LLM search is the shift from ranked links to AI-synthesized answers, where a large language model reads your question, gathers information, and generates one response with citations instead of returning ten blue links. The mechanism underneath matters more than the output, because it explains why a website ranking number one on Google can be invisible inside an AI answer for the same query. Across 680 million citations analyzed in 2026, only 11% of domains were cited by both ChatGPT and Perplexity. The engines do not see the web the same way, and understanding why is the foundation of any LLM search strategy.

This guide explains how LLM search works at the mechanical level: what retrieval-augmented generation actually does, how each major AI search engine retrieves and cites sources differently, why traditional rankings do not transfer, and how the vocabulary (LLMO, LLM SEO, GEO, AEO) fits together. For the strategy and execution layer (how to structure content and earn citations), see RZLT's complete guide to answer engine optimization.

What is LLM search?

Traditional search built an index, matched your query against it, and returned a ranked list of pages for you to choose from. LLM search collapses that final step. Instead of handing you links, the model reads the relevant sources and writes the answer, citing the sources it used.

Most LLM search runs on Retrieval-Augmented Generation (RAG). The pattern has three stages. First, retrieval: the system takes your question, often reformulates it into one or more search-optimized versions, and fetches relevant content from a web index or live search. Second, augmentation: the retrieved content gets injected into the model's context as source material alongside your original question. Third, generation: the model writes an answer grounded in that retrieved content and attaches citations.

The consequence is that two scoring systems now sit between your content and the user. The retrieval stage decides whether your page is even pulled into the model's context. The generation stage decides whether, having been retrieved, your page actually gets cited in the answer. A page can clear the first and fail the second. This is the structural reason rankings and citations diverge: Google's ranking is one signal among many at the retrieval stage, and it has no say at all in the generation stage.

How AI search engines retrieve and cite differently

The single most important fact about LLM search is that the major AI search engines use different retrieval architectures, which is why their citations barely overlap. Optimizing for one does not transfer to the others.

ChatGPT draws on a real-time web index (Microsoft's Bing) when it browses, layered on top of its training data. A large share of its responses still answer from training data alone without any live search, which means for those queries no current content strategy can influence the result. When it does retrieve, it leans toward consensus and reference sources, with Wikipedia accounting for a notable share of its top citations. The practical read: entity recognition and presence in authoritative reference sources matter most here.

Perplexity is built around retrieval. Every query triggers a live web search, and the retrieval is the product rather than an add-on. Because it searches in real time, it weights current accuracy and freshness over accumulated historical authority, and it leans heavily on community sources. A well-structured, fact-dense page published this quarter can outperform an established brand's older content. For brands without decades of authority, Perplexity is often the most winnable platform first.

Claude uses a blend of training data and retrieval, with comparatively conservative citation habits and a preference for depth and well-structured content. It tends to reward technical precision and clearly organized sources over volume.

Google AI Overviews sit closest to traditional search, still pulling substantially from pages that rank organically, though that link is weakening fast: a year ago most AI Overview citations came from top-ten organic results, and by early 2026 that share had dropped sharply. Strong traditional SEO still helps most here, less so everywhere else.

The takeaway is not "optimize for each platform separately" so much as "stop treating AI visibility as one number." Measuring your ChatGPT presence tells you almost nothing about your Perplexity presence. Treating AI search as a single monolithic channel is like running one campaign across LinkedIn and TikTok and expecting matching results.

Why traditional rankings do not transfer to LLM search

A page earns a Google ranking through backlinks, on-page relevance, and domain authority accumulated over time. A page earns an LLM citation by being retrievable and then being the cleanest, most extractable, best-sourced answer to a specific question at the moment of generation. Those are different tests.

This is why a number-one Google result can be absent from an AI answer, and why a page that does not rank in the top twenty can be cited heavily. The retrieval stage may pull your page in on relevance, but if the relevant answer is buried in paragraph six instead of stated up front, the generation stage skips it for a competitor who answered cleanly. Rankings reward the page. Citations reward the segment. How AI is changing SEO comes down to that shift: the unit of optimization moved from the page to the extractable answer inside it.

LLMO, LLM SEO, GEO, AEO: the vocabulary

The terminology is still settling, and the labels overlap heavily. They describe the same underlying shift toward being cited inside AI answers, with differences mostly in emphasis.

LLM SEO and LLMO (LLM optimization) are the broad umbrella terms for optimizing visibility inside large language model outputs. AEO (answer engine optimization) emphasizes direct answers to specific questions. GEO (generative engine optimization) emphasizes longer, multi-source prompts where the AI synthesizes several references into a detailed response. In practice, the tactics converge: structured, extractable, well-sourced content wins across all of them.

Claiming the vocabulary early is itself a strategy. The category is forming now, and the brands that build citation authority before optimization becomes standard practice will hold a compounding advantage, the same way early SEO movers did two decades ago. The mechanism rewards genuine authority that cannot be gamed, because LLMs evaluate sources through several lenses at once. That makes the lead durable once earned.

Where to go next

Understanding how LLM search works is step one. Acting on it is step two. For the execution playbook (how to structure content for citation, how to measure share of voice, and how AEO relates to SEO), read RZLT's complete guide to answer engine optimization. To track where your brand currently shows up across these engines, see RZLT's guide to the top AEO tools for tracking AI search visibility. LLM search is the new front door. The brands that understand the mechanics behind it will be the ones the answer engines reach for.