Across the research done to date, AI systems appear to frequently answer without searching the web at all, cite only a handful of the pages they do read, and in many cases give no clickable link back to a source. This is a reference for the numbers behind that, drawn from peer-reviewed studies where possible and labeled working papers where not.
The headline findings and key takeaways
Gemini gave no clickable citation in 92% of its answers (Strauss et al., Data & Policy, 2026, peer-reviewed).
34% of Gemini answers and 24% of GPT-4o answers were generated without fetching any web content at all (Strauss et al., 2026).
Perplexity's Sonar visited about ten relevant pages per query but cited only three or four (Strauss et al., 2026).
Mean citations per prompt came to about 6.9 for ChatGPT, 12 for Google, and 16 for Perplexity (Zhang et al., 2026, preprint).
AI Overviews appeared on 51.5% of actual user queries in one study (Grossman et al., SIGIR, 2026, peer-reviewed) and 31.2% in another (Ng & Wessel, 2026, working paper).
The sources different engines retrieve, and can therefore cite, barely overlap, with under 0.2 average Jaccard similarity between them (Grossman et al., SIGIR, 2026, peer-reviewed).
Answer engines cite only a few of the sources they retrieve, and position and topical relevance decide which ones (Vishwakarma et al., SIGIR, 2026, peer-reviewed).
How often AI searches the web at all
Before an AI answer can cite anything, it has to fetch a page, and often it doesn't.
Analyzing about 14,000 logged search-assistant conversations, one study found that 34% of Gemini answers and 24% of GPT-4o answers were produced without the system fetching any online content (Strauss et al., Data & Policy, 2026). When a system answers without searching, there is no source to cite.
How often it gives a clickable citation
When the systems do search, they still credit little of what they read. The same study mentioned above found Gemini gave no clickable citation in 92% of its answers. It also measured citation efficiency, the extra citations provided per additional relevant page visited, and found it ranged from 0.19 to 0.45 across models on identical queries, meaning most relevant pages a model reads go uncredited.
On average, a query answered by Gemini or Perplexity’s Sonar left about three relevant websites uncited (Strauss et al., 2026). The authors note this reflects retrieval and product design choices rather than a technical limit.
How many sources it cites, and how much they count
Citation counts vary widely by platform. One analysis of 602 controlled prompts found mean citations per prompt of about 6.9 for ChatGPT, 12 for Google, and 16 for Perplexity (Zhang et al., 2026, preprint). Perplexity's Sonar was separately measured visiting roughly ten pages per query while citing three or four (Strauss et al., 2026).
Raw counts hide a second pattern. The same 602-prompt analysis found ChatGPT cited the fewest sources but drew more heavily on each one, with higher average influence per cited page than Google or Perplexity. The takeaway is that breadth of citation and depth of use are separate things, and a model that lists many sources is not necessarily leaning on any of them.
How often an AI answer even appears
Of course, citation behavior only matters where AI answers actually show up, and that share is now large. A benchmark of 11,500 actual user queries found AI Overviews were generated for 51.5% of them and placed above the organic results (Grossman et al., SIGIR, 2026).
A separate study of more than 2,000 queries put AI Overview prevalence at 31.2%, and argued the feature is deployed selectively by query intent rather than everywhere (Ng & Wessel, 2026, working paper). The two studies used different query samples, so treat the range, not either endpoint, as the finding.
Which sources get cited when they do
Because answer engines cite only a few of the pages they retrieve, being retrieved is not enough. A controlled study running 252,000 paired comparisons across six models found that topical relevance and list position were the biggest drivers of which of two competing sources got cited first, with explicit pricing and a recent timestamp adding smaller gains, and formatting-only edits making little difference (Vishwakarma et al., SIGIR, 2026). In other words, citation is driven by relevance and position, and barely by AI-specific formatting.
Which sources get cited also depends heavily on which engine is answering. The 11,500-query benchmark found the sources retrieved by Google Search, its AI Overview, and Gemini overlapped very little, under 0.2 average Jaccard similarity, with traditional search leaning toward government, education, and institutional sites and the generative engines leaning toward Google-owned content. It also found that sites blocking Google’s AI crawler were significantly less likely to be retrieved by AI Overviews, even when their content was otherwise accessible (Grossman et al., SIGIR, 2026).
Why are these numbers important?
The research currently describes a channel that passes on far less traffic than its reach suggests. Gemini gave no clickable citation in the large majority of its answers, the systems that do cite credit only a few of the pages they use, and any citation sits inside an answer that can resolve the query on its own.
For a brand, that means being cited is a weak proxy for being seen. What the AI says about a brand inside the answer, whether it names and recommends it, reaches the user whether or not a citation is ever clicked. Citations still matter as a signal and a foundation, but treating citation counts as the goal measures the smallest part of the picture.
How this fits with our own research at Traqer
This is one reason why we built Traqer to measure brand mentions separately from citations, rather than blending them into one score. A citation is a link a brand might occasionally earn a click from.
A mention is the model stating, inside the answer itself, what a brand is and whether to recommend it, which is what actually reaches the buyer. We track both per model and at the topic level, because, as the studies above show, citation behavior differs sharply from one system to the next and a single blended number hides that. We treat citations as a foundation signal and an outreach map, not as the commercial outcome.
The limits of this evidence
The studies measure citation in different ways, from clickable links to search-layer citations to pages visited versus credited, so the figures are not directly comparable and are best read as a consistent direction rather than one number.
The attribution study covers a specific set of models and notes that GPT-4o's low uncited gap partly reflects selective disclosure of its logs rather than better attribution.
The per-model citation counts and one of the prevalence figures come from a preprint and a working paper, labeled above, rather than peer-reviewed studies.
These systems change quickly, so every figure is a snapshot of the models and products as tested, not a fixed property of AI search.
Final conclusions
AI answers credit few of their sources. Many are produced without a search, most of the pages that are read go uncredited, and citation counts vary enough across platforms that no single number captures it.
For anyone trying to measure how they show up in AI, the practical conclusion is that citations set a low ceiling on traffic, and what the model says about a brand inside the answer is the larger prize.
Sources:
Strauss, I., Yang, J., O'Reilly, T., Rosenblat, S., & Moure, I. (2026). The attribution crisis in LLM search results: estimating ecosystem exploitation. Data & Policy, 8, e15. Peer-reviewed. https://doi.org/10.1017/dap.2026.10064
Zhang, K., He, X., & Yao, J. (2026). From citation selection to citation absorption: a measurement framework for generative engine optimization across AI search platforms. Preprint (not peer-reviewed). arXiv:2604.25707.
Grossman, R., Liu, S., Chen, M. K., Smith, M., Borcea, C., & Chen, Y. (2026). How generative AI disrupts search: an empirical study of Google Search, Gemini, and AI Overviews. Proceedings of SIGIR '26. Peer-reviewed. https://doi.org/10.1145/3805712.3809667
Ng, R., & Wessel, M. (2026). AI Overview or overreach? Google's strategic deployment of generative AI in search. CRC TR 224 Discussion Paper No. 742 (working paper, not peer-reviewed).
Vishwakarma, R., Kumar, S., & Jamidar, R. (2026). What gets cited: competitive GEO in AI answer engines. Proceedings of SIGIR '26. Peer-reviewed. https://doi.org/10.1145/3805712.3808445