{"id":219,"date":"2026-07-11T21:55:02","date_gmt":"2026-07-11T12:55:02","guid":{"rendered":"https:\/\/www.studio-jt.co.kr\/seo-glossary\/rag\/"},"modified":"2026-11-15T18:46:50","modified_gmt":"2026-11-15T09:46:50","slug":"rag","status":"publish","type":"seo_glossary","link":"https:\/\/www.studio-jt.co.kr\/en\/seo-glossary\/rag\/","title":{"rendered":"RAG"},"content":{"rendered":"<h2 class=\"wp-block-heading\">The technology behind AI search answers<\/h2><p class=\"wp-block-paragraph\">RAG is a technique in which a language model first searches external documents before generating an answer, then writes the answer based on what it retrieved. Google Cloud&#8217;s Vertex AI RAG Engine documentation explains that this approach fills the model&#8217;s context with external information, reducing hallucinations and helping it answer questions more accurately. <a class=\"jt-src jt-src--official\" href=\"https:\/\/docs.cloud.google.com\/vertex-ai\/generative-ai\/docs\/rag-engine\/rag-overview\" target=\"_blank\" rel=\"noopener\">Google (official)<\/a><\/p><p class=\"wp-block-paragraph\">The concept was established in a 2020 paper by Patrick Lewis and colleagues at Facebook AI Research (now Meta AI). The paper reported that retrieving evidence from an external document index and combining it with the generation process produces more specific and factual answers than a purely generative model. <a class=\"jt-src jt-src--research\" href=\"https:\/\/arxiv.org\/abs\/2005.11401\" target=\"_blank\" rel=\"noopener\">RAG paper<\/a><\/p><p class=\"wp-block-paragraph\">Why RAG matters to website owners is clear: to be cited in AI search answers, your documents first need to be easy to retrieve and excerpt. Understanding this structure is the starting point of <a href=\"\/en\/seo-glossary\/geo\/\">GEO<\/a> (Generative Engine Optimization).<\/p><h2 class=\"wp-block-heading\">How it works: the steps Google describes<\/h2><p class=\"wp-block-paragraph\">Google Cloud&#8217;s Vertex AI RAG Engine documentation describes the process in the following steps. <a class=\"jt-src jt-src--official\" href=\"https:\/\/docs.cloud.google.com\/vertex-ai\/generative-ai\/docs\/rag-engine\/rag-overview\" target=\"_blank\" rel=\"noopener\">Google (official)<\/a><\/p><ol class=\"wp-block-list\"><li>Data ingestion &#8211; Collect documents from multiple sources.<\/li><li>Data transformation &#8211; Split documents into pieces called chunks so they can be indexed.<\/li><li>Embedding and indexing &#8211; Convert each chunk into a numeric representation of its meaning and build an index optimized for search.<\/li><li>Retrieval &#8211; Find the chunks most relevant to the question in the index.<\/li><li>Generation &#8211; The language model writes an answer based on the retrieved chunks.<\/li><\/ol><p class=\"wp-block-paragraph\">The key is the unit. RAG does not evaluate a page as a whole. It is less like reading an entire book and more like photocopying only the pages you need and putting them on your desk. How well a single extracted chunk matches the question decides whether it gets cited.<\/p><h2 class=\"wp-block-heading\">AI search services really do search the web<\/h2><ul class=\"wp-block-list\"><li>Google AI Overviews and AI Mode &#8211; Google Search Central documentation explains that they use a query fan-out technique that runs multiple searches on related subtopics at the same time, and that while an answer is being generated, they find supporting web pages and display them as links. <a class=\"jt-src jt-src--official\" href=\"https:\/\/developers.google.com\/search\/docs\/appearance\/ai-features\" target=\"_blank\" rel=\"noopener\">Google (official)<\/a><\/li><li>ChatGPT search &#8211; OpenAI&#8217;s crawler documentation distinguishes the training crawler (GPTBot) from the search crawler (OAI-SearchBot), and explains that OAI-SearchBot is used to surface websites in ChatGPT search results. <a class=\"jt-src