We ran an AI search citation study on 25 questions. Here's what the answers told us.

Last reviewed November 18, 2026

AI search engines rarely agree on which sites to cite, and their picks change within days. We asked ChatGPT, Gemini, Perplexity, and Google’s AI Overviews the same 25 questions about content work, many times each, and recorded every site they cited:

  • Each engine cites mostly its own set of sites: two engines share at most 16% of them.
  • About 8 in 10 sites cited in Google’s AI Overviews also rank on Google’s first page for the same question. For ChatGPT, it’s under 1 in 10.
  • Google’s own AI engines disagree: Gemini and Google’s AI Overviews share only 16% of the sites they cite, and only about 3 in 10 of Gemini’s cited sites rank on Google’s first page.
  • Nine days later, each engine had dropped 42% to 68% of the sites it had cited for a question.
  • ChatGPT never searched the web for 8 of our 25 questions.

What we asked, and where

We asked 25 fixed questions about producing, checking, and organizing content with AI, in US English. ChatGPT, Gemini, and Perplexity answered each question in two rounds, on November 3 and November 12, 2026: 7 times in ChatGPT and Gemini and 5 times in Perplexity, 950 answers in total. On November 16, we searched each question 3 times in Google in a browser and recorded the AI Overview and the regular results: 74 searches. That makes more than 1,000 answers.

Engine Answers Searched the web (round 1 / round 2) Average sites cited per answer Different sites cited
ChatGPT 175 per round 60% / 62% 3.3 / 3.5 142 / 143
Gemini 175 per round 91% / 77% 12.5 / 10.2 498 / 340
Perplexity 125 per round 100% / 100% 10.0 / 10.0 203 / 198
Google AI Overviews 74 searches (November 16) AI Overview shown in 71 (96%) 6.8 134

Gemini and Perplexity cite about 10 sites per answer on average, Google’s AI Overviews about 7, and ChatGPT three to four.

Do AI engines cite the same sites?

AI engines rarely cite the same sites. For the same questions, each pair of engines shared between 5% and 16% of the sites they cited.

Engine pair Sites both engines cited, as a share of all sites either cited (round 1) Round 2
Gemini and Google AI Overviews 11% 16%
Gemini and Perplexity 10% 11%
Perplexity and Google AI Overviews 7% 9%
ChatGPT and Google AI Overviews 5% 8%
ChatGPT and Gemini 5% 7%
ChatGPT and Perplexity 6% 7%

For pairs with Google’s AI Overviews, round 1 uses AI Overviews collected through a data service on November 3, and round 2 uses the AI Overviews from our November 16 browser searches. Part of the change in those rows comes from the switch in how we collected them.

Even Google’s two engines differ. Gemini and Google’s AI Overviews were the closest pair, yet they shared only 11% of their cited sites in the first round and 16% in the second.

The engines don’t just pick different sites; they also lean on different kinds of sites. YouTube and Reddit showed up often in Google’s AI Overviews, and Reddit also in Gemini, but ChatGPT and Perplexity never cited either one.

Only 10 sites were cited by all four engines, among them ahrefs.com, developers.google.com, semrush.com, and zapier.com.

So “AI visibility” isn’t one number. A site can be a regular source in Perplexity and missing from ChatGPT’s answer to the same question, which means each engine has to be checked on its own.

A much larger study points the same way. Wellows’ overlap study looked at 22.7 million citations from five AI engines between January and June 2026 and found that 79.6% of the websites cited for a question appeared in only one engine. Gemini and ChatGPT shared just 6.1% of websites, the lowest of the pairs it compared, while Google’s own AI Overviews and AI Mode shared 24.5%. That study says it didn’t test how much answers vary from run to run, which is what our repeated runs measure.

Do AI engines cite the sites that rank in Google?

Google’s AI Overviews mostly cite sites that also rank in Google for the same question, and ChatGPT, Gemini, and Perplexity mostly don’t. About 8 in 10 sites cited in AI Overviews also appeared on the first page of Google’s regular results for the same question, against about 3 in 10 for Gemini and about 1 in 10 for Perplexity and ChatGPT.

Engine Cited sites that also ranked on Google’s first page
Google AI Overviews 78% (377 of 484)
Gemini 30% (262 of 867)
Perplexity 11% (130 of 1,168)
ChatGPT 9% (47 of 511)

ChatGPT, Gemini, and Perplexity are their November 12 answers, checked against the Google results from November 16.

Gemini is made by Google too, but here it behaves more like ChatGPT and Perplexity: most of the sites these three engines cited didn’t appear on the first page of Google.

For a content team, this means most of the sites AI Overviews cited also ranked on Google’s first page for the same question, so your Google rankings are a good first check there. They tell you little about ChatGPT, Gemini, or Perplexity, which mostly cite sites outside Google’s top results, so those engines need their own checks.

