LLM SEO: How Large Language Models Choose Sources

LLM SEO is the practice of making your pages the sources a large language model retrieves and cites when it answers a question. As of November 2026, most citations come from live retrieval: the assistant runs searches, fetches pages through a crawler or a search index, and quotes the passages that answer each sub-question most directly.

This guide belongs to the generative engine optimization (GEO) hub and covers one mechanism: how answer engines select sources, what the vendors confirm, and how to audit a site against that. The broader how-to lives in the guide to AI search engine optimization. Each claim is marked as officially confirmed, reported by trade media or research, or speculation.

What is LLM SEO, and what else is it called?

LLM SEO, short for large language model search engine optimization, is optimizing content so that AI assistants such as ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews use it as a source. The same work also goes by generative engine optimization (GEO), answer engine optimization (AEO), LLMO and AI SEO.

A large language model is a model trained on very large amounts of text to understand and generate language. On its own it answers from what it learned in training. The assistants people use for research add a search step on top of the model, and that search step is where SEO work has an effect.

Google takes a clear position on the label. Its guide to generative AI features (last updated July 10, 2026) says "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." The techniques overlap heavily with classic SEO; the goal and the measurement differ. The GEO vs SEO comparison covers the labels side by side, so the rest of this guide stays on how sources get chosen.

How do large language models choose which sources to cite?

Large language models choose sources in two ways: from memory, using patterns learned from training data, and from retrieval, by searching an index at answer time and grounding the reply in the fetched pages. Only retrieval produces a clickable citation, which is why LLM SEO targets the retrieval path first.

Google defines retrieval augmented generation as "a technique (also known as grounding) used to improve the quality, accuracy, and freshness of AI responses." The retrieval path has four documented stages:

• Access. The vendor's search crawler must be allowed to fetch the page.
• Index. The page must be in the index the assistant searches. Google's optimization guide says its AI features rely on "our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index."
• Fan-out queries. The assistant rarely searches the user's exact words. Google's AI features documentation says AI Overviews and AI Mode may use "query fan-out," "issuing multiple related searches across subtopics and data sources." Bing's AI Performance report lists these as "Grounding Queries."
• Passage and citation. The answer links the claim to a source. Google's October 15, 2026 announcement describes inline links as "links embedded in words or phrases in an AI response."

Speculation. No vendor documents how passages inside retrieved pages are scored. That self-contained answer paragraphs win citations is practitioner experience, not a published rule.

Which crawlers decide whether an LLM can cite your site?

The crawlers that decide whether an LLM can cite your site are the search and user crawlers, not the training crawlers. OAI-SearchBot, Claude-SearchBot and PerplexityBot fetch pages for search answers, while GPTBot and ClaudeBot collect training data, and each one is a separate robots.txt rule you set on its own.

Officially confirmed, per each vendor's documentation as of November 2026:

• OpenAI. The crawler overview says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features," GPTBot crawls content that may be used for training, and robots.txt changes take about 24 hours to apply. Robots.txt rules may not apply to ChatGPT-User, which runs when a person asks something.
• Anthropic. Its crawler article separates ClaudeBot (training), Claude-User (fetches pages for a user's question) and Claude-SearchBot (indexing for search). Blocking either of the last two may reduce visibility in Claude.
• Perplexity. Its crawler documentation says PerplexityBot surfaces and links websites and is not used to train foundation models; Perplexity-User generally ignores robots.txt.
• Google. Google-Extended controls Gemini training and grounding in Gemini Apps; per Google's crawler list it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal."

Announced, not live. On October 30, 2026, Google said it plans to support a shared AI training control in robots.txt, based on IETF work, from 2027.

How does LLM SEO differ from classic SEO ranking?

LLM SEO differs from classic SEO ranking in the unit that wins: classic search ranks whole pages for one query, while an LLM retrieves pages for several sub-queries and cites the passage that most directly supports one claim. Crawling, indexing and content quality stay the same; the scoring unit, the visible result and the measurement change.

Signal	Classic SEO	LLM source selection	Status
Entry requirement	Crawled and indexed	Fetched by the assistant's search crawler and indexed; snippet eligible for Google	Confirmed in vendor docs
Query matched	The query the user typed	Several fan-out sub-queries per question	Confirmed (Google, Bing)
Unit that wins	A whole page in a ranked list	A passage linked to a phrase or claim in the answer	Confirmed for inline links; passage scoring undocumented
Authority signal	Links and site reputation	Brand mentions track AI visibility more closely than backlinks	Reported (Ahrefs, 2025)
Content format	Relevance and helpfulness	Citations, quotes and statistics raised visibility up to 40% in a lab test	Research (GEO paper, 2024)
Result shown	Blue link with title and snippet	Citation, inline link or unlinked brand mention	Confirmed
Main metric	Rankings, clicks, CTR	AI impressions, citations, mentions	Confirmed reports exist

The research row comes from the GEO paper by Aggarwal and colleagues (KDD 2024). Lab results show direction, not how a live assistant ranks.

