Imagine asking a friend for directions, only to have them repeatedly point you to the same street corner, ignoring every other route. That’s essentially what a new deep-dive into ChatGPT’s search behavior has uncovered: almost a third of the AI’s search queries are repeats of previous questions, and its citation patterns are anything but random. A detailed analysis of 591 responses from ChatGPT’s search feature reveals a systematic bias—some sources get cited every time, while others are used only sporadically, even when they are equally relevant.
This isn’t just a quirk of lazy coding. It’s a window into how large language models (LLMs) like ChatGPT interact with the vast, messy ecosystem of the web, and it raises urgent questions about information diversity, fairness, and the hidden algorithms that decide what we read.
The study, published by a team of independent researchers on Habr, examined 591 responses generated by ChatGPT when it used its built-in web search tool. The goal was to understand how the AI selects sources for its answers—and what happens when the same question is asked multiple times. The results are startling: roughly one-third of all queries were repeats, meaning ChatGPT often re-asked the same search terms internally to gather more data. But the real story is in the citations.
The Citation Conundrum: Who Gets Quoted?
The researchers found that certain websites were cited in nearly every relevant response, while others, with similar authority and content, were ignored almost entirely. For example, major news outlets like Reuters and The Associated Press appeared in over 90% of relevant political news queries. In contrast, smaller, specialized blogs or regional news sites were cited only in about 15% of cases, even when their reporting was more timely or detailed. This isn’t necessarily a sign of malice—it’s a reflection of how ChatGPT’s underlying ranking models work.
The AI doesn’t ‘read’ the web like a human. Instead, it uses a combination of pre-training data and real-time search ranking signals (likely from Bing or a similar engine) to decide which sources to prioritize. Pages with higher domain authority, more backlinks, and faster load times tend to win. But this creates a feedback loop: already dominant sources get more visibility, while new or niche voices are systematically sidelined. The analysis showed that the top 10% of cited sources accounted for over 60% of all citations, a classic power-law distribution.
Why Do Some Queries Repeat?
The finding that one-third of search queries are repeats is equally telling. When ChatGPT needs more information, it doesn’t just refine its question—it often fires off the exact same query again. This happens because the model’s internal search tool is designed to fetch multiple ‘snippets’ or ‘passages’ from the web, and if the first attempt returns incomplete results, it retries. In practice, this means ChatGPT can end up querying the same search engine multiple times for the same request, consuming bandwidth and potentially hitting rate limits. For users, this translates to longer wait times and, occasionally, less diverse answers.
The researchers noted that repeat queries were especially common for complex, multi-step questions, such as “What are the latest breakthroughs in quantum computing?” The AI would search once for a general overview, then re-search the same terms to find specific details from different paragraphs. While this sounds thorough, it also means that if the first search returns a biased set of sources, the repeats will only reinforce that bias.
The Hidden Impact on Information Quality
This citation bias isn’t just an academic curiosity. For businesses, journalists, and everyday users, it means that the information ChatGPT delivers is shaped by an invisible hierarchy of sources. If you ask for the latest news on a climate policy, you’re likely to get quotes from the same three or four global news agencies, even if local environmental groups or academic journals have more nuanced analysis. This can lead to a homogenized view of events, where alternative perspectives are filtered out before you even see them.
Consider a practical example: a user asks ChatGPT for “the best practices for remote team management in 2026.” The AI might cite Harvard Business Review and Forbes repeatedly, while ignoring a well-researched guide from a smaller HR consultancy that actually specializes in hybrid work. The user gets a solid but incomplete answer—and the smaller consultancy loses visibility. Over time, this could concentrate influence in a handful of megasites, making it harder for fresh ideas to break through.
What This Means for Developers and Power Users
If you rely on AI-generated search results for research or content creation, this study offers a cautionary tale. Here are three actionable takeaways:
- Diversify your prompts: Instead of a single generic question, ask ChatGPT to “compare perspectives from academic journals, news sites, and industry blogs.” This nudges the model to pull from different source types.
- Check the sources manually: If ChatGPT cites only a few domains, open those pages yourself and look for counterpoints. The AI’s summary may be accurate but incomplete.
- Use specialized tools: For critical research, don’t rely solely on ChatGPT’s search. Pair it with dedicated search engines or databases that let you filter by domain authority. For example, ASI Biont supports connecting to custom search APIs and data sources, allowing you to bypass default ranking biases—learn more at asibiont.com/courses.
The Bigger Picture: AI and the Attention Economy
This research lands at a critical moment. As LLMs become the default interface for information retrieval, their sourcing decisions will shape public knowledge. If one-third of queries are repeats and citations are concentrated among a few players, we’re building a system that rewards repetition over novelty. The solution isn’t to stop using AI search—it’s to demand transparency. Users should know which sources an AI favors and why. Regulators and developers alike need to ensure that citation algorithms are audited for fairness, just as social media feeds have been scrutinized for algorithmic bias.
The study’s authors suggest that minor tweaks—like randomizing source selection within a set of high-authority pages, or explicitly flagging when a result comes from a repeat query—could dramatically improve diversity. Until then, the burden falls on users to be critical consumers. When you see a ChatGPT response citing the same three sites again, ask yourself: what am I missing?
Conclusion
The revelation that one-third of ChatGPT’s search queries are repeats, combined with a stark citation bias, is a wake-up call for anyone who treats AI-generated answers as gospel. The model isn’t malicious—it’s just reflecting the statistical patterns of the web, which themselves are skewed. But that doesn’t mean we have to accept it. By understanding these biases, questioning the sources, and demanding better tools, we can push for a more equitable information ecosystem. Next time you get a perfectly polished answer from ChatGPT, remember: it might be a polished version of a very narrow sliver of the internet.
Comments