Where AI answers get their sources
If you want to be cited by an assistant, it helps to know how much room there is in an answer. The short version: much less than on a results page, and it is not filled the same way.
The largest study to date compared six assistants (ChatGPT, Gemini, Perplexity, Grok, Google AI Mode and Copilot Search) against Google and Bing across 55,936 queries, 124,287 domains and about 1.4 million citation links. Assistants cited a mean of 4.3 URLs and 3.4 domains per answer, against 10.3 URLs and 7.3 domains for traditional engines, and fewer than ten distinct URLs appeared in 80% of answers (Zhang et al., arXiv:2512.09483).
Four slots, not ten. That is the first thing worth internalising.
Ranking is not the same door
The same study found only 38% of domains appeared in both kinds of engine, while 37% were unique to the assistants, with response-level overlap under 40% for any pair of engines. A smaller 2026 study measured per-model overlap with Google’s results and found medians ranging from 0% for GPT-4o to 14.3% for Perplexity Sonar Pro (Chen et al., arXiv:2601.16858).
Vendor datasets point the same way while disagreeing on the number. Ahrefs reported in March 2026 that 38% of pages cited in AI Overviews also ranked in the top ten, down from about 76% in their July 2025 study (Ahrefs). Treat the precise figures as vendor measurements of a moving target; treat the direction as settled. A meaningful share of what gets cited was never ranking for the question.
What the cited domains have in common
The same large study looked at what distinguishes domains the assistants favour. Those domains “generally exhibit more structured, hierarchical HTML, easier-to-read text, lower domain popularity, and more outlinks to reputable sources”. The authors are careful, and so should you be: this is correlation across sampled URLs, not proof that restructuring your HTML earns citations. It does, though, point the same way as what we know about writing for retrieval.
One widely repeated claim deserves retiring. AI citations are often described as far more concentrated than organic results. Measured by Gini index, the assistants in that study showed lower concentration than Google and Bing, a slightly more even spread of domains. Being small is less of a disqualification here than the folklore suggests.
What to do with this
- Assume four slots. If your category answer already has four good sources, your work is displacement, not addition, which is why displacement belongs in your metrics.
- Stop inferring AI visibility from rankings. They overlap by something between a third and two thirds depending on who measures, so one is not a proxy for the other.
- Check who holds the slots in your category. Often they are not competitors at all but surfaces you do not own.
And be suspicious of precise citation-share statistics, including the ones above. They come from proprietary datasets, they move by large factors between measurement windows, and several figures in wide circulation trace back to no readable source at all.