic result eligibility (Google Search Central, 2025), E-E-A-T fundamentals transfer directly to AI citations.

Critical elements: `article`, `section`, `main`, `header`, `nav`, `aside`, `figure`, `figcaption`, `time`, `cite`, `blockquote`. Provide "anchors" for embeddings and reduce hallucination risk.

Chunking sweet spot: 100–300 tokens (~75–225 words) per section. One topic per section. Descriptive headings (not generic "Introduction"). Tables/lists as atomic chunks. Avoid context-dependent references ("as mentioned above", "see chart below").

Chunking strategies used by LLMs: heading-based (most common), fixed-size with overlap (200–400 tokens + 50 overlap), recursive (paragraph→sentence→word), semantic chunking via embedding similarity, LLM-guided chunking, late chunking.

Search Engine Land (Oct/2025): entity-driven content increases AI citation probability by +35%.

• Conversion (founded 2011 by Diego Ivo, 100+ specialists, R$ 12 billion generated for clients) officially launched in 2025 its first dedicated GEO service, with its entire current portfolio already running GEO initiatives. Diego Ivo: "SEO is going through its biggest inflection point in the last three decades. Those who invest now will come out ahead." Conversion cites Gartner (25% by 2026) and Semrush (AI search traffic will surpass traditional by 2028).
• Wyse (wyse.com.br/geo) positions itself as a pioneer in a proprietary GEO/AEO methodology integrating technical SEO, structured content and digital reputation.
• Bloomin (São Paulo) publishes extensively on GEO and offers services, with positioning in E-E-A-T and digital PR. Data: São Paulo companies that adopted GEO allocate 10–20% of the marketing budget to it.
• Rock Content maintains one of the earliest robust articles in Portuguese on the topic (written in 2024). Giuseppe Caltabiano (VP of Marketing): "Generative Engines represent a transformative shift in the search engine paradigm, offering direct and comprehensive answers to user queries and thus potentially reducing the need to visit sites directly."
• RD Station (Resultados Digitais) covers the topic on its educational blog, more as a trend than an integrated product.
• Neil Patel Brazil has one of the most accessible Portuguese guides equating GEO/AIO/LLMO and translating Eric Siu's comparison table (SEO vs SGE vs GEO).

• ▸Diego Ivo (Conversion), institutional voice on GEO in Brazil.
• ▸Fábio Ricotta (Agência Mestre), traditional SEO with transition to AI.
• ▸Vitor Peçanha (Rock Content), co-founder, educational content.
• ▸Giuseppe Caltabiano (Rock Content), leadership perspective on GE.
• ▸Roberto Dias Duarte (RDD10+), translates international material (e.g., adapted Charlie Guo).
• ▸Camila Porto, Mauricio Cardoso, Alex Peixoto, Diana Martins, active BR SEO pros with posts on LinkedIn in 2025/2026.

Rock Content, Conversion, RD Station, Neil Patel Brasil, Bloomin, Orgânica Digital, bfind, Prosperidade Conteúdos. RDD10+ (Roberto Dias Duarte). Manus (Practical GEO Guide). Wyse blog.

• ▸RD Summit (Brazil's largest digital marketing event, annual).
• ▸SEO Summit and SEOnStage.
• ▸Digitalks.
• ▸MKT4Edu.
• ▸Courses on Alura, Hotmart, Udemy PT-BR.
• ▸YouTube channels and Brazilian LinkedIn groups growing fast in 2025.

• ▸Sabiá (Maritaca AI): family of models natively trained in Brazilian Portuguese; Sabiá-3 launched in 2024. Reference for applications requiring native PT-BR fluency.
• ▸Gemini has particularly strong performance in Portuguese (Google trained with an extensive Brazilian corpus).
• ▸ChatGPT and Claude work well in Portuguese but occasionally show bias toward anglicized constructions; quality varies by domain.
• ▸PT-BR vs PT-PT variations affect relevance: models tend to prioritize PT-BR (larger data volume), which penalizes Portugal-based content on Brazilian queries.
• ▸Brazilianisms, regional slang and PT-BR technical terms (e.g., "celular" vs "telemóvel", "ônibus" vs "autocarro") should be made explicit.

Gradually rolled out in Brazil throughout 2024–2025. In 2025, coverage expanded significantly. Semrush and Conversion have published PT-specific AIO analyses showing patterns similar to the US: saturation in health, education, law and science; severe impact on publishers. G1, UOL and Folha saw measurable declines on informational queries.

• ▸NILC/ICMC-USP (Interinstitutional Center for Computational Linguistics), leading NLP group in Portuguese.
• ▸UFMG (DCC, IR and NLP group).
• ▸UNICAMP (IC).
• ▸UFRGS (INF).
• ▸Maritaca AI (USP/UNICAMP spin-off), Sabiá models.
• ▸Brazilian publications on GEO proper are still scarce in 2025/2026, a clear academic research opportunity.

