Large language models cite sources differently than Google ranks them.
Search Atlas, an SEO software company, compared citations from OpenAI’s GPT, Google’s Gemini, and Perplexity against Google search results.
The analysis of 18,377 matched queries finds a gap between traditional search visibility and AI platform citations.
Here’s an overview of the key differences Search Atlas found.
Perplexity Is Closest To Search
Perplexity performs live web retrieval, so you would expect its citations to look more like search results. The study supports that.
Across the dataset, Perplexity showed a median domain overlap of around 25–30% with Google results. Median URL overlap was close to 20%. In total, Perplexity shared 18,549 domains with Google, representing about 43% of the domains it cited.
ChatGPT And Gemini Are More Selective
ChatGPT showed much lower overlap with Google. Its median domain overlap stayed around 10–15%. The model shared 1,503 domains with Google, accounting for about 21% of its cited domains. URL matches typically remained below 10%.
Gemini behaved less consistently. Some responses had almost no overlap with search results. Others lined up more closely. Overall, Gemini shared just 160 domains with Google, representing about 4% of the domains that appeared in Google’s results, even though those domains made up 28% of Gemini’s citations.
What The Numbers Mean For Visibility
Ranking in Google doesn’t guarantee LLM citations. This report suggests the systems draw from the web in different ways.
Perplexity’s architecture actively searches the web and its citation patterns more closely track traditional search rankings. If your site already ranks well in Google, you are more likely to see similar visibility in Perplexity answers.
ChatGPT and Gemini rely more on pre-trained knowledge and selective retrieval. They cite a narrower set of sources and are less tied to current rankings. URL-level matches with Google are low for both.
Study Limitations
The dataset heavily favored Perplexity. It accounted for 89% of matched queries, with OpenAI at 8% and Gemini at 3%.
Researchers matched queries using semantic similarity scoring. Paired queries expressed similar information needs but were not identical user searches. The threshold was 82% similarity using OpenAI’s embedding model.
The two-month window provides a recent snapshot only. Longer timeframes would be needed to see whether the same overlap patterns hold over time.
Looking Ahead
For retrieval-based systems like Perplexity, traditional SEO signals and overall domain strength are likely to matter more for visibility.
For reasoning-focused models like ChatGPT and Gemini, those signals may have less direct influence on which sources appear in answers.
Featured Image: Ascannio/Shutterstock
