NNT report finds AI cites 11.6% of paragraphs in GEO study

NNT, an AI marketing engineering company, published a Generative Engine Optimization (GEO) report titled "What Content Does AI Cite?" on October 23, 2026, examining which paragraphs in AI-cited content are actually used as answer sources, according to datanet.co.kr.

NNT collected 37 blog posts from one automotive platform that were cited as sources in answers to 10 comparison, price and how-to questions posed to ChatGPT, Gemini and Perplexity. From these it analyzed 2,161 paragraphs, including list items and table rows, while excluding repeated phrases such as ads and calls to action.

Of the analyzed paragraphs, 250 were cited by AI, a citation rate of 11.6%. Paragraphs that scored one standard deviation higher on the information-scope item had citation odds 2.05 times higher, suggesting information density carried more weight than keyword repetition.

Keyword overlap with the question did not help. Paragraphs with no overlap with question words had a 12.5% citation rate, while paragraphs with under 25% overlap had 7.0%. Length also mattered: paragraphs under 30 characters had an 8.1% citation rate, compared with 14.3% for 80-149 characters and 14.5% for 150-249 characters.

Structure showed one of the sharpest gaps. Paragraphs under descriptive subheadings had a 12.4% citation rate, versus 1.6% for paragraphs without subheadings. The report includes a 10-item writing checklist, which recommends stating context within 2-4 sentences and including specific figures, examples and sources.

The report notes a limitation: the ChatGPT sample was restricted to 11 cited paragraphs drawn from 3 how-to questions.