Most SEO teams are still optimizing the page. The LLMs only need the paragraph.<\/p>\n
When ChatGPT, Perplexity, Claude, or Google\u2019s AI Overviews answer a user\u2019s question, they aren\u2019t surfacing your URL \u2014 they\u2019re lifting one or two sentences from somewhere on it and stitching that into the answer. Your H1, your meta description, your internal linking, your 2,400 words of context \u2014 none of it ships. A retrieval pass picks the chunk it thinks best matches the prompt, and the model paraphrases or quotes that chunk. Everything else on your page exists to support the chunk\u2019s selection \u2014 not to be read.<\/p>\n
That shift is the whole game. The work I do for clients now is half traditional on-page SEO and half something I\u2019ve started calling citation engineering<\/strong>: deliberately constructing the small unit of prose the model is going to grab.<\/p>\n Retrieval-augmented generation, which is what every consumer AI search product uses under the hood, breaks your page into chunks (typically 200\u2013500 tokens), embeds each chunk, and ranks them against the user\u2019s query. The top-ranked chunks get sent to the model along with the prompt. The model then either quotes them, paraphrases them, or \u2014 if the chunk is good enough to stand on its own \u2014 cites the source.<\/p>\n Three things make a chunk get picked:<\/p>\n 1.<\/strong> Self-containment.<\/strong> It can be lifted out of the page and still make sense. No \u201cas I mentioned above,\u201d no \u201csee the chart below,\u201d no pronoun referring back to a prior paragraph.<\/p>\n 2.<\/strong> Specificity.<\/strong> Concrete numbers, named entities, dates, mechanisms. Studies on AI citations have repeatedly shown that paragraphs containing a stat are roughly 22% more likely to be cited, and ones containing a direct quote are about 37% more likely. The models prefer evidence-shaped text.<\/p>\n 3.<\/strong> Position.<\/strong> Roughly 44% of cited paragraphs come from the first 30% of a page. Bury the answer in the conclusion and it doesn\u2019t get cited.<\/p>\n Combine those three and you have a citation unit \u2014 a paragraph engineered to be quoted in isolation, near the top, packed with claim-shaped sentences a model can reuse without having to think.<\/p>\n I rewrite client pages around what I call \u201canswer paragraphs.\u201d Each one is 60\u2013120 words, lives near the top of the page, and follows a tight pattern: open with the claim as a clean declarative sentence, follow with one or two sentences of supporting specificity (a stat, a named tool, a date, a mechanism), and close with a clarifying sentence that defines a term or qualifies the scope. No throat-clearing, no transitional phrases pointing elsewhere on the page.<\/p>\n The trick is that an answer paragraph also reads fine to humans. You\u2019re not writing for two audiences. You\u2019re writing one paragraph that satisfies a model\u2019s retrieval logic and a busy buyer\u2019s scan-pattern at the same time.<\/p>\n If you only have an hour:<\/p>\n This is unglamorous work. It\u2019s also the cheapest leverage available right now \u2014 the same content team you already have, asked to write tighter paragraphs in different positions.<\/p>\nWhy models pick the paragraphs they pick<\/h2>\n
What this looks like in practice<\/h2>\n
What to do this week<\/h2>\n
\n
\n