--- title: "Are AI search engines citing AI-generated content? | Underneath" description: "Yes. A 2026 audit of four AI search engines flagged about 16% of readable cited pages as AI-generated, from 7.3% for ChatGPT to 27.8% for Copilot." canonical: "https://underneath.agency/resources/are-ai-search-engines-citing-ai-content" 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 # Are AI search engines citing AI-generated content? Yes. In a 2026 audit of ChatGPT, Copilot, Gemini and Perplexity on politics, health and environment questions, a detection tool flagged about 16% of the cited pages it could read as AI-generated, and none of the sampled pages disclosed it. Rates differed widely by engine. It is one study using one detector, so treat the figures as an estimate, probably on the low side. ## The short version 1. Across 712 real questions, about 16% of the unique sources four AI search engines cited were judged AI-generated ([Allaham and Diakopoulos](https://arxiv.org/abs/2605.23684)). 2. The share ranged from 7.3% of ChatGPT’s cited sources to 27.8% of Copilot’s, with Gemini at 14.7% and Perplexity at 9.4% (Allaham and Diakopoulos). 3. 97.1% of the AI-generated sources came from outside the 25 most-cited websites, so they sit in the long tail rather than on familiar domains (Allaham and Diakopoulos). 4. A separate detector flagged 8.90% of pages in Google and Gemini results as rewritten to win AI citations, rising to 16.36% among pages updated in 2026 ([Chu and colleagues](https://arxiv.org/abs/2608.16824)). 5. In a simulation, a single planted page at the top of the results led the most vulnerable AI assistants to recommend a fake product in 27% of cases ([Luo and Chen](https://arxiv.org/abs/2606.13610)). ## How much of what AI search cites is AI-generated? About one in six cited pages, in the only audit that has measured it directly. [Allaham and Diakopoulos](https://arxiv.org/abs/2605.23684) at Northwestern University put 712 real user questions about US politics, health and the environment to ChatGPT, Copilot, Gemini and Perplexity, through each product’s normal interface. The 2,848 answers cited 26,266 unique web pages. The team could extract text from 72.9% of them and ran that text through Pangram, a detection tool that estimates whether writing was produced by AI. About 16% of the readable cited sources were classed as likely or highly likely AI-generated. None of it was labeled. The authors checked 200 flagged pages by hand for any statement that AI had been used, and found no such disclosure. ## Which engines cite the most AI-generated pages? Copilot, by a wide margin, and ChatGPT the least. Each figure below is the share of that engine’s cited sources that the detector flagged. | Engine | Cited sources flagged as AI-generated | |---|---| | Copilot | 27.8% | | Gemini | 14.7% | | Perplexity | 9.4% | | ChatGPT | 7.3% | ChatGPT had the lowest share despite citing the most pages, an average of 14.7 unique sources per answer. By topic, the paper reports shares of 16.3% for health, 13.3% for the environment and 11.1% for politics. ## Where do the AI-generated sources come from? Mostly from obscure sites, not the big domains engines cite most. The 25 most-cited websites accounted for 23.8% of all citations, but only 2.9% of the AI-generated sources. The other 97.1% came from the long tail of rarely cited domains. The problem is spread thin rather than concentrated. 28.0% of all cited domains, 1,754 sites, had at least one cited page flagged as AI-generated. They included niche advocacy sites, Reddit, some news sites and even government health pages. For the kinds of sites engines lean on overall, see [which websites AI search relies on](https://underneath.agency/resources/which-websites-do-ai-search-engines-cite). One clear case: the eighth-largest source of flagged pages, with 28 cited pages, describes itself on its homepage as an “AI-powered research tool.” ## How reliable are these estimates? Reasonably, as a lower bound, but they rest on one detector and three topics. The authors tested Pangram first: it classified 100% of 200 human-written test texts correctly. But on 105 articles from sites that journalists had reported as publishing AI-generated news, it missed 31.4%, so the true share of AI-generated citations is probably higher. Other limits matter. About 27% of cited pages could not be read at all, often PDFs or videos. The questions covered only politics, health and the environment, in English, and the authors stress that AI-generated content is not necessarily low quality. The paper is a preprint. ## Is content written to win AI citations growing too? It appears so, according to one detection study. [Chu and colleagues](https://arxiv.org/abs/2608.16824), at the CISPA Helmholtz Center for Information Security, built a detector for [pages rewritten to be picked up by AI search](https://underneath.agency/resources/how-much-web-content-is-optimized-for-ai-search), a practice called generative engine optimization (GEO). They ran it on 10,095 pages returned by