Does E-E-A-T matter for LLM SEO?

E-E-A-T matters for LLM SEO indirectly. Google says E-E-A-T is not a specific ranking factor, yet its AI features retrieve pages through the same core ranking systems that look for experience, expertise, authoritativeness and trust. Outside Google, no assistant documents an E-E-A-T style signal, so treat it as a quality bar rather than a lever.

E-E-A-T stands for experience, expertise, authoritativeness and trustworthiness. Google's helpful content guidance (last updated October 5, 2026) states that "E-E-A-T itself isn't a specific ranking factor" and that "trust is most important." It asks publishers to make clear who wrote a page, how it was made and why it exists. Its AI features guide adds a direct instruction: "Focus on developing unique, expert-led content that provides value beyond common knowledge." The Google AI content guidelines explain how quality raters apply that bar.

Outside signals matter too. The Ahrefs study of 75,000 brands (May 2025) found Spearman correlations with AI Overview brand visibility of 0.664 for branded web mentions, 0.527 for branded anchors, 0.392 for branded search volume, 0.326 for Domain Rating and 0.218 for backlink count. Ahrefs stresses these are correlations, not causes, and about 26% of the brands in its 2025 sample had no AI Overview mentions at all.

How do you run an LLM SEO audit?

An LLM SEO audit checks, for each priority question, the four gates a page must pass before an assistant can cite it: crawler access, index presence, passage fit and outside corroboration. The Source Gate Audit fixes the first gate that fails, because a better paragraph cannot help a page that is never retrieved.

Gate	Test	Fail signal	First fix
1. Access	robots.txt, CDN and firewall rules for each search crawler	Search crawler disallowed or blocked	Allow search crawlers; decide on training crawlers separately
2. Index	URL Inspection in Search Console and Bing Webmaster Tools	Not indexed, noindex, nosnippet or low max-snippet	Fix indexing and snippet controls
3. Passage	Read the section meant to answer the question	Answer buried or spread over sections	One self-contained answer under a clear heading
4. Corroboration	Search the brand and topic on other sites	No independent mentions for the topic	Earn genuine mentions with data others cite

What should an LLM SEO strategy measure?

An LLM SEO strategy should measure visibility inside AI answers first and treat AI referral clicks as a diagnostic, because the first-party reports count impressions and citations rather than visits. Google, Bing and OpenAI each expose one piece: AI impressions, AI citations with grounding queries, and tagged ChatGPT referrals.

Officially confirmed. Google's Search Console Generative AI performance reports, introduced June 3, 2026 and live for all websites since August 31, 2026, show impressions, pages, countries, devices and dates. Bing's AI Performance report (February 2026) shows Total Citations, Average Cited Pages, Grounding Queries and page-level citation activity across Copilot and Bing AI summaries. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral URLs; the guide to ChatGPT referral traffic covers why some of it still lands as direct.

Reported. Search Engine Journal (November 6, 2026) summarized October 2026 data from Marie Haynes across 100 sites, where ChatGPT sent 0.62% of Google organic's traffic volume. The same article quotes Kevin Indig on his search for AI influence in Direct traffic: "An effect could still exist. I just couldn't isolate it."

For prompt based tracking at scale, the guide to AI visibility tools compares the options. Run the same question set every month, because answers change from run to run.

Common mistakes in LLM SEO

The common mistakes in LLM SEO come from mixing up training and retrieval, or from measuring with one screenshot. Most of them block the access or index gate without anyone noticing, or produce a visibility number nobody can repeat, so the fixes are simple once the mechanism is clear.

• Blocking search crawlers while meaning to block training. A blanket bot rule can stop OAI-SearchBot or Claude-SearchBot. Set training and search crawlers separately.
• Adding llms.txt for Google. Google writes that Google Search does not use these files. Spend the time on indexing and content.
• Chunking pages into fragments. Google's optimization guide says "there's no requirement to break your content into tiny pieces." Answer first under each heading and keep the page whole.
• Treating one answer as evidence. AMEC, quoted by Search Engine Journal, warns that "evaluations performed once and stored on a shelf are decorative, not protective." Repeat a fixed question set.
• Buying ads to get cited. OpenAI stated in January 2026 that "ads do not influence the answers ChatGPT gives you"; ChatGPT ads are a separate placement.
• Blocking AI crawlers by IP. Anthropic advises against it, because the bot may no longer read robots.txt at all.
• Chasing backlinks alone. The Ahrefs correlations put brand mentions ahead of link counts, so earn mentions on the sites assistants retrieve.