ChatGPT pulls dozens of pages for each query but only cites about half of them. Behind the scenes it tags each URL with a ref_type—search, news, reddit, youtube or academia—and decides which ones to open and reference based on the title, URL and snippet before ever fetching full content. In a dataset of 1.4 million ChatGPT 5.2 prompts, 88.46% of cited links came from the general search pool (25.6 million data points), while news barely hit 12%, and YouTube, Reddit and academic sites hovered under 2% each. Even though Reddit makes up 67.8% of non-cited URLs, it’s almost never credited in final answers, serving mainly to shape context behind the scenes. You might think fields like snippets or publication dates would skew citation rates, since non-cited pages showed snippets 14.8% of the time versus 4.36% for cited ones and pub_dates 92.7% versus 36%. But that gap collapses once you strip out Reddit’s API-fed content. Within the search vertical alone, snippets are nearly non-existent for both cited and non-cited links (around 2.5% vs. 0.1%), and pub_date sits at 33.8% for cited versus 49% for non-cited. Those early patterns turn out to be artifacts of how data flows through ChatGPT’s pipeline, not real signals of citation preference. The one clear driver of citation is semantic relevance. Using cosine similarity on embeddings, Ahrefs found that cited URLs have higher title similarity to both the original prompt (0.602 vs. 0.484) and to internally generated “fanout” queries (max match 0.656). In other words, it’s not the presence of a snippet or a date stamp that wins a citation—it’s matching the model’s sub-questions closely. Focusing on titles that align with those hidden fanout queries looks like the most reliable way to get ChatGPT to link back to your content.