--- title: "LLM optimization" description: "LLM optimization from ShoutRank. Write and structure pages so a language model can lift the offer without inventing it." canonical: https://shoutrank.com/llm-optimization markdown: https://shoutrank.com/llm-optimization.md --- # LLM optimization LLM optimization from ShoutRank. Write and structure pages so a language model can lift the offer without inventing it. LLM optimization is how a page is written and structured so a model can quote the offer, the limit, and the proof. Models compress. A page that buries the answer under a brand story gets compressed into something you did not say. A page that states the audience, the job, and the limit in plain lines is easier to lift intact. This is on-page work with a specific reader in mind. It supports [ChatGPT](https://shoutrank.com/chatgpt-seo), [Gemini](https://shoutrank.com/gemini-seo), [Perplexity](https://shoutrank.com/perplexity-seo), and overviews. It does not replace [technical SEO](https://shoutrank.com/technical-seo). A beautiful passage in a document the crawler never receives does not matter. ShoutRank does not guarantee a ranking or a citation. ## The problem ### The differentiator is a metaphor. The homepage is a slogan. The model reaches for a competitor’s docs, where the feature is named in a sentence. ### Tables and facts are images. Pricing, specs, and comparisons live in pictures. The text alternative is empty. The model guesses. ### The page tries to instruct the model. Hidden lines tell a crawler what to say. Readers who view source, and the models that ignore the trick, are both worse off. ## How ShoutRank approaches it We pick the URLs that should be quotable and rewrite the extract: a definition, who it is for, who it is not for, and a fact a person can verify. Internal links point at that URL with words a model and a person both understand. We also check machine-readable copies where the site already publishes them, including Markdown twins on this site. A twin that contradicts the HTML is a bug. ## What the work includes ### Extract The lines that should survive compression, written on the page, not in a hidden field. ### Structure notes Headings, lists, and tables in HTML, with alt text where an image carries a fact. ### Contradiction list Places the site disagrees with itself, including old posts and profiles. ### Quotable proof A source, a date, or a method next to any claim you want repeated. ## Process ### Discovery Choose the URLs an answer should be able to cite. Leave the rest alone. ### Audit Read them as a model would: first screen, headings, lists, and missing text alternatives. ### Strategy Write the extract in the language of the buyer, including the limit. ### Implementation Publish it in the HTML. Remove hidden instructions if any were added earlier. ### Monitoring Re-fetch the URL and confirm the lines are in the document, not only in a screenshot. ### Optimization Tighten pages where answers still invent a feature. ### Reporting The extracts shipped, and any answer that still contradicts them. ## Why ShoutRank ### Visible text only If a person cannot see the line, we do not count it as optimization. ### Limits included A model that repeats an over-claim creates support debt. We write the boundary into the extract. ### Checked in the HTML We refetch the page. A CMS preview is not evidence. ## How success is measured Assigned URLs whose opening now states the offer and the limit. Answers that quote or contradict those lines. Facts that were trapped in images and are now in text. A named client result is published when that client clears the figure. Until then the report is the measurement. ## Questions ### Is LLM optimization schema markup? Schema can clarify a page. It does not supply a missing explanation. We use it when it matches visible text, and we do not mark up claims the page does not show. ### Should every page be rewritten for models? No. Only the URLs assigned to a query or a prompt. Rewriting the archive for its own sake creates competing extracts. ### Do you add FAQ schema everywhere? Only where the questions are real and visible. A block of invented questions is not a deliverable. ## Related - [On-page SEO](https://shoutrank.com/on-page-seo): The assignment of a query to a URL. - [AI SEO](https://shoutrank.com/ai-seo): The program this writing serves. - [Pages that get cited](https://shoutrank.com/insights/structure-pages-for-citations): The field note on structure.