More companies are trying to figure out how to get mentioned in AI answers. If someone asks ChatGPT or Google AI Mode for the best software, healthcare provider, college, insurance company, or local service, you want your brand to be part of the answer.
One way that happens is through the sources the engines retrieve. If a comparison site, review platform, industry publication, or other third-party page keeps getting cited, being included there puts your brand in the material the engine is using to build its answer.
The citation itself is not the goal. The goal is getting your brand mentioned or recommended, which raises a more useful question: which citation sources are actually worth pursuing?
We ran 2,080 prompts across 32 categories, repeated every prompt six times on both ChatGPT Search and Google AI Mode, and collected 24,960 answers. The results showed a clear recurring citation layer, but they also showed why chasing every site that happens to get cited today is probably a mistake.
36% of websites accounted for 93% of AI citations
After cleaning the data, we found 8,004 websites cited across the study.
Most barely showed up. Sixty-four percent of those websites appeared on only one distinct prompt, yet together they generated just 7.3% of all citation activity.
The other 36% appeared across multiple prompts and accounted for 92.7% of citations.
That tells us there is a huge long tail of sources that appear once or for one narrow question, while a much smaller group keeps getting retrieved.
It does not mean those recurring sources are automatically better. It does mean that a website cited once should not be treated the same way as a website appearing across dozens or hundreds of prompts.
ChatGPT and Google AI Mode use very different sources
The biggest difference in the study was between the engines themselves.
First-party and vendor-owned websites made up 52.3% of ChatGPT citations compared with 30.4% in Google AI Mode. Official, institutional, and reference sources were also much more common in ChatGPT.
AI Mode went the other direction with social and UGC platforms. Large social and UGC sources represented only 1.1% of ChatGPT citation activity but 31.7% of AI Mode citations.
The platform-level numbers make the difference obvious. AI Mode cited YouTube 8,744 times in our dataset. ChatGPT cited it twice. Reddit appeared across 692 AI Mode prompts compared with 61 ChatGPT prompts.
That does not mean one engine is using good sources and the other is using bad ones. It means they are drawing from very different parts of the web.
That matters because a site can look like a major “AI citation source” while almost all of its visibility is coming from one engine. FitSmallBusiness appeared across 168 ChatGPT prompts and zero AI Mode prompts. US News appeared across 192 AI Mode prompts and zero ChatGPT prompts.
So I would be careful with any analysis that blends all AI citations into one score.
Some sources really do show up across both engines
We also wanted to find the websites that kept appearing across both systems rather than being carried almost entirely by one.
Forbes appeared across 357 ChatGPT prompts and 280 AI Mode prompts. NerdWallet was 308 and 249. Trustpilot was 84 and 101. Healthline was 78 and 60. G2 was 140 and 57, while Bankrate was nearly even at 55 and 56.
BBB, ConsumerAffairs, Course Report, Teladoc Health, LendingTree, and Capterra also appeared repeatedly across both engines.
I find these sources more interesting because two different retrieval systems are repeatedly finding them. That still does not make them automatically “better” or prove they will remain important over time, but it is a stronger signal than raw citation volume from one engine.
Traditional authority played a role here, but it was far from the whole story. When we compared citation recurrence with DataForSEO backlink and organic-search metrics, the relationship was only weak to moderate.
Some very large websites barely appeared. Harvard had roughly 688,000 referring domains and showed up across two prompts. MIT appeared across five.
At the same time, smaller category-specific sites could perform extremely well. Course Report had roughly 4,100 referring domains but appeared across 40 ChatGPT prompts and 56 AI Mode prompts.
That is a good reminder that authority is not just about scale. It is also about relevance to the category.
Some very weak sites are getting a surprising amount of visibility
This is where things get interesting.
We found websites with almost no traditional authority showing up repeatedly in AI citations. CodingBootcamp.net had 53 referring domains and appeared across 32 prompts. HotelPriceWatch.com had 42 referring domains and appeared across 23. TXRAC.com had 67 referring domains and appeared across 41.
Some of those sites also had very shallow citation footprints. Roughly 95% of CodingBootcamp.net’s citation activity came from one URL. HotelPriceWatch was above 91%.
Compare that with NerdWallet, where hundreds of different URLs were cited and the most-cited page represented only about 1.8% of the site’s citation footprint.
That feels like an important distinction. A site repeatedly supplying useful pages across its entire domain looks different from one page getting picked up over and over.
At the same time, low authority does not automatically mean low quality. We found small organizations that were clearly legitimate authorities in their niche, including CIRR.org, a nonprofit standards organization publishing audited bootcamp outcomes.
So I would not use a simple DR or referring-domain threshold to decide whether a citation source is valuable.
The more difficult question is whether some of the weaker, easily manufactured sources showing up today are temporary retrieval quirks.
My expectation is that at least some of them will eventually be filtered more aggressively as the engines improve. That is based on experience, not something this first study proves. In fact, some of the weak and heavily search-targeted sites are performing very well right now.
That is why we froze the experiment. We saved the exact prompt set and source cohorts so we can run the same test again later and see which sources actually persist.
What I would do right now
I would not optimize around citation count by itself, and I would not treat ChatGPT and Google AI Mode as the same citation ecosystem.
I would start with the sources that repeatedly matter in the categories my customers care about. Some will be obvious authorities. Others will be strong niche sites. And sometimes there will be a weak listicle that happens to be getting cited today.
You may be able to get your brand listed there, and it may help in the short term.
But before investing much time or money into it, I would ask a simple question:
Would I still want my brand mentioned on this website if ChatGPT stopped citing it tomorrow?
If the answer is yes, there is probably value beyond the current retrieval behavior. If the only reason the placement matters is that one engine happens to cite the page today, I would treat it very differently.
Methodology
We collected 24,960 responses from 2,080 prompts across 32 categories. Each prompt was run six times on ChatGPT Search and six times on Google AI Mode using a consistent U.S./English configuration.
After normalization and removal of non-editorial intermediary URLs, the dataset contained 102,366 citation events across 8,004 registrable domains. Source classification covered roughly 89% of citation volume, and traditional authority metrics came from DataForSEO backlink and organic-search data.
This is a single September 2026 snapshot, so it tells us what the engines were retrieving during this collection window, not which sources will remain important over time. That is what the next collection is designed to test.