The overlap shows that rankings and AI Overview citations go together, not that a ranking causes a citation. So keep doing SEO for Google’s AI Overviews, and add GEO work on top of SEO for ChatGPT, Gemini, and Perplexity.

How much do AI citations change between runs?

AI citations change a lot from run to run in Gemini, Google’s AI Overviews, and ChatGPT, and hardly at all in Perplexity. 40% to 62% of the sites those three engines cited for a question appeared in only one of the 7 runs.

Engine Of the sites cited for a question: in only one run (round 1 / round 2) In all runs (round 1 / round 2)
Gemini 62% / 58% 1% / 1%
Google AI Overviews 49% / 51% 7% / 7%
ChatGPT 41% / 40% 13% / 8%
Perplexity 3% / 0% 94% / 98%

For a content team, that means a single check of a question can’t tell you whether you’re “in” an engine’s answer. In three of the four engines, asking the same question again changes many of the cited sites, so one check can show you as cited when you rarely are, or miss you when you usually are. Perplexity is the exception and gave almost the same sources every time we asked on the same day. How many times you need to ask, per engine, is worked out in our guide to tracking AI citations.

How much did AI citations change in a week?

Between November 3 and November 12, each engine stopped citing 42% to 68% of the sites it had cited for a question. ChatGPT cited 58% of its first-round sites again for the same question, Perplexity 52%, Google’s AI Overviews 42%, and Gemini 32%.

Engine Site and question pairs, round 1 (all 25 questions) Cited again in round 2 Share New in round 2
ChatGPT 173 100 58% 85
Perplexity 238 123 52% 113
Google AI Overviews 245 104 42% 149
Gemini 620 200 32% 228

Perplexity’s sources hardly vary within a day: 94% of its sites appeared in every run in the first round and 98% in the second. So most of its change between rounds is change over time. For the other three engines, part of the change comes from the run-to-run variation within each round, so their change over time is smaller than the table shows.

Gemini changed most: it searched in 77% of answers, down from 91%, and cited 340 different sites, down from 498.

Some patterns held. ChatGPT skipped the web for the same 8 questions in both rounds, and YouTube appeared in 73 AI Overviews in the first round and 74 in the second.

A site and question pair is one site cited for one question. It counts as cited again if the site appeared in any second-round run for the same question.

When does ChatGPT search the web?

In our runs, ChatGPT searched for questions about tools, Google’s policies, and GEO, and skipped the web for most how-to questions about quality and teams. Overall, it searched in about 6 of 10 answers: 60% in the first round and 62% in the second. OpenAI says ChatGPT searches automatically when a question might benefit from information on the web. For 8 of our 25 questions, it didn’t search in any of the 7 runs, in either round. In the first round it also skipped the web search for one question, which got a search in the second round.

When ChatGPT doesn’t search, it answers from what its model learned in training and cites no sites. In our runs, these 8 questions got no citations in ChatGPT at all, so no page, new or old, could be cited there for them.

Searching doesn’t guarantee a citation either. In the 88 answers per round where we could see which pages ChatGPT retrieved, it retrieved about 16 pages per answer in the first round and 17 in the second, but cited the sites of only 37% and 39% of them. In other words, about two in three pages ChatGPT retrieved while answering never appeared as a source. Being found by ChatGPT and being cited by it are two different results.

Which sites does each engine cite?

Reddit was the most-cited site, followed by HubSpot’s blog, Semrush, Google’s developer documentation, and Medium, but each engine leans on a different set. These counts cover ChatGPT, Gemini, and Perplexity in both rounds and the 71 AI Overviews from November 16.

Site Answers citing it ChatGPT Gemini Perplexity Google AI Overviews
reddit.com 105 0 72 0 33
blog.hubspot.com 80 6 4 70 0
semrush.com 77 30 31 10 6
developers.google.com 69 43 13 10 3
medium.com 69 0 47 0 22
youtube.com 61 0 5 0 56
searchengineland.com 44 12 7 25 0
ahrefs.com 43 9 11 20 3
zapier.com 43 10 15 15 3
dev.to 43 0 3 40 0
jasper.ai 40 15 23 0 2

YouTube ranks lower here only because we have 71 AI Overviews against 950 chat answers; it appeared in most AI Overviews. Most sites appear rarely: 34% of the 1,033 sites cited appeared in just one answer. For a single question, Gemini cited between 7 and 36 different sites across its runs in the first round.

Google’s AI Overviews cite YouTube and Reddit

YouTube appeared in about 8 in 10 AI Overviews, and Reddit in almost half.

Platform AI Overviews citing it (of 71)
YouTube 56 (79%)
Reddit 33 (46%)
Medium 22 (31%)
Facebook 18 (25%)
Quora 14 (20%)
LinkedIn 13 (18%)

We measured which platforms AI Overviews cite, not whether a brand’s own videos or posts there get cited.

