Key Takeaways
- The principle: An AI does not cite a page; it cites a passage. This passage must be locatable (under a clear title), self-contained (understandable on its own), factual (figures, definitions), and attributable (author, date, source).
- Twelve rules: One idea per section, question in the title, answer at the top, self-contained paragraphs, clearly stated topic, cited facts, tables and lists, explicit definitions, FAQs, author and date, clean HTML, no hidden content.
- What doesn't work: long introductions, storytelling without facts, pronouns without referents, and text embedded in images or JavaScript.
- The test: Copy a single paragraph into a wizard and ask it what the paragraph is about and who is making the claim.
Language models read a page differently than humans do: they break the text down into sections, assess the relevance of each section to the question asked, and select the ones that provide an answer, attributing them to the source. Content that is well-structured for LLMs is therefore content in which each section answers a specific question, stands on its own, and contains verifiable facts. These rules also benefit the reader and SEO. This article provides a twelve-point checklist and an example; the technical implementation (markup, HTML) is in Structuring Your Pages for AI.
What a template does to a page
- Layout: The page is divided into sections (by headings, paragraphs, and table cells).
- Selection: For the question asked, the passages that are most semantically similar are selected—usually two or three per page.
- Extraction: The student rephrases or quotes the passage; it must include the subject, the statement, and, if possible, the number.
- Credit: The source is identified (domain, title, sometimes author); indicators of reliability (author, date, consistency with other sources) influence the choice.
Consequence: The page as a whole matters less than the quality of its individual sections. A 3,000-word article with no self-contained sections is cited less often than an 800-word page where each paragraph answers a question. Search engines are described in Understanding AI Engines.
The 12-point grid
| # | Rule | Why | How to check |
|---|---|---|---|
| 1 | One idea per section | The structure follows the headings; a section that combines two topics results in ambiguous passages | Each H2 can be summed up in one sentence |
| 2 | The question in the title (H2 or H3) | Semantic anchor point for selection | The title uses the wording provided by users |
| 3 | The direct answer in the first paragraph | This is the excerpt | 40 to 60 words, independent, including the main point |
| 4 | Standalone paragraphs | A passage must be understood out of context | No “as mentioned above,” no “that” without a referent |
| 5 | The topic mentioned in each paragraph | Attribution: “GreenRed” rather than “we” or “the tool” | Each paragraph contains the subject's name |
| 6 | Facts: figures, dates, ranges, sources | AI systems prioritize verifiable statements | At least one fact per section |
| 7 | Tables and lists for comparisons and steps | Cell-by-cell extractable structure | HTML tables with headers, no images |
| 8 | Explicit Definitions | "X is..." in one sentence is the most commonly used format | The term in bold, with its definition in the same sentence |
| 9 | An FAQ with 3 to 6 questions | Corresponds to conversational phrasing | FAQ Page Markup, Answers of 40 to 70 Words |
| 10 | Author's Name and Date Updated | Reliability and Freshness | Visible on the page and in the Article markup |
| 11 | Clean HTML | Robots read HTML, not the rendered output | Headings in H tags, text in the HTML—not as images or loaded via JavaScript |
| 12 | Nothing Hidden | Collapsed accordions, tabs, and click-to-expand content are read if they are in the HTML; otherwise, they are ignored | Check the source code |
A Before-and-After Example
| Before | After |
|---|---|
| H2: Our Principles for Monitoring Positions “We believe it’s essential to monitor your positions regularly. As we’ve seen, this allows you to plan ahead. Our approach is different, and our clients appreciate it.” | H2: How Often Should You Monitor Your SEO Rankings? “Weekly tracking of SEO rankings is sufficient for an SME: daily fluctuations of 1 to 3 positions are just noise, and a trend is best assessed over a four-week period. GreenRed tracks rankings daily but reports weekly and monthly changes, with an alert triggered if a site drops more than 5 spots.” |
The second passage has a question-form title, a direct answer, a number, the subject named, and stands on its own: it can be extracted and attributed. The first passage says nothing that an AI could quote.
Length, tone, and vocabulary
- Length: 800 to 2,000 words per article, organized into 5 to 8 sections; depth is important, but each section should stand on its own.
- Tone: factual and direct; AI systems don’t need to be convinced—they need to be informed. The tone for AI engines is discussed in Should you adjust your tone when using AI engines?.
- Vocabulary: the users' perspective, with technical terms explained; synonyms help ensure consistency across various formulations.
- Originality: a perspective, a piece of information, or an example that other sources don't have; AI systems look for additional information, not repetitions.
The Two-Minute Test
Copy a single paragraph from your page into ChatGPT or Claude, and ask: “What is this passage about, who is stating it, and what fact does it contain?” If the answer is accurate, the passage is extractable. If the AI asks for context or can’t identify the subject, rewrite it. Repeat this with three randomly selected paragraphs: it’s the fastest structural audit there is. To take it further, apply the checklist using a systematic brief; see The GEO Content Brief.
How GreenRed Can Help You
Instead of juggling multiple tools, GreenRed's AI writer brings these metrics together in a single dashboard, compares them over time, and shows you which actions to prioritize. You can try it for free—no credit card required—from the page Rates.
Frequently Asked Questions
What is the ideal length for content cited by AI?
The length of the page is less important than the length of the sections: a direct answer of 40 to 60 words under each heading, sections of 150 to 300 words, and a total of 800 to 2,000 words per page. AI systems extract two or three passages per page; each must be understandable on its own.
Should we write differently for AI and for Google?
No: The rules that make a passage eligible for a snippet (question in the title, direct answer, facts, tables, FAQs, author) are also the ones that result in Google’s optimized snippets and improve readability. The only difference is the emphasis on self-contained paragraphs and the explicitly stated topic.
Is AI-generated content cited by AI systems?
They are considered valid if they follow the rules above and provide original information; they are not valid if they simply repeat what is already stated on ten pages, which is the typical flaw in texts generated without data. The origin of the text matters less than its accuracy, structure, and value.
Can AI read accordions and tabs?
Yes, if their content is present in the page’s HTML—even if it’s collapsed; no, if it’s loaded on click via JavaScript. Check the source code: if the accordion text appears there, it is read. Accordion-style FAQs using the FAQPage markup are a best practice.