LLM SEO: How To Get Cited Inside ChatGPT, Claude, and Gemini

LLM SEO is the practice of getting your content cited, paraphrased, or recalled inside the output of a large language model. The output can be a search-attached answer (ChatGPT with search, Perplexity, AI Overviews), a pure chat response with no web search at all (regular ChatGPT, Claude, the Gemini app), an agent's tool call, or a voice assistant reading from an LLM. LLM SEO is the umbrella label that covers all of those surfaces, and it recognises two distinct paths to citation: retrieval, where the model fetches your page live, and training-corpus recall, where the model 'remembers' your content from pretraining.

GEO targets the synthesised generative answer specifically: AI Overviews, ChatGPT with search, Perplexity. AEO targets any surface where the answer is lifted into place (featured snippets, People Also Ask, voice answers, AI Overviews, chat answer cards). LLM SEO is broader than both. It includes the cases where there is no live web fetch at all — a user asking ChatGPT a question without enabling search, a Claude conversation that never touches a tool, the Gemini app answering from its own knowledge. Those interactions still mention brands and cite ideas, and the training-corpus path is the only lever that moves them. The disambiguation page at /en/ai-search/geo-vs-aeo-vs-llm-seo/ treats the overlap in more depth.

Retrieval is when the model fetches your page live during the chat session. The model reads the rendered page, extracts spans, and cites them in the response. ChatGPT with search, Perplexity, Claude with web search, Gemini Deep Research, and AI Overviews all use retrieval. Training-corpus recall is when the model answers from its own weights without a live fetch. The model 'remembers' that your brand is associated with a topic, an approach, or a fact, because that association was present often enough in the training data. Regular ChatGPT, Claude without search, and the Gemini app without tool use all rely on this path. The two paths reward different work: retrieval rewards on-page structure and extractable answers; training-corpus recall rewards scale, consistency, and brand co-occurrence on trusted domains across the public web.

Not directly, but the levers are knowable. Models pretrain on a snapshot of the open web, weighted toward trusted domains and recent crawls. To be present in the snapshot, your content needs to be reachable (no JavaScript-only rendering of the answer, no blocking of GPTBot/CCBot/Google-Extended unless you intend to), repeatedly mentioned in the same context across many pages, and consistent in how it names entities. The Brand-Authority play is to earn mentions on Wikipedia, primary research outlets, journals, podcasts with transcripts, and forums whose archives are crawled. The Brand-Consistency play is to use the same canonical names and entity references everywhere your content appears, so the model encodes a clean association rather than a fuzzy cluster. Neither is fast, both compound.

Google did not use the phrase 'LLM SEO' specifically. On 15 May 2026, Google Search Central published its first official guide on optimising for generative AI features in Google Search, framing AEO and GEO as 'still SEO' from its perspective. The guide explicitly rules out llms.txt, content chunking, AI-specific schema, and AI rewrites as Google-specific levers, while keeping people-first content, semantic HTML, and high-quality media as the foundation. The guide is silent on the non-Google LLMs (ChatGPT, Claude, Gemini app, Perplexity), which use different retrieval and extraction stacks. The practical reading for LLM SEO: on Google's AI surfaces, the work is foundational SEO; on the other LLM surfaces, the LLM-SEO playbook applies per-engine.

By default, yes. Blocking GPTBot via robots.txt removes you from the snapshot OpenAI uses to train future models and, for some product configurations, from the live retrieval pool too. Blocking ClaudeBot and Google-Extended has comparable effects for Anthropic and Google. The case for blocking is narrow: high-value proprietary content where you would rather not be paraphrased without attribution, sites under explicit publisher agreements that compensate for the training use, or jurisdictions where the legal position requires it. For the typical content site, the volume of citation-shaped traffic from these engines exceeds the marginal value of training-data exclusion, so the default is to allow.

llms.txt is a short, structured Markdown file at your site root that lists your highest-value pages with one-line descriptions, grouped by topic. The proposal was introduced in 2024 and adoption is partial. Some retrieval engines and agent products read it as a discovery signal, treating it analogously to a structured sitemap focused on the LLM's reading pattern. Google's May 2026 guide is explicit that llms.txt is not used by Google's AI features. CTAIO Labs ran a 30-day field test across three sites and measured small but real per-engine deltas on ChatGPT and Perplexity in 'llms.txt: 30-Day Citation Experiment'. The current recommendation: ship llms.txt for the engines that read it (low effort, modest upside), and treat the Google line as authoritative for Google's surfaces.