Gemini cites Reddit and Medium

Gemini cited Reddit in 72 of its 350 answers and Medium in 47, more than any company’s own site. Semrush (31) and Jasper (23) followed. Like Google’s AI Overviews, Gemini mixes community platforms with articles, but it almost never cited YouTube.

Perplexity cites HubSpot’s blog and developer platforms

Perplexity cited HubSpot’s blog in 70 of its 250 answers, by far its favorite source, followed by dev.to (40), deepsmith.ai (30), microsoft.com (26), upwork.com (25), and tryprofound.com (25). It never cited YouTube, Reddit, or Medium.

ChatGPT cites Google’s documentation and SEO tool companies

ChatGPT’s most-cited site was developers.google.com, in 43 of its 350 answers, followed by surferseo.com (30), semrush.com (30), and frase.io (28). Like Perplexity, it never cited YouTube, Reddit, or Medium.

What kind of articles get cited?

The cited articles matched the type of question. For the three questions about choosing tools, 41% of the cited pages were “best” or “top” lists (543 of 1,312) and 26% were comparisons (341). For the nine how-to questions, 42% were guides (1,378 of 3,292) and only 7% were lists. Across all 25 questions, guides were the largest group we could classify, at 23% of cited pages.

Tool companies got cited mostly for guides and research, not for lists of tools. Semrush, Ahrefs, Frase, Surfer, Jasper, and HubSpot had 394 pages cited, and only 19 of them were “best” or “top” lists. The rest included how-to guides and data studies, such as a Semrush study on whether AI content ranks.

On a question about Google’s own rules, the official source won. When we asked “Does Google penalize AI-generated content?”, ChatGPT cited developers.google.com in all 7 runs of the first round, ChatGPT, Gemini, and Perplexity cited it in every run of the second, and Google’s AI Overviews cited it in all 3 of our November 16 searches.

For a content lead, this means the format should match the question: in our data, lists and comparisons made up two-thirds of the pages cited for tool questions, and guides made up the largest share for how-to questions.

How to check and improve your AI citations

Start by measuring where you stand, then match your content to what each engine cites:

  1. List the 10 to 20 questions that matter most to your business, in the words your buyers would use with an AI tool.
  2. Ask each question in ChatGPT, Gemini, Perplexity, and Google’s AI Overviews three to five times per engine, on the same day, logged out, and in a new chat each time, as we did. For 20 questions that’s 240 to 400 checks, so a team of one may start with 10 questions or use a tracking tool.
  3. Record which sites each engine cites, and the type of page, such as list, guide, comparison, video, or forum thread. Count how many of the runs cite your site: once in 5 runs is luck, 4 in 5 is a position worth protecting.
  4. Check whether the sites cited in Google’s AI Overviews also rank in Google for the same question. In our data, about 8 in 10 did, against 3 in 10 or fewer for the other engines, so for AI Overviews check your Google rankings for those questions first.
  5. Match the format to the question. Write researched lists and comparisons for “which tool” questions and guides for “how to” questions, and fill the gaps where competitors’ pages get cited and yours don’t.
  6. Plan YouTube and Reddit content for the questions where Google’s AI Overviews or Gemini cite those platforms. Our guide to GEO for content teams covers that work as part of a wider plan.
  7. Repeat the check every month, or every week while you learn how fast your answers change, and compare each round with the last.

What we changed because of it

We run Content Runbook on the same data, so the findings changed our own plan:

  • We aim our quality and team pages at Google, Gemini, and Perplexity first, and don’t track ChatGPT for those pages until a later round shows it searching for them, because it skipped the web for 6 of the 10 quality and team questions in both rounds.
  • Our page on whether Google penalizes AI content quotes and links Google’s own documentation, because every engine cited Google’s documentation for that question.
  • We check citations every week until early December, then monthly, because each engine dropped 42% to 68% of its cited sites for a question within nine days.
  • Brand accounts on Reddit and LinkedIn stay in our plan, because both show up in Google’s AI Overviews for our topics.
  • We’re deciding whether to start a YouTube channel sooner than planned, because YouTube appeared in about 8 in 10 AI Overviews.

How we ran the study

We asked the same 25 questions in every round. We wrote them to match what content teams ask about AI, across six topics, and we don’t change their wording between rounds, so the rounds stay comparable.

