--- title: "What does content that AI engines prefer look like? | Underneath" description: "AI engines favor pages that state the conclusion first, cover the topic fully, explain how and why, and pack checkable facts such as definitions and figures." canonical: "https://underneath.agency/resources/what-content-do-ai-engines-prefer" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # What does content that AI engines prefer look like? AI engines prefer content that states its main point first, covers the topic fully, explains how and why, and packs in checkable facts such as definitions and figures. Layout alone does little; what matters is that each section holds information an answer can reuse. Most of this evidence comes from lab tests and observational studies, so treat it as a pattern to test, not a guarantee. ## The short version 1. When researchers at Carnegie Mellon taught a system the rules AI engines seem to reward, rewrites following them raised visibility by 35.99% on average in [lab tests](https://arxiv.org/abs/2510.11438). 2. Cited pages with definitions shaped answers 57.33% more, and pages with comparisons 55.28% more, in [an analysis of pages cited by ChatGPT, Google and Perplexity](https://arxiv.org/abs/2604.25707). 3. Question-and-answer formatting on its own did not help in that analysis; it went with 5.74% less influence. 4. Google’s AI Overviews rarely copy sentences: the median share of five-word sequences lifted verbatim from a cited page was 0.0% in [our study](https://underneath.agency/research/ai-overview-cited-pages-study). ## What patterns do AI engines reward in a page? They reward pages that are comprehensive, factual, neutral in tone, clearly organized and specific. An automated system learned these rules from tens of thousands of AI engine choices. Wu and colleagues at Carnegie Mellon University built AutoGEO. For each question, it compares the page an AI engine used most with the one it used least, then writes down the differences as rules. The common rules included covering the topic comprehensively, attributing claims to credible sources, keeping a neutral tone without promotion, and using clear headings and lists. For in-depth research questions on Gemini, it also learned three more rules. State the key conclusion at the beginning, explain causes and mechanisms, and back claims with specific data or named examples. Rewrites that followed the rules raised the lab visibility scores by 35.99% on average, without lowering the quality of the AI’s answers. The tests used fixed sets of five candidate pages per question, not the live web. ## Should you put the conclusion first? Probably, though the direct evidence is thin. “Conclusion first” is one of the learned rules, but no study has tested it alone on live engines. The AutoGEO authors show one worked example. Their rewritten version presents the main thesis upfront, discusses the topic more thoroughly and explains the underlying how and why. That is a single paragraph, judged by the authors, not a measurement. The GEO-16 audit by [Kumar and Palkhouski](https://arxiv.org/abs/2509.10762) recommends the same thing: an answer-first summary, compact paragraphs and descriptive headings. Its authors present this as a design principle rather than a tested result. Leading with the answer costs little, so the downside of trying it is small. ## Which kinds of content get used most in AI answers? Pages with definitions, comparisons, figures and code get reused most. Citations used only as bare references count least. Zhang and colleagues analyzed 18,151 pages cited by ChatGPT, Google AI Overviews and Perplexity across 602 prompts. They scored how much of each answer’s wording, structure and position could be traced to each cited page. Pages with definitions, comparisons, numbers, how-to steps or code scored higher. | Content on the cited page | Influence on the answer, against pages without it | |---|---| | Definitions | +57.33% | | Comparisons | +55.28% | | How-to steps | +41.20% | | Question-and-answer format | −5.74% | By role, citations used as definitions averaged an influence score of 0.1531, against 0.0529 for citations used only as references. The authors’ reading is that evidence works as reusable building blocks, while question-and-answer formatting is “only a surface wrapper”. They stress these are associations, not tested causes. ## Does longer, more structured content win? Depth and structure go together with being used, but length alone does not win. Once pages are compared fairly, word count fades. In the Zhang analysis, the most influential quarter of cited pages averaged 1,943.30 words, against 169.82 for the least influential. They also had 12.50 times as many headings. The authors warn this “should not be reduced to a rule that longer content always wins”. Our own Google data adds a check. In our study of 3,096 ranking pages, word count did not matter once ranking position and the search were held constant. But cited pages with an HTML table were 14.7 points more likely to have the answer’s wording traced to them. Other page features, such as dates and structured data, are covered in [on-page signals linked to AI citations](https://underneath.agency/resources/on-page-signals-linked-to-ai-citations). ## Do all AI engines want the same things? Mostly, within the