LLM SEO: How Large Language Models Actually Pick the Sources They Cite

Every article on "LLM SEO" tells you the same tactical playbook — answer-first structure, schema markup, entity consistency, cluster content. The tactics work. But most articles omit why they work, and that omission is expensive. Without the underlying model, you cannot decide which of five conflicting tactics to prioritise when they disagree, cannot judge whether a new platform (say, a hypothetical enterprise AI search released next quarter) requires new tactics or the same ones, and cannot tell whether an agency selling you "a fully custom LLM SEO framework" is doing anything meaningful or repackaging the basics.

The underlying model is simple, and it's shared by every major AI search product on the market — ChatGPT Search, Perplexity, Google AI Overviews and AI Mode, Gemini, Claude, and Copilot. They all use retrieval-augmented generation, or RAG. The differences between them are tuning parameters on the same base pipeline. Once you understand the pipeline, tactical decisions become obvious — and you stop needing to relearn AEO every time a new platform launches.

Retrieval-augmented generation is the technique that lets a language model answer with fresh, cited facts instead of hallucinating from its training data. It runs in five distinct stages, and each stage is a filter you can pass or fail.

Step 1 — Chunking. Somewhere before your page ever enters an AI search product, it has been crawled by a web-scale spider (BingBot for ChatGPT/Copilot, PerplexityBot for Perplexity, Googlebot for Google AI Overviews, and so on). The HTML gets stripped, structured content is extracted, and the page is split into chunks — usually 200-500 tokens (roughly 150-380 words) each. Different vendors chunk differently: some by heading, some by fixed token count, some by semantic paragraph. What matters for you: the chunk boundary decides what gets embedded as a unit. A 900-word section with no subheadings will be arbitrarily split — and the split may cut across your answer.

Step 2 — Embedding. Each chunk is passed through an embedding model, which converts the text into a 768- or 1536-dimension vector. Semantically similar chunks land near each other in this vector space — a chunk about "customer acquisition cost" ends up near "CAC" and "cost per new customer" even without shared keywords. Every chunk in the vendor's index is embedded and stored, indexed for fast nearest-neighbour search.

Step 3 — Retrieval. When a user query arrives, the query is embedded using the same model, and the vector store returns the top-k most similar chunks. K is usually 20-100. Modern production RAG systems run this in parallel with BM25 (keyword) search and fuse the results — a pattern called hybrid retrieval — because pure vector search misses exact-match queries and pure keyword search misses semantic matches. Reciprocal Rank Fusion is the most common fusion method.

Step 4 — Reranking. The top-k candidates from retrieval are passed through a cross-encoder reranker — a more expensive model that scores each candidate against the query with full attention. This is where a lot of the ranking work actually happens. Reranking catches relevance that embedding similarity misses: a chunk that mentions the query terms but doesn't answer the question gets down-ranked; a chunk that directly answers gets promoted. The output is a shortlist of maybe 5-15 chunks that will be shown to the language model.

Step 5 — Generation and citation. The shortlisted chunks are injected into the language model's context along with the query, and the model writes the answer, citing sources it used. This is where the last drop-off happens: research on ChatGPT found only about 15% of retrieved pages actually appear as citations in the final answer. The model reads them all, but attributes only what it actually pulled from.

Because the pipeline has five stages, there are five distinct failure modes — and five distinct optimisations. Most AEO advice conflates them.

Fail at chunking = your important content is split badly. Fix: use clear H2 and H3 headings every 250-350 words. Never bury a critical statistic in the middle of an 800-word section — chunkers may cut it away from its context. If a chunk cannot stand alone as a coherent answer, it won't get cited alone.

Fail at embedding = your content is semantically ambiguous or entity-fuzzy. Fix: use precise, unambiguous language for entities. "CRO" alone is ambiguous (Conversion Rate Optimisation? Chief Revenue Officer?). Spell it out at least once per section. Use canonical entity names ("Google Analytics 4", not "GA"). This tightens the embedding representation and makes your chunk more likely to be a nearest neighbour to the query.

Fail at retrieval = your page is not in the vendor's index, or your chunks are not near the query in vector space. Fix: get indexed by every major crawler (BingBot, PerplexityBot, Googlebot). Cover query intent explicitly — if users ask "how much does X cost", the phrase "how much does X cost" should literally appear on your page in an H2 or opening sentence. This helps both BM25 (keyword) retrieval and semantic retrieval.

Fail at reranking = you retrieved but got down-ranked. This is where content quality separates. The reranker asks: does this chunk actually answer the query? If your "answer" is 300 words of preamble followed by 30 words of substance, the reranker will down-rank you in favour of a chunk that answers in the first sentence. This is the mechanical reason "answer-first structure" is repeated in every AEO guide — it's the single biggest lever for reranker score.