Google Search and by Gemini for 1,000 real queries. The detector flagged 8.90% of pages overall: 8.14% of Google Search pages and 9.09% of the pages Gemini cited in its answers. Among pages that declared when they were last updated, the rate rose from 7.02% for 2024 to 12.80% for 2025 and 16.36% for 2026. Only 19.57% of pages carried such a date, so the trend rests on a minority of pages. The flagged pages’ own evidence was often weak: 69.34% of the citations inside them pointed to sources with little editorial accountability or that were hard to inspect. The authors call all of these estimates, not definitive measurements. Self-serving content reaches AI answers in other forms too. In [our study of “best of” lists cited by AI](https://underneath.agency/research/self-promoting-best-lists-study), 24.2% of cited numbered lists with an identifiable publisher ranked that publisher first, though such lists made up only 1.1% of all citations. ## Why does this matter for brands? Because planted or synthetic pages can shape what AI assistants recommend. In a simulation by [Luo and Chen](https://arxiv.org/abs/2606.13610), researchers rewrote real retrieved web pages to promote fake products and tested 12 AI models on 225 real products. A single altered page at the top of the results fooled the most vulnerable models in 27% of cases. Replacing the top three pages raised the worst model’s rate to 73.8%. The same page placed lower in the results was nearly harmless, with fooled rates of 1–4%. This was a controlled test, mostly in Chinese, not an audit of live engines. Still, together with the audits above, it shows that your accurate pages compete for citation with mass-produced ones, and that whatever ranks first can carry outsized weight. We look at deliberate misuse in [whether GEO can push false information](https://underneath.agency/resources/can-geo-push-false-information-into-ai-answers). ## What should you do about it? Know which pages are cited for your category, and give engines better material than the synthetic alternatives. 1. Run your buyers’ key questions through the main assistants and look at who wrote the cited pages. Flag anonymous or obviously machine-written sources describing you or your category. 2. Publish original, checkable material: your own data, named experts and specific facts that a generic AI-written page cannot supply. 3. If you use AI to draft content, have a subject expert edit and fact-check it. Errors on your own pages can be repeated in AI answers with your name attached. 4. Watch for fake reviews, clone sites or misleading comparison pages about your products, and report them to the platforms hosting them. For help monitoring the sources AI search uses about your brand, see our [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization). ## What does the research not tell us yet? The evidence is early: one citation audit, one page-level detector study and one simulation. - No study has measured AI-generated citations for product, software or local-business questions. - No one has tested whether cited AI-generated pages are less accurate than human-written ones. - We found no evidence on whether engines favor or penalize AI-written pages when choosing what to cite. - The rising trend in pages written for AI citation is based on self-declared update dates on a fifth of pages, not on citations over time. - Detectors disagree and miss some AI text, so all shares here are estimates. ## Frequently asked questions ### Does ChatGPT cite AI-generated websites? Sometimes. In the 2026 audit, 7.3% of ChatGPT’s cited sources were flagged as AI-generated, the lowest of the four engines tested. ### Can tools tell whether a cited page was written by AI? Partly. The detector used had no false alarms on 200 human-written texts, but it missed 31.4% of articles from known AI content sites. ### Do websites disclose that their content is AI-generated? Rarely, if ever, in this audit. Of 200 cited pages flagged as AI-generated and checked by hand, none disclosed the use of AI. ### Is AI-generated content more likely to be cited by AI search? The research has not answered this. Studies so far measure how much AI-generated content is cited, not whether engines prefer it over human writing. ## Sources - Allaham and Diakopoulos (2026), [Synthetic Sources?: Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources](https://arxiv.org/abs/2605.23684), arXiv:2605.23684. - Chu, Leng, Li, Shen, Shen and Zhang (2026), [GEO-Flag: Detecting and Measuring GEO-Optimized Web Content](https://arxiv.org/abs/2608.16824), arXiv:2608.16824. - Luo and Chen (2026), [One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders](https://arxiv.org/abs/2606.13610), arXiv:2606.13610. - Underneath (2026), [How many “best of” lists cited by AI rank their own brand first?](https://underneath.agency/research/self-promoting-best-lists-study) --- This is the Markdown twin of https://underneath.agency/resources/are-ai-search-engines-citing-ai-content. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.