By Ryan Kings, Founder & CTO at AEOForged · Published 27 July 2026

What Sources Do AI Engines Like ChatGPT, Gemini, and Perplexity Actually Cite?

Direct answer: We asked ChatGPT, Gemini, and Perplexity 96 questions about answer engine optimization — across four separate studies, covering tool selection, agency trust, brand authority, and category legitimacy — and recorded every source they cited. Of 3,563 total citations, 89.5% pointed to editorial content (blogs, publications, vendor pages). Only two domains — youtube.com and linkedin.com — appeared in the top five cited sources across all four studies. And in same-day repeat tests, the leading cited source changed 97.5% of the time.

Each study targeted a different topic within the AEO and AI-visibility category, but all four used the same method: real questions submitted to real answer engines, with the cited sources recorded and classified. The four studies were run independently — different question sets, different topics, separate analysis — which makes the points where they agree more meaningful than any single study's results.

Key takeaways

• Editorial content dominates — 89.5% of all 3,563 recorded citations came from blogs, publications, and vendor content pages. This held in an 84.8%–92.0% band across all four studies.
• YouTube and LinkedIn are universal — both placed in the top five cited domains in every study. Reddit made it in three of four.
• Citation rankings are volatile — the top-cited domain changed on 39 of 40 same-day repeat tests across the four studies.

The four studies at a glance

Each study asked 24 questions to three answer engines (ChatGPT, Gemini, Perplexity), recording which domains were cited in each response. A subset of questions was repeated on the same day to test whether the top-cited source stayed consistent.

Each row links to the full study, which includes its own methodology and raw dataset.

What types of content do AI engines cite most?

Adding up the citation counts from all four studies gives a category-wide breakdown that no single 24-question study can show on its own:

The pattern was consistent: editorial content captured between 84.8% and 92.0% of citations in each individual study. That narrow range across four different question sets suggests this is a stable property of how these engines source answers in this category, not a quirk of any single study.

Which domains appear across all four studies?

A single study can tell you who won that particular set of questions. Comparing the top-cited domains across four independent studies tells you which sources are structurally dominant in the category.

• youtube.com — top five in all four studies (22.8%, 17.4%, 17.4%, 29.7% of answers citing it).
• linkedin.com — top five in all four studies (21.7%, 17.4%, 34.8%, 23.1%).
• reddit.com — top five in three of four. It led two studies outright (33.7% and 30.8%) but dropped to roughly 16th in the tool-evaluation study, where vendor comparison content dominated instead.
• blog.hubspot.com, medium.com, and semrush.com each appeared in the top 10–15 of three studies.

For brands competing for AI citations in this space, YouTube and LinkedIn presence appears to be a baseline requirement regardless of the question type. Reddit matters most for trust and legitimacy questions, less for direct tool comparisons.

How stable are AI citation rankings?

Each study included a same-day repeat test: the first 10 questions were asked three times on the same day to the same engine, and we checked whether the top-cited source stayed the same. Combined across all four studies, the leading domain changed between repeats in 39 of 40 test groups — a 97.5% flip rate.

This does not mean AI engines are unreliable in general. It means the specific domain that leads a given answer changes frequently, even within a single day. For anyone tracking AI visibility, a single check is a snapshot of a moving target, not a stable ranking.

Methodology

Each study submitted 24 questions to three AI answer engines (ChatGPT, Gemini, and Perplexity), recorded every source URL cited in the responses, and classified each source by type (editorial, social, Reddit, YouTube, documentation, review site). A 10-question variance panel repeated questions three times on the same day to measure citation stability. All four studies were run during July 2026.

This synthesis adds up the published results from those four studies. No new questions were asked to any engine. Where per-study numbers can be meaningfully combined (citation counts by source type, variance-panel flip counts), they are summed. Where they cannot (each study's leading-domain share has its own confidence interval and sample size), they are shown per-study. Each source study publishes its own methodology and dataset with per-question, per-engine detail.

Limitations

• All four studies cover the AEO and AI-visibility category specifically. The source mix in other categories (e.g. health, finance, travel) may differ significantly.
• Three engines were measured (ChatGPT, Gemini, Perplexity). Other AI answer surfaces may cite different source types.
• The studies were all conducted in July 2026. Citation patterns may shift as engines update their retrieval and ranking systems.
• 96 questions across four studies is a meaningful sample for this category, but not exhaustive. Each individual study's methodology page discloses its own sample-specific limitations.

Frequently asked questions

Why does editorial content dominate so consistently?

All three engines drew the large majority of their citations from blogs, vendor content, and publications rather than user-generated content or video — a pattern that held within 84.8%–92.0% across all four studies.

Should brands avoid Reddit and YouTube based on this?

No. youtube.com appeared in the top five cited sources in every study, and reddit.com in three of four. Both are part of the citation landscape — the data shows they are a baseline for competing in this category, not that they are unimportant.

Where can I see the raw data?

Each of the four studies linked in the table above publishes its own dataset with per-question, per-engine, per-run results.