# Which Sources Each AI Engine Cites, and What to Do About It > ChatGPT, Perplexity and Google lean on different sources. What the citation data shows, how old it is, and where to put your effort first. URL: https://llmagnet.com/blog/sources-each-ai-engine-cites Author: Ben Cohen, Founder & CEO at LLMagnet Published: 2026-10-03 Updated: 2026-11-10 Category: Research ChatGPT, Perplexity and Google’s AI features lean on different sources, so work aimed at one engine won’t automatically carry over to the others. In a vendor test of 100,000 prompts, only 11% of domain citations appeared in both ChatGPT and Perplexity. Treat that as a direction, not a constant: most of the data below is from 2025, and engines change. ## Contents 1. What the 11% Figure Measures 2. The Top Sources, Side by Side 3. ChatGPT: Reference Sources and Brand Mentions 4. Perplexity: Reddit, Reviews and Recent Pages 5. Google AI Mode and AI Overviews: Wikipedia, YouTube, Reddit 6. Grok and Claude: Thin Data 7. Which Sources Fit Your Business 8. A Worked Example: Reading One Audit 9. Common Mistakes 10. Where to Put Your Effort First 11. How to Measure It 12. FAQ ## What the 11% Figure Measures [Profound](https://www.tryprofound.com/blog/citation-overlap-strategy), which sells AI visibility tracking, ran 100,000 prompts through ChatGPT and Perplexity. It reported 37.4% of domain citations as ChatGPT only, 51.6% as Perplexity only and 11% as appearing in both. The article is from July 2025. It doesn’t give the collection period or define overlap, and it doesn’t say whether it counted unique domains or total citations. ![Share of domain citations across 100,000 prompts: Perplexity only 51.6 percent, ChatGPT only 37.4 percent, cited by both 11 percent.](https://llmagnet.com/images/blog/charts/engine-sources-overlap.png) *Source: Profound, July 2025. Vendor data.* Domain overlap isn’t answer overlap. Two engines can cite different sites and still name the same brands, so the figure tells you about sources, not about who gets recommended. The same article gives overlap for other pairs, from 6% between Google AI Overviews and Microsoft Copilot to 16.4% between Perplexity and AI Overviews. So the low ChatGPT and Perplexity number isn’t an odd result from one pairing. Why the pools differ is mostly inference, because engines don’t publish their retrieval weights. One difference is measurable. In [Ahrefs’ study of 15,000 long-tail prompts](https://ahrefs.com/blog/ai-search-overlap/) (vendor data, published August 2025), 28.6% of the URLs Perplexity cited ranked in Google’s top 10 for the same prompt, against about 8% for ChatGPT’s in-text citations. Ahrefs suggests Perplexity’s focus on citations may explain the closer match, and notes that it runs its own crawler-based index. The engines also offer different numbers of places to appear. Profound counted the domains cited in an average response. ![Average domains cited per response: Google AI Overviews about 7.7, Perplexity about 7.3, ChatGPT about 5.0, Microsoft Copilot about 2.5.](https://llmagnet.com/images/blog/charts/engine-sources-slots.png) *Source: Profound, July 2025. Vendor data.* Fewer places make a harder contest. If your category’s Copilot answers cite two or three sites, your own page may struggle to be one of them, and a listing on a site that’s already cited could be the more realistic target. ## The Top Sources, Side by Side [Profound’s analysis of 680 million citations](https://www.tryprofound.com/blog/ai-platform-citation-patterns) (vendor data, August 2024 to June 2025) covers ChatGPT, Google AI Overviews and Perplexity. It reports two different numbers for each leading source, and they are easy to confuse. One is a share of the engine’s ten most-cited sources. The other is a share of all its citations. This table uses the second, because that’s the one that tells you how much of the pool a site really holds. | Rank | ChatGPT | Perplexity | Google AI Overviews | | --- | --- | --- | --- | | 1 | Wikipedia, 7.8% | Reddit, 6.6% | Reddit, 2.2% | | 2 | Reddit, 1.8% | YouTube, 2.0% | YouTube, 1.9% | | 3 | Forbes, 1.1% | Gartner, 1.0% | Quora, 1.5% | | 4 | G2, 1.1% | Yelp, 0.8% | LinkedIn, 1.3% | | 5 | TechRadar, 0.9% | LinkedIn, 0.8% | Gartner, 0.7% | ![Each engine’s top source as a share of all its citations: Wikipedia 7.8 percent of ChatGPT citations, Reddit 6.6 percent of Perplexity citations, Reddit 2.2 percent of Google AI Overview citations.](https://llmagnet.com/images/blog/charts/engine-sources-leaders.png) *Source: Profound, 680 million citations, August 2024 to June 2025. Vendor data.* Three things follow from these lists: - **The long tail is most of it.