April 23, 20265 min read

How Generative Engine Optimization Works: Why LLMs Don't Cite What Google Indexes

An LLM never searches your query as-is. It decomposes, reads selectively, extracts in chunks. Understanding that logic changes how you structure a site for AI.

Go from theory to practice

74 structured lessons across 6 parts, from technical foundations to long-term practice.

Access the training

Key Takeaways

  • An LLM never answers the exact query it receives. It breaks it down into several sub-queries (fan-out queries) before reading the web.
  • The LLM doesn't read your page as a whole. It chunks it into pieces that get scored independently. Many Google-ranked pages fail to produce a single citable chunk.
  • Six families of factors decide a citation: topical authority, brand-topic-proof associations, freshness, structure, third-party sources, brand signals.
  • According to Princeton's GEO study, optimized pages gain 30 to 40% more AI visibility than unoptimized content, even at the same SEO rank.
  • SEO remains the foundation. GEO adds a layer of optimization for extractability and perceived credibility by the LLM.

Part of our Generative Engine Optimization complete guide.

Your site ranks well on Google. You type your main query into ChatGPT. Your brand is nowhere. Instead: three competitors, and a blog post you've never seen.

That's not a bug. That's the mechanism. An LLM doesn't search the same way a classic search engine does, and it never searches your query as-is. Understanding the logic, even without diving into every step, is enough to see why Google rank is no longer a reliable predictor of AI visibility.

An LLM reads the web in pieces, not pages

A RAG LLM (Retrieval-Augmented Generation) like ChatGPT Search, Perplexity, Gemini or Google AI Overview always follows the same general logic: it reformulates the user's query into multiple angles, reads a selection of pages on the web, slices what it reads, keeps the best passages, and assembles an answer.

Two structural differences with Google change everything.

One: the query is never run as-is. The LLM generates several sub-queries to cover the implicit angles of the initial question. These are fan-out queries. If you optimize only for your main keyword, you miss most of what the LLM actually searches.

Two: the page isn't the unit of evaluation. The LLM chunks your content based on its structure. Every chunk is scored in isolation on its ability to answer. A page that reads well for a human can produce zero citable chunks if it isn't structured for extraction.

Fan-out queries: the concept that flips the table

If you remember one idea from GEO, make it this one. When a user asks "the best CRM for a freelancer," the LLM doesn't search that question. It searches several versions of that question, across multiple angles, to cover what the user actually wants to know: invoicing, project tracking, pricing, alternatives, complexity level.

Each angle is a distinct sub-query. Each sub-query filters a different slice of pages. The final answer aggregates several sources, each selected on a different sub-query.

Direct consequence: a site that covers a topic broadly, across multiple angles, beats a site with one perfect page on the main keyword. That's a full inversion of the "one page = one keyword" logic of classic SEO.

Perplexity is the only mainstream LLM that exposes fan-out queries in its UI. ChatGPT used to show them publicly until GPT-5.3, then OpenAI tucked them behind the API. Monitoring tools like Mentionable hit the API on your behalf to retrieve them on your target prompts.

Chunking: why structure beats content

The LLM doesn't evaluate your page as a whole. It slices it into chunks via your headings, paragraphs, and lists. Each piece is then scored independently on its ability to answer a specific fan-out query.

Dense content, with no clean hierarchy and ambiguous pronouns, produces unusable chunks. The LLM can't isolate a passage that says "this service also offers..." without knowing what "this service" refers to. It skips to a competitor whose chunk explicitly names the brand.

Content structure is therefore a GEO lever with immediate impact, fully under your control. It's one of the reasons two sites with comparable content at identical SEO rank can have radically different AI visibility.