Two stacks. Dedicated LLM visibility platforms (Profound, Peec AI, Otterly, AthenaHQ, Scrunch, Evertune, Bluefish, Goodie AI, Rankscale, Semji) score a fixed query set across the engines that matter weekly, reporting how often your pages appear as cited sources. CTAIO Labs benchmarked ten of them in '10 LLM Visibility Tools on 3 Real Brands'; the scored shortlist of the six that earned a recommendation is at /en/radar/geo-tools/. The second stack is GA4 referral tracking for chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, and copilot.microsoft.com — coverage is partial because not every agent passes a referrer, but the trend is informative. For the training-corpus path specifically, the closest measurable proxy is unbranded queries that surface your brand inside a chat answer with no search invoked; the LLM visibility tools cover this when their query set is configured to suppress the search tool.

It depends on the path. Retrieval-path results show up in days to weeks: ChatGPT and Perplexity reindex their cited-source pool roughly weekly, llms.txt updates can be picked up within days, and a new high-quality answer block on an already-trusted page can earn citations on the next reindex. Training-corpus results show up across model generations, which means months to years. The current generation of frontier models was trained on data through 2024 or early 2025, depending on the model; the next generation will use later 2025 and 2026 data. Content published today aimed at training-corpus recall is investing in the next pretraining run, not this week's citations.

Yes, with proportionality. The share of search-shaped traffic that already runs through some form of LLM is large enough in 2026 that ignoring LLM SEO leaves measurable revenue on the table. AI Overviews now appear on a meaningful slice of informational queries on Google (third-party tracker estimates range from roughly 13% to over 30% depending on vertical and methodology); ChatGPT and Perplexity together send referral traffic to most B2B and high-consideration B2C verticals. The play is to layer LLM SEO onto an already-functioning SEO programme rather than treating it as a separate stack. The on-page work overlaps substantially with AEO and GEO, so the marginal investment is modest if the SEO foundation is in place.

LLM SEO is the practice of getting your content cited, paraphrased, or recalled inside the output of a large language model. The output can be a search-attached answer (ChatGPT with search, Perplexity, AI Overviews), a pure chat response with no web search at all (regular ChatGPT, Claude, the Gemini app), an agent's tool call, or a voice assistant reading from an LLM. LLM SEO is the umbrella label that covers all of those surfaces, and it recognises two distinct paths to citation: retrieval, where the model fetches your page live, and training-corpus recall, where the model 'remembers' your content from pretraining.

GEO targets the synthesised generative answer specifically: AI Overviews, ChatGPT with search, Perplexity. AEO targets any surface where the answer is lifted into place (featured snippets, People Also Ask, voice answers, AI Overviews, chat answer cards). LLM SEO is broader than both. It includes the cases where there is no live web fetch at all — a user asking ChatGPT a question without enabling search, a Claude conversation that never touches a tool, the Gemini app answering from its own knowledge. Those interactions still mention brands and cite ideas, and the training-corpus path is the only lever that moves them. The disambiguation page at /en/ai-search/geo-vs-aeo-vs-llm-seo/ treats the overlap in more depth.

Retrieval is when the model fetches your page live during the chat session. The model reads the rendered page, extracts spans, and cites them in the response. ChatGPT with search, Perplexity, Claude with web search, Gemini Deep Research, and AI Overviews all use retrieval. Training-corpus recall is when the model answers from its own weights without a live fetch. The model 'remembers' that your brand is associated with a topic, an approach, or a fact, because that association was present often enough in the training data. Regular ChatGPT, Claude without search, and the Gemini app without tool use all rely on this path. The two paths reward different work: retrieval rewards on-page structure and extractable answers; training-corpus recall rewards scale, consistency, and brand co-occurrence on trusted domains across the public web.