jt-src--official\" href=\"https:\/\/developers.openai.com\/api\/docs\/bots\" target=\"_blank\" rel=\"noopener\">OpenAI (official)<\/a><\/li><li>Perplexity &#8211; Perplexity&#8217;s crawler documentation states that PerplexityBot is designed to surface and link websites in search results and is not used for model training. <a class=\"jt-src jt-src--official\" href=\"https:\/\/docs.perplexity.ai\/guides\/bots\" target=\"_blank\" rel=\"noopener\">Perplexity (official)<\/a><\/li><\/ul><h2 class=\"wp-block-heading\">Does ranking high mean being cited? Official guidance and data disagree<\/h2><p class=\"wp-block-paragraph\">The common belief that good standard SEO alone gets you into AI search is only half right. Google&#8217;s official guidance and large-scale measurements paint different pictures.<\/p><figure class=\"wp-block-table jt-verdict\"><table><tbody><tr><th>Google&#8217;s official position<\/th><td>Google says there are no extra requirements to appear in AI Overviews or AI Mode, and that following standard SEO best practices such as crawlability and indexability is enough. <a class=\"jt-src jt-src--official\" href=\"https:\/\/developers.google.com\/search\/docs\/appearance\/ai-features\" target=\"_blank\" rel=\"noopener\">Google (official)<\/a><\/td><\/tr><tr><th>Expert measurements<\/th><td>In an analysis of about 860,000 SERPs, only 38% of AI Overview citations came from top-10 pages, down to about half of the 76% seen a year earlier. Ranking does not equal citation. <a class=\"jt-src jt-src--data\" href=\"https:\/\/ahrefs.com\/blog\/ai-overview-citations-top-10\/\" target=\"_blank\" rel=\"noopener\">Ahrefs data<\/a><\/td><\/tr><\/tbody><\/table><\/figure><p class=\"wp-block-paragraph\">The two positions don&#8217;t directly contradict each other. Standard SEO is still the foundation of search and indexing. But reaching the top of the rankings doesn&#8217;t automatically bring citations, because Google uses query fan-out to also consult results for many sub-questions beyond the original one. In a GEO experiment published in 2024 by researchers at Princeton and IIT Delhi, the cite-sources technique raised the generative engine visibility of the site ranked 5th in the SERP by 115.1%, while lowering the No. 1 site&#8217;s visibility by 30.3% on average, leveling the field in favor of lower-ranked pages. <a class=\"jt-src jt-src--research\" href=\"https:\/\/arxiv.org\/abs\/2311.09735\" target=\"_blank\" rel=\"noopener\">GEO paper<\/a> In practice, it is safer to treat improving rankings and becoming easy to cite as two separate jobs.<\/p><h2 class=\"wp-block-heading\">Experimentally proven ways to increase citations<\/h2><p class=\"wp-block-paragraph\">The same GEO study tested nine techniques on GEO-bench, a benchmark of 10,000 queries, to separate methods that actually work from those that don&#8217;t. <a class=\"jt-src jt-src--research\" href=\"https:\/\/arxiv.org\/abs\/2311.09735\" target=\"_blank\" rel=\"noopener\">GEO paper<\/a><\/p><ul class=\"wp-block-list\"><li>Add statistics &#8211; Back your claims with specific numbers.<\/li><li>Add quotations &#8211; Quote relevant experts or sources directly.<\/li><li>Cite sources &#8211; Clearly state the sources behind your claims.<\/li><\/ul><p class=\"wp-block-paragraph\">These three techniques had the biggest effect; the top technique raised Position-Adjusted Word Count by 41% and the subjective impression score by 28% over the baseline. <a class=\"jt-src jt-src--research\" href=\"https:\/\/arxiv.org\/abs\/2311.09735\" target=\"_blank\" rel=\"noopener\">GEO paper<\/a> In contrast, keyword stuffing had no effect or even backfired. Overall visibility gains from the techniques reached up to 40%, but varied by domain and query type, so a field-specific approach is needed.