- A new academic preprint found that roughly 16% of sources cited by major AI search platforms showed signs of AI-generated content.
- Microsoft Copilot had the highest rate in the dataset at 27.8%, meaning more than one in four analyzed sources showed signs of AI generation.
Roughly one in six web pages cited by major AI search platforms shows evidence of being AI-generated.
That is the finding from a new academic preprint, and it points to a problem content teams and SEO professionals cannot ignore: AI search systems are pulling from a source pool that is increasingly filled with AI-produced content.
The research comes from Mowafak Allaham and Nicholas Diakopoulos. Their paper, “Synthetic Sources?: Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources”, is currently available on arXiv and has not yet completed formal peer review.
What the researchers found
The team submitted 712 real-world questions covering politics, health and the environment to ChatGPT, Microsoft Copilot, Google Gemini and Perplexity.
The four platforms produced 2,848 answers and cited 26,266 unique URLs in total. Of those, 19,154 sources were successfully analyzed using Pangram, an AI content detection system.
Approximately 16% were classified as likely or highly likely to contain AI-generated material.
The per-platform breakdown was:
- Copilot: 27.8%
- Gemini: 14.7%
- Perplexity: 9.4%
- ChatGPT: 7.3%
Copilot’s figure is the most striking. More than one in four sources it cited in this dataset showed signs of AI generation, a rate nearly four times higher than ChatGPT’s.
That gap raises questions about how differently each platform filters, retrieves and selects cited sources.
The detection caveat matters
The researchers were clear about one important limitation: AI detection tools are imperfect.
Pangram can classify a page as likely AI-generated, but it cannot confirm exactly how the content was produced or whether the information is accurate.
AI-assisted content is also not automatically low quality. A company can use AI to draft a piece and then have an expert verify, rewrite and improve it before publication.
The study measured signs of AI generation, not factual accuracy or editorial quality. Those are separate questions the research did not attempt to answer.
Why this creates risk for businesses
The larger issue is what happens when unverified AI-generated content enters the citation loop at scale.
Once a platform like Copilot or Gemini cites a page, that information becomes part of an AI-generated answer. If the original source contains errors, those errors now have a path to reach users through another system.
For businesses, the risk is direct.
If a third-party page describes a company’s product incorrectly, misquotes a price or places it in the wrong category, an AI platform may repeat that distorted information in an answer.
That makes consistency across credible third-party sources more important. AI platforms have more reference points to compare when the same accurate information appears in multiple places.
Citations are concentrated, but the long tail still matters
The study also found that citations were heavily concentrated.
Within the dataset, 59.1% of domains were cited only once and 16.5% were cited twice. At the same time, the top 25 domains accounted for 23.8% of all citations across the four platforms.
That mirrors a familiar pattern from traditional search: established publishers capture a large share of visibility.
But the long tail also matters. Many smaller and less prominent websites did receive citations, which is important for businesses that are not major publishers.
As we covered in earlier reporting on what may influence AI citation decisions, niche relevance and content specificity can matter even for sites without large authority footprints.
A business does not need to be the biggest source on the web. It needs to be the clearest and most specific source for the exact question an AI platform is trying to answer.
What marketers should take from this
The study is limited to three topic areas: politics, health and the environment. Citation behavior in ecommerce, SaaS, local services or B2B marketing may look different.
The paper is also a preprint, and the detection method carries uncertainty.
Still, the finding is hard to ignore. AI search platforms are already citing a meaningful amount of content that shows signs of AI generation.
For content and SEO teams, the question is not simply whether to use AI in production.
The better question is whether the final page contains something an AI platform has a real reason to cite: original data, verifiable claims, specific facts, expert analysis or information not already repeated across hundreds of competing pages.
Generic content that mirrors what every other site publishes is now competing in a source pool increasingly filled with content produced the same way.