** Even Wikipedia on ChatGPT is under 8%. More than 90% of each engine’s citations go somewhere other than its top source. - **A source can matter on one engine and barely on another.** Wikipedia is 7.8% of ChatGPT’s citations and 0.6% of AI Overviews’. - **Only three sites make all three top-10 lists:** Reddit, Forbes and NerdWallet. Profound’s overlap article advises starting with domains cited across several engines. In practice that points to community discussion and editorial coverage, with the caveat that NerdWallet is a finance site and your category will have its own equivalents. Profound doesn’t say how its prompts were chosen or how they split by topic, so don’t treat these lists as your category’s lists. Treat them as a reason to go and check. ## ChatGPT: Reference Sources and Brand Mentions Wikipedia made up 47.9% of the citations among ChatGPT’s ten most-cited sources, but 7.8% of all ChatGPT citations. The first number is a share of a top-10 list. Don’t read it as nearly half of everything ChatGPT cites. Being named and being linked are also different outcomes. [BrightEdge](https://www.brightedge.com/resources/weekly-ai-search-insights/chatgpt-brand-mentions-vs-citations-what-triggers-visibility) (vendor data, published August 2025, “thousands” of prompts) found ChatGPT produced 2.4 brand mentions per prompt and 0.74 citations, a ratio of about 3.2 to 1. In 44% of prompts it named no brand at all. “Best” queries averaged 4.8 mentions, about twice the baseline, but a quarter of them still named none. So check commercial prompts separately from informational ones, or the second group will drag your average down and hide what’s happening. What to do this week: - **Check your reference entries.** If your company has a Wikipedia article or Wikidata item, correct errors through the proper process, and disclose your connection. [Wikipedia’s conflict of interest guideline](https://en.wikipedia.org/wiki/Wikipedia:Conflict_of_interest) strongly discourages editing an article about your own company. Post the correction on the article’s talk page with the `{{edit COI}}` template and a reliable source. If you’re paid to do it, you must say who pays you. Don’t write your own article. - **Pitch the trade publications your buyers read.** Reference-style and editorial sources are where a brand gets described in third-party words. - **Track mentions and citations separately.** A brand can be named often and linked rarely, and the two need different fixes. ## Perplexity: Reddit, Reviews and Recent Pages The same Profound dataset shows Reddit at 46.7% of citations among Perplexity’s top ten sources and 6.6% of all its citations. Reddit is important, but it isn’t half of what Perplexity cites, and most citations go elsewhere. Perplexity also leans toward newer pages. In [Ahrefs’ study of about 17 million cited URLs](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content) (vendor data, published July 2025), the pages it cited were on average 250 days newer than organic results, counting age from when Ahrefs’ crawler first saw each page. ChatGPT’s lean was stronger, and the top three citations in Google AI Overviews were slightly older than organic results. ![Average age of cited pages in days, from when Ahrefs’ crawler first saw each page: Google AI Overviews top three citations 1,432, organic results 1,416, Perplexity 1,166, Gemini 1,118, Copilot 1,056, ChatGPT in-text references 1,023, ChatGPT citations 958.](https://llmagnet.com/images/blog/charts/engine-sources-age.png) *Source: Ahrefs, July 2025, per-platform table. Vendor data. Age is counted from when Ahrefs’ crawler first saw the page. AI Overviews covers its top three citations only.* Read the scale before you act on this. The newest group is still over two and a half years old on average, so a useful older page isn’t disqualified anywhere. We’d treat freshness as a tie-breaker on topics that change, not a reason to rewrite everything. What to do this week: - **Find the threads.** Run your category’s buying questions on Perplexity and note which subreddits and forums appear. Answer real questions there, say who you work for, and follow each community’s rules. Planted posts can be removed and can damage the brand. - **Fill out review profiles** on the sites Perplexity shows for your category, such as G2, Capterra or Trustpilot. - **Update pages when something has changed.