<\/p><h2 class=\"wp-block-heading\">What it means for websites: content is read in chunks<\/h2><p class=\"wp-block-paragraph\">Search Engine Land&#8217;s content chunking guide explains that AI systems extract content by passage rather than by whole page, and that a clear heading hierarchy makes this extraction far more efficient. <a class=\"jt-src jt-src--data\" href=\"https:\/\/searchengineland.com\/guide\/content-chunking-seo\" target=\"_blank\" rel=\"noopener\">Search Engine Land<\/a><\/p><ul class=\"wp-block-list\"><li>Answer-first paragraphs &#8211; Don&#8217;t bury the answer inside the paragraph; give it in the first sentence. Wherever the chunk is cut, the key claim survives inside it.<\/li><li>Self-contained sections &#8211; Each chunk should make complete sense on its own. A piece that only makes sense with the surrounding paragraphs is hard to excerpt.<\/li><li>Clear heading structure &#8211; Use h2 for major topics and h3 as boundaries for subunits, so each chunk is labeled with its topic.<\/li><\/ul><pre class=\"wp-block-code\"><code>&lt;!-- A structure that stays complete even when cut into chunks --&gt;\n&lt;h2&gt;Refund policy&lt;\/h2&gt;\n&lt;p&gt;Custom-made products can be fully refunded only within 24 hours of payment.&lt;\/p&gt;\n\n&lt;!-- A structure to avoid - a design div instead of a heading; machines lose the boundary and topic --&gt;\n&lt;div class=\"section-title\"&gt;Refund policy&lt;\/div&gt;<\/code><\/pre><h2 class=\"wp-block-heading\">Practical checkpoints for building and running a site<\/h2><p class=\"wp-block-paragraph\">From the perspective of building and running a site, half of your RAG readiness is decided at the markup and rendering stage, before any content is written. The basis for each item is noted below.<\/p><ul class=\"wp-block-list\"><li>Serve complete HTML from the server &#8211; Search Engine Land&#8217;s AI crawler optimization guide explains that most <a href=\"\/en\/seo-glossary\/ai-crawler\/\">AI crawlers<\/a> don&#8217;t execute JavaScript, so content injected by JavaScript after load is effectively invisible to them. <a class=\"jt-src jt-src--data\" href=\"https:\/\/searchengineland.com\/guide\/optimize-for-ai-crawlers\" target=\"_blank\" rel=\"noopener\">Search Engine Land<\/a><\/li><li>Provide important content as text &#8211; Google Search Central documentation advises providing important content in text form and keeping standard basics such as <a href=\"\/en\/seo-glossary\/crawling-indexing\/\">crawlability and indexability<\/a> in order to appear in AI features. <a class=\"jt-src jt-src--official\" href=\"https:\/\/developers.google.com\/search\/docs\/appearance\/ai-features\" target=\"_blank\" rel=\"noopener\">Google (official)<\/a><\/li><li>Keep one topic per paragraph &#8211; The Search Engine Land guide sets one idea per paragraph as a principle and recommends keeping paragraphs short. <a class=\"jt-src jt-src--data\" href=\"https:\/\/searchengineland.com\/guide\/content-chunking-seo\" target=\"_blank\" rel=\"noopener\">Search Engine Land<\/a><\/li><li>Review which AI crawlers robots.txt allows &#8211; According to OpenAI&#8217;s documentation, the training crawler (GPTBot) and the search crawler (OAI-SearchBot) are controlled separately, and blocking OAI-SearchBot keeps your site out of ChatGPT search answers. <a class=\"jt-src jt-src--official\" href=\"https:\/\/developers.openai.com\/api\/docs\/bots\" target=\"_blank\" rel=\"noopener\">OpenAI (official)<\/a><\/li><li>Control exposure with existing controls &#8211; Google Search Central documentation explains that existing controls such as nosnippet, data-nosnippet, max-snippet, and noindex apply to AI Overviews and AI Mode as well. <a class=\"jt-src jt-src--official\" href=\"https:\/\/developers.google.com\/search\/docs\/appearance\/ai-features\" target=\"_blank\" rel=\"noopener\">Google (official)<\/a><\/li><\/ul><h2 class=\"wp-block-heading\">Common mistakes<\/h2><ul class=\"wp-block-list\"><li>Putting the conclusion at the end of the paragraph &#8211; Natural for human readers, but the opposite of the chunk-excerpting advice to put the answer first.<\/li><li>Making headings with divs and classes instead of heading tags &#8211; Both the chunk boundary and the topic label disappear.