Not directly, but the levers are knowable. Models pretrain on a snapshot of the open web, weighted toward trusted domains and recent crawls. To be present in the snapshot, your content needs to be reachable (no JavaScript-only rendering of the answer, no blocking of GPTBot/CCBot/Google-Extended unless you intend to), repeatedly mentioned in the same context across many pages, and consistent in how it names entities. The Brand-Authority play is to earn mentions on Wikipedia, primary research outlets, journals, podcasts with transcripts, and forums whose archives are crawled. The Brand-Consistency play is to use the same canonical names and entity references everywhere your content appears, so the model encodes a clean association rather than a fuzzy cluster. Neither is fast, both compound.

Google did not use the phrase 'LLM SEO' specifically. On 15 May 2026, Google Search Central published its first official guide on optimising for generative AI features in Google Search, framing AEO and GEO as 'still SEO' from its perspective. The guide explicitly rules out llms.txt, content chunking, AI-specific schema, and AI rewrites as Google-specific levers, while keeping people-first content, semantic HTML, and high-quality media as the foundation. The guide is silent on the non-Google LLMs (ChatGPT, Claude, Gemini app, Perplexity), which use different retrieval and extraction stacks. The practical reading for LLM SEO: on Google's AI surfaces, the work is foundational SEO; on the other LLM surfaces, the LLM-SEO playbook applies per-engine.

By default, yes. Blocking GPTBot via robots.txt removes you from the snapshot OpenAI uses to train future models and, for some product configurations, from the live retrieval pool too. Blocking ClaudeBot and Google-Extended has comparable effects for Anthropic and Google. The case for blocking is narrow: high-value proprietary content where you would rather not be paraphrased without attribution, sites under explicit publisher agreements that compensate for the training use, or jurisdictions where the legal position requires it. For the typical content site, the volume of citation-shaped traffic from these engines exceeds the marginal value of training-data exclusion, so the default is to allow.

llms.txt is a short, structured Markdown file at your site root that lists your highest-value pages with one-line descriptions, grouped by topic. The proposal was introduced in 2024 and adoption is partial. Some retrieval engines and agent products read it as a discovery signal, treating it analogously to a structured sitemap focused on the LLM's reading pattern. Google's May 2026 guide is explicit that llms.txt is not used by Google's AI features. CTAIO Labs ran a 30-day field test across three sites and measured small but real per-engine deltas on ChatGPT and Perplexity in 'llms.txt: 30-Day Citation Experiment'. The current recommendation: ship llms.txt for the engines that read it (low effort, modest upside), and treat the Google line as authoritative for Google's surfaces.

Two stacks. Dedicated LLM visibility platforms (Profound, Peec AI, Otterly, AthenaHQ, Scrunch, Evertune, Bluefish, Goodie AI, Rankscale, Semji) score a fixed query set across the engines that matter weekly, reporting how often your pages appear as cited sources. CTAIO Labs benchmarked ten of them in '10 LLM Visibility Tools on 3 Real Brands'; the scored shortlist of the six that earned a recommendation is at /en/radar/geo-tools/. The second stack is GA4 referral tracking for chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, and copilot.microsoft.com — coverage is partial because not every agent passes a referrer, but the trend is informative. For the training-corpus path specifically, the closest measurable proxy is unbranded queries that surface your brand inside a chat answer with no search invoked; the LLM visibility tools cover this when their query set is configured to suppress the search tool.

It depends on the path. Retrieval-path results show up in days to weeks: ChatGPT and Perplexity reindex their cited-source pool roughly weekly, llms.txt updates can be picked up within days, and a new high-quality answer block on an already-trusted page can earn citations on the next reindex. Training-corpus results show up across model generations, which means months to years. The current generation of frontier models was trained on data through 2024 or early 2025, depending on the model; the next generation will use later 2025 and 2026 data. Content published today aimed at training-corpus recall is investing in the next pretraining run, not this week's citations.

Yes, with proportionality. The share of search-shaped traffic that already runs through some form of LLM is large enough in 2026 that ignoring LLM SEO leaves measurable revenue on the table. AI Overviews now appear on a meaningful slice of informational queries on Google (third-party tracker estimates range from roughly 13% to over 30% depending on vertical and methodology); ChatGPT and Perplexity together send referral traffic to most B2B and high-consideration B2C verticals. The play is to layer LLM SEO onto an already-functioning SEO programme rather than treating it as a separate stack. The on-page work overlaps substantially with AEO and GEO, so the marginal investment is modest if the SEO foundation is in place.