How ChatGPT Picks Sources: Reverse-Engineering AI Citations
Our Research Methodology
To understand how ChatGPT selects sources for its responses, we ran over 10,000 queries across 15 industry verticals and analyzed the resulting citations. We tracked which domains were cited, how they were referenced (direct recommendation, supporting evidence, or passing mention), and what content characteristics correlated with higher citation rates.
Our research covered queries from simple factual lookups to complex comparison and recommendation requests, giving us a comprehensive view of ChatGPT's citation behavior across different query types.
Authority Signals That Matter
The strongest predictor of citation in ChatGPT responses was what we call 'entity authority' — a combination of how frequently a brand or domain appears in the model's training data, how many independent sources reference it, and whether its claims are corroborated by other authoritative sources.
Domains with strong Wikipedia presence, consistent mentions across news outlets, and high-quality backlink profiles from .edu and .gov domains were cited 4x more frequently than domains without these signals. This suggests that ChatGPT's source selection heavily weights the same authority signals that Google has used, but applies them at the entity level rather than the page level.
Content Structure and Extractability
Content that was structured for easy extraction — clear headings, concise definitions, bullet-pointed feature lists, and comparison tables — was significantly more likely to be cited than equivalent content buried in long-form prose.
Specifically, pages that included a clear, self-contained definition or description in the first two paragraphs were cited 3x more often than pages where the key information was scattered throughout the content. This aligns with how retrieval-augmented generation (RAG) systems work: they retrieve relevant passages, so content that concentrates key information in extractable chunks performs better.
The Recency Factor
For topics where information changes frequently, ChatGPT showed a strong preference for recently published or updated content. Pages updated within the last 90 days were cited 2.5x more often than older content covering the same topics.
This has important implications for content strategy: regularly updating your key pages with current data, examples, and references can significantly improve your citation rate. It's not enough to publish authoritative content once — maintaining freshness is essential for sustained AI visibility.
Implications for Your GEO Strategy
Based on our findings, we recommend a three-pronged approach to improving ChatGPT citations: First, build entity authority through consistent presence across authoritative sources. Second, structure your content for extractability with clear definitions, comparisons, and data points. Third, maintain content freshness through regular updates with current information.
Refora tracks all of these factors automatically, showing you exactly where your content excels and where it falls short compared to competitors who are being cited more frequently.
Refora Research
Author
Refora Research publishes original data and analysis on AI search trends, citation patterns, and generative engine optimization strategies.