Ahrefs analyzed 1.4 million ChatGPT prompts. Ranking in classic search still gates most citations, but which page gets cited has changed.
Ahrefs analyzed 1.4 million ChatGPT prompts to answer a specific question: when ChatGPT cites a source, where did that source actually come from? The retrieval system tags every URL it pulls in with an internal label, called a ref_type, marking which channel it came through: general search, news, Reddit, YouTube, or academic sources. Search dominates. In Ahrefs' own words, 88% of the URLs that end up cited by ChatGPT are taken directly from that general search index, the same index that powers classic organic rankings.
What the study actually measured
The study, run by Ahrefs' Louise Linehan and data scientist Xibeijia Guan and published in April 2026, looked at ChatGPT's retrieval pipeline across five ref_types. Search accounted for 25.6 million of the data points analyzed and dominated the citation share. News made up 12.01% of citations, Reddit 1.93%, YouTube 0.51%, and academic sources 0.40%, despite Reddit alone contributing over 16 million retrieved URLs to the dataset, more volume than every other non-search channel combined.
That gap between volume and citations is the more interesting finding. Reddit accounts for 67.8% of every URL that ChatGPT retrieves but doesn't end up citing. ChatGPT pulls Reddit threads in constantly, apparently uses them to gauge consensus and build context, and then almost never gives Reddit credit in the actual answer. As the researchers put it, it learns from the crowd and cites another source. If your content strategy assumes a Reddit-heavy approach to AI visibility, this specific study is a reason to be skeptical of exactly how much citation credit that approach earns.
Why this isn't the same as "keep doing SEO exactly like before"
None of this means ranking your target keyword is enough on its own, and a separate, complementary Ahrefs study makes that limit clear. That second study, based on 863,000 SERPs and 4 million citations from Google's AI Overviews specifically (a different product from ChatGPT, though built on a related retrieval mechanism), found that only 37.9% of AI Overview citations also ranked in the top 10 organic results for that exact query, down sharply from about 76% roughly a year earlier.
The explanation Google itself provides is query fan-out: rather than pulling citations only from the literal query a person typed, Google's own documentation confirms that AI Overviews split a query into multiple related sub-questions and pull citations from whichever pages rank well across that broader set, not necessarily the original search term. A page can fail to rank for someone's exact phrase and still get cited, because it ranks well for one of the fan-out sub-questions the system generated behind the scenes. The practical shift this implies is that topical coverage across a cluster of related queries now matters more than optimizing a single target keyword in isolation.
Putting the two findings together
Read together, the two studies describe a system that still runs on classic search infrastructure but doesn't reward classic single-keyword ranking the way it once did. The retrieval layer, confirmed independently by Ahrefs' ChatGPT research, draws the overwhelming majority of its citable material from the same general search index that ranks pages for ordinary Google queries. Being invisible to that index still means being invisible to ChatGPT's citations in the vast majority of cases. But once a page is inside that retrieval pool, what determines whether it actually gets cited has shifted toward matching the AI's internally generated sub-questions, not just the literal query someone typed into the search box.
That's a different mechanism from what this blog has already covered on backlinks and brand mentions, which looked at what correlates with a brand actually getting mentioned by name once it's already part of the conversation. This is about something upstream of that: whether a page makes it into the pool of retrievable, citable sources in the first place, and the data says the general search index, the same one classic SEO has always targeted, is still doing almost all of that gatekeeping.
What this means in practice
The old SEO fundamentals, targeting relevant keywords, matching search intent, and fixing the technical issues that keep pages out of the index, remain the entry ticket into ChatGPT's citation pool, because that pool is drawn overwhelmingly from the same general search index those fundamentals were always built for. What has changed is the ceiling above that entry ticket: ranking on page one for one specific phrase no longer guarantees a citation, because the system is increasingly pulling from a wider set of related queries a person never actually typed. The practical response is to keep building topical depth across a cluster of related searches rather than treating any single keyword ranking as the finish line.
None of this replaces the need for real backlinks, either. A page that never earns the rankings to enter the general search index in the first place never gets the chance to compete for a citation on the fan-out queries either.