<\/li><li>Keeping key information only inside JavaScript tabs or accordions &#8211; Content that isn&#8217;t in the initial HTML might as well not exist for crawlers that don&#8217;t run JavaScript.<\/li><li>Blocking all AI crawlers at once and still expecting citations &#8211; Training crawlers and search crawlers are different, so decide your crawl policy first.<\/li><\/ul><p class=\"wp-block-paragraph\">In short, RAG answers the question of what unit AI uses to read your site. Remembering just one fact, that the unit is the chunk and not the page, changes how you structure paragraphs and design headings.<\/p><h2 class=\"wp-block-heading\">References<\/h2><ul class=\"wp-block-list\"><li><a href=\"https:\/\/developers.google.com\/search\/docs\/appearance\/ai-features\" target=\"_blank\" rel=\"noopener\">AI Features and Your Website &#8211; Google Search Central<\/a><\/li><li><a href=\"https:\/\/docs.cloud.google.com\/vertex-ai\/generative-ai\/docs\/rag-engine\/rag-overview\" target=\"_blank\" rel=\"noopener\">Vertex AI RAG Engine overview &#8211; Google Cloud<\/a><\/li><li><a href=\"https:\/\/developers.openai.com\/api\/docs\/bots\" target=\"_blank\" rel=\"noopener\">OpenAI crawlers overview (GPTBot, OAI-SearchBot, ChatGPT-User) &#8211; OpenAI<\/a><\/li><li><a href=\"https:\/\/docs.perplexity.ai\/guides\/bots\" target=\"_blank\" rel=\"noopener\">Perplexity Crawlers &#8211; Perplexity<\/a><\/li><li><a href=\"https:\/\/arxiv.org\/abs\/2005.11401\" target=\"_blank\" rel=\"noopener\">Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., 2020) &#8211; arXiv<\/a><\/li><li><a href=\"https:\/\/arxiv.org\/abs\/2311.09735\" target=\"_blank\" rel=\"noopener\">GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) &#8211; arXiv<\/a><\/li><li><a href=\"https:\/\/ahrefs.com\/blog\/ai-overview-citations-top-10\/\" target=\"_blank\" rel=\"noopener\">38% of AI Overview Citations Pull From The Top 10 &#8211; Ahrefs<\/a><\/li><li><a href=\"https:\/\/searchengineland.com\/guide\/content-chunking-seo\" target=\"_blank\" rel=\"noopener\">Use content chunking to structure information for better UX, rankings, and AI visibility &#8211; Search Engine Land<\/a><\/li><li><a href=\"https:\/\/searchengineland.com\/guide\/optimize-for-ai-crawlers\" target=\"_blank\" rel=\"noopener\">AI crawler optimization: How to optimize a website for AI crawlers &amp; AI agents &#8211; Search Engine Land<\/a><\/li><\/ul>","protected":false},"excerpt":{"rendered":"<p>RAG (Retrieval-Augmented Generation) is a technique in which AI searches external documents before writing an answer and then generates the answer based on what it retrieved. According to each company&#8217;s official documentation, Google AI Overviews, ChatGPT search, and Perplexity all search and excerpt web documents to build their answers. Because documents are excerpted in chunks, answer-first paragraphs and a clear heading structure are the starting point for earning AI citations.<\/p>\n","protected":false},"template":"","glossary_categories":[15],"class_list":["post-219","seo_glossary","type-seo_glossary","status-publish","hentry","glossary_categories-geo-ai"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/seo_glossary\/219","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/seo_glossary"}],"about":[{"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/types\/seo_glossary"}],"version-history":[{"count":1,"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/seo_glossary\/219\/revisions"}],"predecessor-version":[{"id":2304,"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/seo_glossary\/219\/revisions\/2304"}],"wp:attachment":[{"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/media?parent=219"}],"wp:term":[{"taxonomy":"glossary_categories","embeddable":true,"href":"https:\/\/www.studio-jt.co.kr\/en\/wp-json\/wp\/v2\/glossary_categories?post=219"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}