• ▸Brazil is consistently top 5 globally in ChatGPT usage (often top 3 on some metrics), high adoption for a middle-income country.
• ▸AI adoption grows 4× faster in low/middle-income vs high-income countries (NBER), Brazil benefits from this trend.
• ▸StatCounter Brazil shows ChatGPT as the dominant AI referrer, with Perplexity and Gemini growing fast.

Main gap in Brazil: lack of PT-BR-specific quantitative benchmarks and original academic research on LLM citation in Portuguese.

Universities: Princeton (Narasimhan, Deshpande, Murahari), Georgia Tech (Rajpurohit), IIT Delhi (Aggarwal), Stanford (Liu, Liang), Harvard (Kumar, Lakkaraju), Columbia, MIT.

Industrial labs: Allen Institute for AI (Kalyan), Google DeepMind, OpenAI, Anthropic, Meta AI, Microsoft Research.

1. Automated GEO: RL and LLM-guided optimization (AutoGEO).
2. Adversarial robustness: defense against LLM manipulation.
3. Domain-specific GEO: e-commerce (E-GEO), healthcare, legal.
4. Multimodal GEO: image, video, audio as sources.
5. Agentic pipelines: GEO for autonomous agents (not just chatbots).
6. Fair attribution: how to distribute fair credit across cited sources.

• ▸Longitudinal studies (how responses evolve month by month).
• ▸Research in non-English languages, PT, Spanish, Hindi.
• ▸Attribution economics, revenue-share models between publishers and LLMs.
• ▸Causal impact assessment (almost everything is correlational).
• ▸Domain-specific benchmarks (health, finance).

Gartner (press release Feb/2024, consistent in 2025–2026):

"By 2026, traditional search engine volume will drop 25%, with search marketing losing market share to AI chatbots and other virtual agents.", Alan Antin, VP Analyst.

Gartner also projects 70% of consumers with "some confidence" in AI search and that tech/services companies with >US$ 5B in revenue will spend 10% of marketing on content monitoring and brand reputation tied to AI-generated content.

Semrush projects that AI search traffic will surpass traditional search by 2028.

SparkToro (Rand Fishkin, 2025): if AI tool growth doubles annually, AI can rival traditional search in 6–10 years. But currently: Google does ~210× more searches/day than ChatGPT; ChatGPT has less desktop traffic than DuckDuckGo. Traditional isn't in absolute decline, it's losing relative share.

• ▸58% in 2024 → 60% in 2025 (Semrush/Datos).
• ▸Semrush discoveries: when a keyword gains an AIO, zero-click drops slightly (33.75% → 31.53%), AIO isn't a direct cause of zero-click, it's a consequence of intent type.
• ▸AI Overviews work similarly to featured snippets, they accelerate a trend that has existed since 2016.

• ▸OpenAI Operator (Jan/2025), autonomous agent that navigates and executes actions.
• ▸Anthropic Computer Use, Claude controlling the desktop.
• ▸Google Project Mariner / Deep Research, multi-step agents.
• ▸Perplexity Comet Browser, AI-first browser.
• ▸ChatGPT Atlas/Agent Browsers, AI-centered browsers.

Implications: agents don't need human UX; they consume content in nodes (JSON, Markdown, semantic HTML). Agent-optimized content delivery becomes a differentiator. Adobe LLM Optimizer already offers "content delivery optimization", AI-readable versions served directly to crawlers.

• ▸Google Lens growing fast.
• ▸ChatGPT with images (Ghibli mode went viral in 2025, boosting downloads).
• ▸Voice commerce via Alexa, Google Assistant, Siri, ChatGPT voice mode.
• ▸Video search optimization (VideoObject, transcripts, timestamps).
• ▸Speakable schema for TTS and voice.

• ▸NYT vs. OpenAI/Microsoft (2023–2026, NYT won preliminary ruling in 2025).
• ▸NYT vs. Perplexity (Dec/2025), accuses RAG of reproducing verbatim content.
• ▸Chicago Tribune, Reddit, Encyclopedia Britannica, Merriam-Webster, Nikkei, Asahi Shimbun, News Corp vs. Perplexity.
• ▸Authors/publishers vs. Anthropic, US$ 1.5B settlement in 2025 (pirated books).
• ▸Getty Images vs. Stability AI.
• ▸Amazon threatens Perplexity (Aug/2025) for shopping agent posing as human.

• ▸OpenAI + NewsCorp, Axel Springer, Le Monde, Financial Times.
• ▸Meta + CNN, Fox News, People Inc, USA Today (Dec/2025), crawl in exchange for links, citations and cash.
• ▸Microsoft Publish Content Marketplace (Sep/2025), FT, Reuters, People Inc, Axel Springer; pays publishers a commission when content feeds Copilot responses.
• ▸Reddit-Google deal (2024), US$ 60M/year.
• ▸Perplexity Comet Plus, US$ 42.5M pool for creators (2025).