** Add new data or corrected details. Changing the date alone isn’t an update. ## Google AI Mode and AI Overviews: Wikipedia, YouTube, Reddit For AI Overviews, Profound’s data puts Reddit at 21.0% and YouTube at 18.8% of citations among the top ten sources. As shares of all AI Overview citations, those fall to 2.2% and 1.9%. For AI Mode, [Storyboard18 reported Ahrefs data](https://storyboard18.com/brand-marketing/ai-modes-top-100-sources-revealed-wikipedia-youtube-google-dominate-citations-80373.htm) from 5.5 million queries. The September 2025 article doesn’t say when the queries were collected. Its top-100 list put Wikipedia first, followed by YouTube, blog.google and Reddit. [Ahrefs’ current table](https://ahrefs.com/blog/most-cited-domains-ai-mode) (vendor data, US queries, November 2026) ranks google.com first by mention share, then Reddit, YouTube, Facebook and Wikipedia. Treat any order as specific to its dataset. Google says there are [no additional requirements](https://developers.google.com/search/docs/appearance/ai-features) to appear in AI Overviews or AI Mode, and that no special schema is needed. Overlap with Google’s results is also higher here, though the studies differ in method: in [Ahrefs’ study of 863,000 keyword SERPs](https://ahrefs.com/blog/ai-overview-citations-top-10) (vendor data, published March 2026), 37.9% of AI Overview citations also appeared in the top 10 SERP blocks for that query. What to do this week: - **Pick one or two topics that benefit from being shown,** such as setup or comparisons, and publish a video. Put an accurate transcript on your own site. - **Do the ordinary search work.** Clear authorship, cited sources, accurate dates and crawlable pages are ordinary search hygiene. - **For local businesses,** make sure your Google Business Profile is complete and correct. ## Grok and Claude: Thin Data We couldn’t verify comparable citation-share data for Grok or Claude, so we won’t describe a pattern for either. Test them directly with your own prompts. The best-known accuracy test is about citing news, not brands. The [Tow Center at Columbia Journalism Review](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php) ran 1,600 queries across eight tools in February 2025, asking each to identify the source of an article excerpt. Grok 3 answered 94% incorrectly and Perplexity 37%. Out of the 200 prompts tested on Grok 3, 154 citations led to error pages. Those tools have changed since, so treat the numbers as history. The lesson still holds: open the link before you count a citation. ## Which Sources Fit Your Business The table below is our own reading of Profound’s full top-10 lists (the table above shows the first five of each), plus Google Business Profile for local businesses. It isn’t a measured result, and Profound didn’t publish data by industry. Use it to decide what to look for first when you check your own prompts. | If you are | Check these first | A first move | | --- | --- | --- | | **B2B software** | G2, Gartner, TechRadar, Forbes, LinkedIn | Complete your G2 profile and ask recent customers for reviews | | **Consumer ecommerce** | Reddit, YouTube, PCMag, TechRadar | Send product to reviewers who already rank in your category | | **Local or travel** | Yelp, TripAdvisor, Google Business Profile | Correct your listings, then answer open reviews | | **Finance** | NerdWallet, Forbes, Business Insider | Find which comparison pages list your competitors and not you | Size changes the plan as well. A team of one or two should pick a single source type on a single engine and stay on it for a quarter. A larger team can run one workstream per source type, but should still report by engine, so a gain on Perplexity can’t hide a loss on ChatGPT. ## A Worked Example: Reading One Audit The company and figures below are made up for illustration. Fernhollow sells scheduling software to dental clinics. It runs 25 buyer prompts on three engines and records what it finds. | Engine | Named in | Linked in | Most common sources cited instead | | --- | --- | --- | --- | | ChatGPT | 6 of 25 | 1 of 25 | Wikipedia, a dental trade magazine | | Perplexity | 1 of 25 | 1 of 25 | Reddit threads, G2, Capterra | | AI Overviews | 3 of 25 | 2 of 25 | Competitor pages, YouTube demos | Each row points to a different job. ChatGPT names Fernhollow but rarely links it, so the brand is known and the gap is in sources that describe it. Perplexity barely knows it, and the replacements are forums and review sites. On AI Overviews the competition is other vendors’ pages, which makes it mostly ordinary page work. Sales calls show that most prospects mention Perplexity and Google. So Fernhollow starts with its G2 and Capterra profiles, then rewrites its two comparison pages. It leaves Wikipedia alone, since it doesn’t have an article and isn’t notable enough to earn one yet. ## Common Mistakes - **Quoting the top-10 share as the total.