same kind of question. Rules shift more by topic than by engine. In the AutoGEO study, the [rules learned from Gemini and GPT](https://underneath.agency/resources/do-ai-engines-prefer-same-content) overlapped by 78.95% on research questions. Rules learned from the two general question sets overlapped by 88.24%, but overlap with shopping questions fell to 34.78% and 40.00%. Shopping rules favored step-by-step guidance and product details such as model numbers and specifications, over in-depth explanation. A controlled test by [Vishwakarma and colleagues](https://arxiv.org/abs/2605.25517) found the same split. For product reviews, missing prices and missing specifications cost pages the first citation. Their test examined 18 content factors across 252,000 trials, in a simulation run by the software company Sprinklr. ## Is “answer-ready” content always a good idea? Only when the facts are real. AI-optimized writing is now detectable, and some of it props up weak claims. Researchers at the CISPA Helmholtz Center built a detector for content written to win AI citations. Run on 10,095 pages retrieved by Google Search and Gemini, it flagged an estimated 8.90% as optimized, rising to 16.36% among pages modified in 2026. The authors treat these as estimates, since the detector is not error-free. On the flagged pages, 69.34% of the sources they cited were rated low on verifiability: little editorial accountability, or hard for readers to check. The AutoGEO team also tested manipulative rewrites that try to hijack the AI. These raised visibility scores but always made the AI’s answers worse. Dense, factual pages are what engines reuse; invented facts dressed in the same format are a reputational risk. ## What should you do about it? Write each important page as a dense, honest briefing on one question. In practice: 1. Open with the answer or conclusion in a sentence or two. 2. Cover the topic fully, including the how and why behind the answer. 3. Add definitions, comparisons, specific figures and step-by-step instructions where they fit, each with a source. 4. Put product facts such as prices and specifications in plain text or simple tables. 5. Keep the tone neutral; cut sales language and hedging. 6. Do not convert pages to question-and-answer format expecting a lift on its own; see [whether FAQ pages help citations](https://underneath.agency/resources/do-faq-pages-help-ai-citations). If you want help applying this across your key pages, see our [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization). ## What does the research not tell us yet? No study has rewritten real pages to this pattern and tracked live AI citations over time. - **Lab versus live.** AutoGEO’s gains come from fixed sets of five pages per question, with the target page already included. - **Association versus cause.** The influence analysis observed existing pages. Better-written pages may also come from stronger brands. - **Conclusion first, alone.** Its effect has not been isolated from the other rules. - **Industry differences.** Shopping, research and local questions reward different things, and many sectors are untested. - **Measurement.** The influence score is the authors’ own proxy built from word overlap and position, not a view inside the AI. ## Frequently asked questions ### How should I structure content for AI search? Lead with the answer, then cover the topic fully with definitions, comparisons and specific figures under clear headings. In one analysis of cited pages, definitions went with 57.33% more influence on the answer. ### Does FAQ formatting help with AI citations? Not on its own. Pages in question-and-answer format showed 5.74% less influence on AI answers than other pages in an analysis of ChatGPT, Google and Perplexity citations. ### Is longer content better for AI engines? Only when the length carries more useful structure and facts. The most influential cited pages averaged 1,943.30 words, but in our Google study word count stopped mattering once ranking was accounted for. ### Do AI engines copy text from my page? Rarely word for word. In our study of Google’s AI Overviews, the median share of five-word sequences copied verbatim was 0.0%, so clear facts matter more than catchy phrasing. ## Sources - Wu, Zhong, Kim and Xiong (2025), [What Generative Search Engines Like and How to Optimize Web Content Cooperatively](https://arxiv.org/abs/2510.11438), arXiv:2510.11438. - Zhang, He and Yao (2026), [From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms](https://arxiv.org/abs/2604.25707), arXiv:2604.25707. - Chu and colleagues (2026), [GEO-Flag: Detecting and Measuring GEO-Optimized Web Content](https://arxiv.org/abs/2608.16824), arXiv:2608.16824. - Kumar and Palkhouski (2025), [AI Answer Engine Citation Behavior An Empirical Analysis of the GEO16 Framework](https://arxiv.org/abs/2509.10762), arXiv:2509.10762. - Vishwakarma, Kumar and Jamidar (2026), [What Gets Cited: Competitive GEO in AI Answer Engines](https://arxiv.org/abs/2605.25517), arXiv:2605.25517. - Underneath (2026), [What pages cited by AI Overviews have in common: 3,096 pages](https://underneath.agency/research/ai-overview-cited-pages-study) --- This is the Markdown twin of https://underneath.agency/resources/what-content-do-ai-engines-prefer. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.