ChatGPT Cites Only 1.93% of Reddit Pages — What 1.4M Prompts Reveal About AI Citation Mechanics

A new study from Ahrefs, published April 15, 2026, offers the most granular look yet at how ChatGPT decides which sources to cite — and which to silently consume. The research analyzed 1.4 million ChatGPT 5.2 desktop prompts from February 2025, tracking every URL retrieved and whether it in the end received a citation.

The headline finding: ChatGPT's overall citation rate sits at almost exactly 50/50 — 49.98% of retrieved URLs get cited, 50.02% do not. That average, though, masks dramatic disparities by source type, with Reddit sitting at the extreme end of the "used but never credited" spectrum.

1. The Ahrefs Study: Methodology and Scale

The study, authored by Louise Linehan and Xibeijia Guan, used cosine similarity scores computed from open-source embeddings to approximate ChatGPT's internal semantic matching process. This let the researchers reverse-engineer how the model evaluates title-to-query relevance when deciding which retrieved pages to cite. Across 1.4 million prompts, ChatGPT retrieved an average of roughly 16.5 URLs per prompt , nearly identical for both cited pages (16.57) and non-cited pages (16.58). The retrieval step itself is source-agnostic. The discrimination happens downstream, during the citation selection phase. Not at the retrieval gate. After it.

2. Citation Rates by Source Type , The Reddit Anomaly

Standard web search results are cited at an 88.46% rate , nearly 9 in 10 retrieved search pages make it into the response. Reddit, by contrast, is cited only 1.93% of the time, despite representing an enormous share of the retrieval pool. The visualization of non-cited URL distribution makes Reddit's role even starker: Reddit functions as a massive context reservoir. ChatGPT reads it voraciously to understand sentiment, user experiences, and conversational knowledge , then cites more "authoritative" web search results instead. More than two-thirds of all uncited URLs in ChatGPT responses come from Reddit. OpenAI paid for that data partnership and then buried the attribution. Make of that what you will.

3. What Actually Gets Cited: URLs, Titles, and Fanout Queries

Beyond source type, three specific factors significantly predict citation probability.

URL Structure: 8.67 Percentage Point Advantage

Pages with natural language, descriptive URL slugs (e.g., `/how-to-tune-meta-descriptions`) achieve an

89.78% citation rate compared to 81.11% for opaque or non-semantic URLs (e.g., `/article/58291`). That 8.67 percentage point gap is a real optimization lever , something our URL Slug Generator can help with directly.

Title-to-Query Semantic Alignment

Cosine similarity scores between page titles and queries tell a clear story:

The Fanout Query Mechanism ,

The Most Actionable Finding ChatGPT doesn't simply match pages against the user's original prompt. It generates internal "fanout" sub-queries , decomposing the user's question into specific information needs , and then matches pages against these narrower sub-questions. The highest citation probability goes to pages whose titles align with these granular sub-queries (0.656 cosine similarity) rather than the broad original prompt (0.602). This aligns with the

agentic search patterns we've been tracking, where AI systems decompose queries before acting.

4. Page Age and Authority Signals

The study shows a strong preference for established content. Cited pages in the search category have a median age of approximately

500 days (~1.3 years), with cited pages observed as old as 2,700+ days (7.4 years). Non-cited pages tend to be significantly younger. This contrasts with earlier research: a previous Ahrefs study from July 2025 found a median cited page age of 958 days. The shift toward newer-but-still-established pages may reflect ChatGPT's evolving retrieval calibration, but the core pattern holds , fresh-off-the-press content rarely gets cited. This age-bias compounds the broader citation dynamics we analyzed in an earlier 815K-page study. For news content, the pattern inverts slightly: cited news pages have a median age of ~200 days while non-cited news pages skew older at ~300 days, suggesting recency matters more within the news category where timeliness is intrinsic to value.

What This Tells Us About AI Authority Signals

ChatGPT uses page age as a proxy for content that's been validated over time , pages that have accumulated backlinks, user engagement, and indexing history. The implication for evergreen content strategies is direct: building authoritative, lasting content pays compounding dividends in the AI era, just as it does in traditional SEO. Same game. Higher stakes. If you're working on tuning for AI citation likelihood, our

LLM Citation Checker can help score your content across ChatGPT, Perplexity, and Gemini.

5. Actionable SEO Implications for AI Citation Optimization

The Ahrefs study translates into a concrete optimization checklist for sites that want to be cited by AI systems.

Tune titles for sub-questions, not just topics. Think about the specific fanout queries a user's broad question might decompose into. Structure content to answer these narrow, specific sub-questions explicitly , and reflect that specificity in your `