The 6 families of citation factors

Once a chunk is extracted, the LLM ranks it on six dimensions. Named:

  1. Topical authority of your site as a whole
  2. Brand-topic-proof associations (is your brand linked to your topic through verifiable facts)
  3. Freshness of the content and its last update
  4. Structure of the content and its readability by the LLM
  5. Third-party sources mentioning you elsewhere on the web
  6. Brand signals aggregated across the web

Some factors play out within a week (structure, freshness). Others build over three to six months (topical authority, associations). Others still compound over the long term (third-party sources, brand signals).

Prioritization, the method for acting on each, and the order to tackle them when starting from zero are among the most covered topics in a proper GEO training. That's the transition from theoretical understanding to an executable plan.

What it actually changes for you

Three major implications, without going into the how.

One: Google rank is no longer a reliable predictor of AI visibility. You can be #1 on your main keyword and invisible on ChatGPT, Perplexity and Gemini. The inverse is also true.

Two: covering a topic broadly beats covering one keyword perfectly. Content architecture becomes at least as important as individual article quality.

Three: your page structure is likely your largest unused lever. Most B2B sites in 2026 have good content poorly structured for chunking. Fixing it moves AI visibility within weeks.

Understanding the mechanism is a start. Translating it into a plan that holds together (audit, restructuring, measurement) is another matter. For the full path, the GEO training walks the 74 lessons in the right implementation order.

Further reading

Frequently Asked Questions

How does an LLM like ChatGPT decide which source to cite?
The LLM breaks the user's query into several sub-queries (fan-out queries), pulls SERP results for each, chunks those pages using their structure, scores those chunks on relevance, credibility and freshness, then assembles the winning chunks into a synthesized answer with citations. Only chunks that survive scoring show up in the final answer.
What is a fan-out query?
A fan-out query is a sub-query the LLM generates automatically from the user's initial question. For a single question like 'best CRM for freelancers', the LLM can spin up several angles (invoicing, project tracking, pricing, commitment-free options). Each sub-query runs separately, and their results are aggregated into the final answer.
Why isn't my well-ranked Google page getting cited by ChatGPT?
Three common reasons. One: your page ranks on your main keyword but not on the fan-out queries the LLM actually runs. Two: your content isn't structured for extraction, so no chunk is self-contained enough to be cited. Three: your topical authority and brand signals are weak, and the LLM prefers sources it recognizes as references in your niche.
Does classic SEO still matter for LLM visibility?
Yes, SEO is the foundation. RAG LLMs rely on Google and Bing results to decide which pages to crawl. Ranking well on classic SEO increases the probability that the LLM reads you. But it's no longer enough: GEO adds a layer of optimization for extractability and perceived credibility by the LLM.
How many pages does an LLM read before answering?
A RAG LLM has a limited reading budget per query. It retrieves a large number of SERP results but only opens a handful of pages in detail, selected based on title and meta description quality. A strong title and relevant meta description directly increase your odds of being read, not just listed as a source.
Which factors weigh the most on a citation?
Six families: topical authority, brand-topic-proof associations, freshness, content structure, third-party sources, brand signals. Some play out in the short term and are fully under your control (structure, freshness). Others build over the long term (topical authority, third-party sources). The hierarchy and the how-to sit among the most covered topics in any serious GEO training.
How do I see which fan-out queries an LLM runs for my topic?
Perplexity exposes the sources and the intermediate searches it runs in its interface. Type a query related to your space and watch the listed sources. ChatGPT hid these behind its API since GPT-5.3. Monitoring tools like Mentionable hit the LLM APIs directly to retrieve those sub-queries and cross them with your content.
Alexandre Rastello
Alexandre Rastello
Founder & CEO, Mentionable

Alexandre is a fullstack developer with 5+ years building SaaS products. He created Mentionable after realizing no tool could answer a simple question: is AI recommending your brand, or your competitors'? He now helps solopreneurs and small businesses track their visibility across the major LLMs.

Published April 23, 2026

Apply GEO with a clear method

The Mentionable GEO training walks the 6 pillars in order: technical foundations, SEO, LLM mechanics, content strategy, tooling, practice management.