Cloudflare vs. Perplexity (Aug/2025): Cloudflare accused Perplexity of Chrome user-agent spoofing, ASN rotation, ignoring robots.txt. Cloudflare delisted Perplexity from the Verified Bots program. Matthew Prince (Cloudflare): "It shouldn't be that you can use your monopoly position of yesterday in order to leverage and have a monopoly position in the market of tomorrow." Since Jul/2025, Cloudflare blocks 400B+ crawl attempts and introduced experimental pay-per-crawl.

• ▸EU AI Act (in effect 2024–2026 gradually), requires transparency in training data.
• ▸Brazil PL 2338/2023 (AI Civil Framework), under Senate review; establishes copyright and transparency.
• ▸Copyright rulings: in Anthropic, legally acquired books may be fair use; pirated infringes copyright.

• ▸Gen Z uses AI as first option for research, especially education (Pew 2025: 58% of US adults <30).
• ▸Trust in AI answers has grown, but remains moderate.
• ▸Surface multiplication for discovery: Reddit, TikTok, YouTube, LinkedIn, Substack are starting to compete with Google for the research "entry point".

• Rand Fishkin (SparkToro): "AI chat tools are growing, but their influence is limited compared to traditional search." Recommends focusing on "where your audience actually spends time" via SparkToro or Similarweb; build products "that market themselves".
• Aleyda Solís (Orainti): "Answer Engine Optimization is about clarity, context, and corroboration. The more your facts align with trusted sources, the more AI trusts your content to speak for you." In a Humans of Martech interview (Jan/2026): AI crawlers often blocked without the owner knowing; JavaScript hiding content; importance of topics over individual prompts; community (Reddit, YouTube comments) as a strong signal for LLMs.
• Lily Ray (Amsive): warns against aggressive GEO tactics that can trigger spam policies; AI citation often mirrors organic SEO strength, fundamentals matter.
• Jason Barnard (Kalicube): "Search, Answer, and Assistive Engine Optimization is the 2025 reality."
• Gary Illyes (Google): "Google doesn't support LLMs.txt and isn't planning to."
• Sam Altman (OpenAI, TED 2026): ~10% of the world population uses ChatGPT systems.

• ▸Adversarial prompt injection (Kumar & Lakkaraju 2024): "strategic text sequences" can inject bias into recommendations.
• ▸Negative SEO via LLM manipulation: planting misleading content in sources LLMs consume.
• ▸LLM inconsistency (SparkToro): impossible to optimize for a single response; only works at statistical scale.
• ▸Lily Ray: aggressive GEO tactics can trigger Google spam policies and lead to penalties.

• ▸HubSpot, Daily Mail, CNN (-27 to -38%), Men's Journal (-415% variable) document severe decline.
• ▸Attribution problem: GA4 doesn't consistently isolate AI-referred traffic. Google Search Console groups AI Mode clicks in the "Web" type, with no separate filter.
• ▸AI exposure without a click still influences purchase, but isn't directly measurable.
• ▸"Great Decoupling": search usage grows, site clicks drop.

• ▸NYT vs. Perplexity (2025): Perplexity allegedly hallucinated information attributing it to the NYT, damaging the brand.
• ▸Liu et al. (2023): 25.5% of citations do NOT support the claim they reference.
• ▸Reputational risk for brands misrepresented by LLMs.

• ▸Wikipedia, Reddit, YouTube, Quora are the most cited sources, creating structural bias against brand-owned content.
• ▸Chen et al. (2025): AI Search favors "Earned Media" over brand-owned/social.
• ▸English dominance: models perform worse and cite fewer sources in PT, Spanish, Hindi.

• — Is it ethical to optimize for LLMs?
• — "Algorithmic manipulation" vs "genuinely more useful engineering"?

Consensus: answer-first, E-E-A-T, schema techniques are ethical and value-adding; keyword stuffing / prompt injection are manipulative.

• ▸Publisher median: -10% YoY traffic in H1/2025.
• ▸News publishers: -7%; non-news content: -14%.
• ▸HubSpot: -70 to -80%.
• ▸CNN: -27 to -38%.
• ▸Some recipe sites: -50 to -70%.

• ▸Model cutoff dates: Gemini/ChatGPT sometimes cite content from months before the present.
• ▸RAG limitations: retrieval doesn't always find the ideal source.
• ▸Context window: even with 1M+ tokens, LLMs "forget" in the middle of long conversations.
• ▸Tool fragmentation (Genezio): six visibility tools produce six different narratives for the same brand.

The AI visibility market went from zero to ~US$ 848 million in less than two years. Clear segmentation:

• ▸Schema validation: Google Rich Results Test, Schema.org validator, Schema Markup Validator.
• ▸llms.txt: directory.llmstxt.cloud, llmstxt.directory (Mintlify auto-gen).
• ▸AI bot detection: Knowatoa AI Search Console (24 UAs), BlogPros Robots.txt AI Checker.
• ▸Crawl/render audit: Screaming Frog SSR mode, Sitebulb, Ahrefs.
• ▸Content chunking audit: Presenc AI RAG Fetchability, AirOps.
• ▸Entity/Knowledge Graph: InLinks, Kalicube Pro, Wikidata.
• ▸PAA/Question mining: AlsoAsked, AnswerThePublic.