** It happens because the larger number makes a better headline. It sends budget to one site on the belief that it holds half the pool. - **Copying someone else’s source list.** General lists pool every kind of prompt. A cited source for consumer travel tells you little about procurement software. - **Testing only on the engine the team uses.** People check the assistant they already have open. Your buyers may be on another one. - **Counting a citation without opening it.** Some links are dead or point to the wrong page, and they inflate your count. ## Where to Put Your Effort First We don’t have reliable data on which engine each industry’s buyers use, and a general rule would hide that. Measure your own gap, then match the work to it. ![Four steps: write 20 to 30 buyer prompts, run them on each engine, read the sources cited instead of you, then start with the source type that appears most often on the engine your customers use.](https://llmagnet.com/images/blog/charts/engine-sources-plan.png) *The order we recommend. It’s our own process, not a measured result.* 1. **Write 20 to 30 prompts** that real buyers ask, including comparisons and “best for” questions. 2. **Run them on each engine** and record whether you are named, linked, or absent. 3. **Read the sources that are cited instead of you.** That list tells you where to work: Reddit threads, review sites, trade press, video or your own pages. 4. **Start with the source type that appears most often** on the engine your customers mention in sales calls. ## How to Measure It One engine’s result doesn’t stand in for the others. [Zen Media](https://martechedge.com/news/zen-media-finds-ai-visibility-can-vary-sharply-across-ai-engines) (vendor data) analyzed three 2026 AI visibility campaigns and reported that one brand appeared in 15.5% of prompts on its weakest engine and 59.5% on its strongest, across ChatGPT, Gemini, Claude, Perplexity and Grok. The report calls these figures campaign-specific. One check doesn’t stand for the month either. [Ahrefs tracked 43,000+ keywords](https://ahrefs.com/blog/ai-overview-change) for a month (vendor data, published November 2025) and found AI Overview text differed from the previous observation about 70% of the time, with 45.5% of cited URLs new each time. Ahrefs set no threshold for how big a difference had to be, and the meaning stayed similar (cosine similarity of 0.95), so expect the citations to move more than the message does. Run the same prompts weekly for at least four weeks, and compare averages rather than single results. For each engine, keep three numbers: the share of prompts where you’re named, the share where you’re linked, and the five domains cited most often in your place. The third is the one that tells you whether your work is reaching the right sources. Believe a change only if it’s larger than the swing you saw between repeat runs of the same prompts. LLMagnet measures AI visibility across seven engines and has an agent that prepares fixes for you to approve. Its [free audit](https://ai-visibility.llmagnet.com) is one place to start. ## FAQ ### How much do ChatGPT and Perplexity citations overlap? In a Profound test of 100,000 prompts, published in July 2025, 11% of domain citations appeared in both engines, 37.4% in ChatGPT only and 51.6% in Perplexity only. It is vendor data and measures domains, not whether the same brands are recommended. ### Does Wikipedia make up almost half of ChatGPT’s citations? No. In Profound’s analysis of 680 million citations, Wikipedia was 47.9% of citations among ChatGPT’s ten most-cited sources, but 7.8% of all ChatGPT citations. ### How much does Perplexity rely on Reddit? Profound’s data from August 2024 to June 2025 put Reddit at 46.7% of citations among Perplexity’s top ten sources and 6.6% of all its citations. Reddit matters, but most of Perplexity’s citations go elsewhere. ### What does Google AI Mode cite most? Ahrefs data from 5.5 million AI Mode queries, reported by Storyboard18 in September 2025, ranked Wikipedia first, YouTube second, blog.google third and Reddit fourth. Ahrefs’ November 2026 US table ranks google.com first by mention share, then Reddit, YouTube, Facebook and Wikipedia. Treat any order as specific to its dataset. ### Does ChatGPT mention brands without linking to them? Yes. BrightEdge reported 2.4 brand mentions and 0.74 citations per ChatGPT prompt in August 2025, about 3.2 mentions per citation. It is vendor data from thousands of prompts. ### Can one GEO strategy cover every engine? Partly. Clear facts, useful pages and third-party coverage are sensible work on any engine, but the sources each engine favors differ. Test the same buyer prompts on every engine and work on the sources cited in place of you.