LLM Citation Mechanics

Most discussions of 'LLM SEO' collapse citation into a single magical event. It is not. For RAG-based systems (which now includes ChatGPT search, Perplexity, Gemini, Claude with web access, and Google AI Overviews) citation is the output of three sequential steps.

Step 1: Retrieval

The system constructs a search query (sometimes the user's literal query, often a rewritten variant) and retrieves a candidate set — typically 5 to 20 URLs. If you are not in the candidate set, you cannot be cited. Retrieval is dominated by the same signals as classical web search: title tag, H1, body keyword density, link equity, and freshness.

Step 2: Ranking inside the candidate set

The model re-ranks the candidates based on relevance to the answer it is constructing. This step rewards source diversity (one Wikipedia + one industry source + one local source beats three industry sources), recency stamps, and apparent authority (well-known domains beat unknown ones, all else equal).

Step 3: Selection during generation

Even after ranking, the model picks which sources to actually cite based on which sentences it chose to include in the answer. This is where quotability dominates: a short, declarative sentence with a named subject is far more likely to make it into the final answer than a long, conditional one.

1. 1Put the entity name in the H1, not just the title tag. Retrieval over entity-keyword queries weights H1 more heavily than the title.
2. 2Include a 4-bullet TL;DR immediately after H1. Selection step preferentially extracts from this position.
3. 3Stamp the page with a visible 'Last updated' date in the body, not just in metadata. Recency ranking responds to in-body dates.
4. 4Use a dedicated Schema.org @type matching the content (Article, HowTo, FAQPage, MedicalWebPage, etc.). Generic WebPage schema underperforms in retrieval.
5. 5Author byline with a real name and link to a bio page. Selection step rewards perceived authorship.
6. 6Cite primary sources inline with anchor text that matches the source's brand. The model uses your outbound citations to evaluate your trustworthiness.
7. 7Avoid hedging adverbs ('typically,' 'usually,' 'often,' 'might') in the TL;DR. They get filtered out and reduce extractability.

• •Stuffing FAQ schema with answers that do not appear in the visible body. LLMs ignore (or worse, downgrade) the page.
• •Embedding 'as cited by ChatGPT' or similar self-promotional badges in the body. No measurable lift.
• •Aggressive cross-linking from every page to a target page using identical anchor. Helps classical rank, hurts perceived diversity in LLM ranking step.
• •Long single-paragraph 'Q&A' blocks with no headings. Retrieved poorly, selected almost never.
• •Adding a 'For LLMs' section at the bottom of pages. No measurable lift; arguably weakens substrate quality by adding redundant content.