Fail at citation = you survived to the shortlist but the model didn't pull from you. This is often about authority and specificity. If two chunks say the same thing and one is from a domain the model recognises as an authority on the topic (Wikipedia, the primary research source, a domain the entity is known for), the model attributes to the authoritative one. Being the primary source — not the aggregator — is what wins here.

All the major AI search products use the same base architecture. What differs is the tuning — which index they retrieve from, how many sources they cite, which authority signals they weight, and how they handle freshness.

• ChatGPT Search. Retrieves via Bing. Cites an average of ~15 sources per response but often shows only 4-8 inline. Weights Reddit (roughly 40% of LLM references across all products per Semrush's 150,000-citation study) and Wikipedia heavily. Freshness-sensitive for news queries; less so for evergreen how-to content.
• Perplexity. Uses its own proprietary crawler (PerplexityBot) plus partnerships. Shows citations more prominently than any other product — the UI is source-first. Cites Reddit less than ChatGPT, weights news publishers more heavily. Freshness-sensitive; content older than 12-18 months on trending topics gets discounted.
• Google AI Overviews and AI Mode. Retrieves via Google's own index — meaning traditional Google SEO signals (backlinks, E-E-A-T, Core Web Vitals) all still apply. AI Overviews now trigger on approximately 48% of tracked search queries globally and 26.8% in India as of early 2026. Cites 3-5 sources typically. Weights YMYL (Your Money Your Life) authority the hardest.
• Gemini. Cites an average of 3 sources per response — the smallest citation footprint of the major products. Draws from a smaller pool that includes Wikipedia, Reddit, and YouTube. If you rank in Gemini for a query, you often own most of the answer.
• Claude. Uses Brave Search when web access is enabled. Newer to public web citation than the others; the citation pool is smaller and more news-publisher-weighted.

The strategic implication: pick the two or three platforms your buyers actually use, get your pages indexed by their crawlers, and optimise your content structure so it survives their rerankers. The tactics converge more than they diverge — because the mechanism is the same.

If the RAG pipeline is the substrate, three content properties determine whether you survive it well.

Structure — the chunkability of your page. A page written as one flowing narrative will be chunked arbitrarily. A page written with clear H2s every 250-400 words, an H3 hierarchy under them, and an answer-first opening to each section will chunk cleanly into semantically coherent units. Each unit becomes an independently retrievable answer. Two pages of equal quality but different structure will have very different citation rates — the well-chunked one wins.

Entity clarity — how unambiguously your content resolves the entities it mentions. If you write "we helped a fintech achieve 3x growth", nothing about that chunk is retrievable for anyone searching for anything specific. If you write "we helped Razorpay's SME segment grow onboarding conversion by 3x through a 12-week CRO engagement", every noun is a retrievable entity. Named entity density lifts embedding precision, which lifts retrieval rank, which lifts citation probability.

Semantic density — how much distinct information per paragraph. A paragraph that says three concrete things beats a paragraph that says one thing three ways. Rerankers reward information density because the model is looking for chunks that give the most answer in the smallest window. Purple-prose intros, hedge phrases, and repetitive framing get down-ranked. The instinct many writers have — build up before making the point — is the exact opposite of what wins AEO.

1. From keyword density to entity density. Old SEO counted target keyword occurrences. LLM SEO counts named entities — people, companies, products, standards, benchmarks — because entities are what embedding models resolve. Use canonical spellings, first-mention full names, and consistent naming across your site.

2. From long meandering intros to answer-first paragraphs. The first 40-60 words after any H2 are usually what gets embedded as the primary answer chunk for that section. Every H2 should have a direct answer in that window, followed by supporting detail.

3. From single-topic pages to hub-and-spoke clusters. Topical breadth signals authority to embedding-based systems even more than to Google. A hub page linking to 4-6 tactical deep-dives on the same topic cluster gets cited more than a single comprehensive page of equal word count, because rerankers can retrieve the specific spoke that matches the specific query. This article is a spoke of our AEO strategic overview; the ChatGPT SEO deep-dive is another.

4. From backlink obsession to citation-source hygiene. Backlinks still matter for Google AI Overviews because Google's index still weights them. But for ChatGPT, Perplexity, Claude, and Gemini, being cited by the sources those models trust — Wikipedia, Reddit threads, industry newsletters, LinkedIn thought leadership — matters more than DoFollow backlink counts. This means digital PR shifts from "get a link" to "get quoted with your name attached".

5. From vanity-metric reporting to citation-first measurement. Rankings and clicks lag AI visibility by weeks or months. If you wait for referral traffic to prove AEO is working, you'll conclude it isn't. The leading indicator is citation rate — how often your brand or content appears in AI answers to your target prompts. See how to measure AEO — brand visibility in AI answers for the full measurement stack.

None of this replaces classical SEO. It layers on top. The tactics that made you rank in Google — clean site architecture, fast pages, honest schema, real E-E-A-T, quality writing — are still necessary. LLM SEO adds a layer of structural and semantic discipline on top: write so a chunker can slice you cleanly, embed you precisely, and rerank you highly.