AI-written pages were the most-used source in 27 of 50 prompts, Reihs finds in 43,200 answers

Yannick Reihs leads AI Search strategy at Radyant, an agency specialized in SEO and AI Search. In September 2026, he tested 80 buyer questions across six AI tools, including ChatGPT and both Google AI search tools. Each question ran 90 times per tool, giving 43,200 answers, tracked with Peec AI. He compared 656 pages to see which ones the tools used as sources. Pages were scored without knowing how often they had been used. Pangram checked which text appeared to be AI-written.

What the study found

• AI-written pages were the most-used source for 27 of 50 questions. They were among the three most-used sources for 36 of 50. On each site, pages with more AI-written text were no less likely to be used.
• Better pages were used more often. The tools regularly used 29% of low-scoring pages, 66% of middle-scoring pages and 88% of high-scoring pages. Here, “regularly” means at least 20 times during the month. Once pages answered the question well, the extra effect of a higher quality score was too small to measure.
• Pages did better when they answered the exact question, used the format it asked for and put its wording in the title. In Reihs's study, pages on one human-written site matching the question's words were used in roughly half the answers. A page whose title used an expert term instead appeared in just 7%.
• Losing Google rankings also hurt a site's chances in Google's AI answers. One large SaaS blog had none of its 15 tested pages in Google's top 20. AI Overviews never used those pages as sources. Claude still used them in one in five answers that involved web search.
• The AI detector could spot AI-written text even on good pages. Pangram flagged all 50 known AI-written pages and none of the 16 known human-written pages.
• Reihs found that none of the five domains with Google data showed a drop in impressions compared with September last year. They all grew, by 15% to 614%. This does not prove that every approach to AI content is safe.

What the study cannot tell us

The sites covered different topics and page types, so the study cannot fully separate their effects. One AI model scored the pages, with no second scorer to check its work. Only one AI detector was used. Bing and Brave rankings were not checked. The page scores reflect early-October snapshots, taken after the September tracking period.

The Brief's take

We think the useful first check is whether a page answers the buyer's exact question in the format they need. In this study, that mattered more than whether AI wrote the text. A page being used as a source does not always mean the answer visibly cites it.

The study also shows a risk for mass-produced pages with little value: when Google stops ranking them, its AI answers may stop using them too. Three of the sites studied are Radyant's own AI content projects. Read this as one agency's findings, with the limits above.