# Question Topic
1 What’s the best workflow for producing blog content with AI in a small marketing team? Workflows
2 How should an agency set up a human-in-the-loop process for AI-written content? Workflows
3 How do I write a content brief that gets good results from AI writing tools? Workflows
4 What is content operations, and how does AI change it? Workflows
5 How do I scale content production with AI without losing quality? Workflows
6 Give me a checklist for reviewing AI-generated articles before publishing. Quality
7 How do I fact-check content written by AI? Quality
8 How do I create a style guide that AI writing tools can follow? Quality
9 How can I measure the quality of AI-assisted content? Quality
10 What are the most common mistakes in AI-generated blog posts? Quality
11 What does an AI content editor do? Teams
12 How should I restructure my content team now that we use AI? Teams
13 How much does it cost to produce an article with AI compared to a freelance writer? Teams
14 How many articles a month can a five-person content team produce with AI? Teams
15 What should an AI usage policy for a content team include? Teams
16 What are the best AI writing tools for content agencies? Tools
17 Which AI tools help with SEO content briefs? Tools
18 Claude or ChatGPT: which is better for writing long-form content? Tools
19 Are AI content detectors accurate? Tools
20 Does Google penalize AI-generated content? AI search
21 How do I get my content cited by ChatGPT and Perplexity? AI search
22 What is generative engine optimization, and how is it different from SEO? AI search
23 How do I track whether AI search engines cite my website? AI search
24 Is there a good free template for an AI content brief? Templates
25 What should a content SOP for AI-assisted writing include? Templates

Engines. We collected ChatGPT and Gemini answers from their logged-out web apps and Perplexity answers through its API, using the Sonar model, all through DataForSEO. Every question ran in US English, 7 times in ChatGPT and Gemini. Perplexity got 5 runs to save cost, since it costs more per answer, and its sources turned out almost identical across runs.

Google. On November 16, we searched each question 3 times in Google in a browser, from the US, with results not personalized, and recorded the AI Overview’s sources and the first page of regular results. We left out one search that loaded right after Google blocked the browser, which leaves 74. We first collected AI Overviews through DataForSEO too, but its Google pages didn’t match what searchers in the US see, so every other AI Overviews and Google ranking figure on this page comes from the browser check. The run-to-run, nine-day, and article-type figures for AI Overviews, and the first-round engine overlap, still come from the DataForSEO collection of November 3 and 12, because one day of browser searches can’t measure them.

Counting. Site counts are by site, not page, and a site counts once per answer however many times it’s linked. Article types and ChatGPT’s retrieved pages are counted by page. Links to m.youtube.com count as youtube.com. When ChatGPT answers without searching, its app lists its own home page as a source, and we leave that out. AI Overviews show site names rather than addresses, so we matched names to sites and left out 32 of 613 sources we couldn’t match.

Google rankings. For each site an engine cited, we checked whether it appeared among the regular results on the first page of Google for the same question, in our November 16 browser searches. For AI Overviews, we used the results on the same page as the overview; for the other engines, any of the 3 searches for that question.

Google’s redirect links. In the November 3 and 12 collection, AI Overviews link some sources through a google.com address that hides the site: 18 links in the first round, 189 in the second. We recovered the site by matching the link’s title to a regular result on the same page, to the same title elsewhere in our data, or to the source name Google shows next to the link. We couldn’t recover 2 links in the first round and 30 in the second, and left them out of the counts.

Article types. We sorted the cited pages by words in their titles, such as “best” or “top” for lists, “vs” for comparisons, and “how to” or “guide” for guides. Titles that matched no type, 38% of cited pages, stay unclassified.

Cost. The first round cost $3.17, including $0.28 for failed AI Overview requests that we ran again. The second cost $2.88. The browser searches cost nothing extra.

Who ran it. Content Runbook, an independent publication. We don’t sell an AI visibility tool or a writing tool, and no company paid for this study.

Data. Download the per-site counts for each engine: first round, November 3, 2026 (CSV) and second round, November 12, 2026 (CSV). The first-round file predates our November 6 recount, so it still shows YouTube at 72 of 175 AI Overviews; the corrected count is 73 of 173 (see corrections below).

How to cite this page. Content Runbook, “AI search citation study,” November 2026, https://contentrunbook.com/ai-search-citation-study/.

What this data can’t tell you

This is two rounds nine days apart, so it shows change between two dates, not a trend. Our Google figures come from one day of browser searches, 3 per question. We’re adding a round every week until early December, then one a month, which will show whether ChatGPT keeps skipping the same questions and whether the rate of change holds.

The questions are 25 short, fixed questions we wrote, in US English. We didn’t test longer prompts, follow-up questions, or other countries and languages, which may give different answers. Logged-out apps and Perplexity’s API can also cite differently from the apps people use while signed in.

“Cited again” counts a site if it appears in any run of the next round, so for the engines with a lot of run-to-run variation, part of the change between rounds is that variation.

Recovering the sites behind Google’s redirect links is a best match by title or source name, and 30 second-round links stay unknown. Sorting articles by title words is rough, and more than a third of cited pages fit no type.

Site counts are by site, so two different articles from the same site count as one; only the article types and ChatGPT’s retrieved pages are counted by page.

Corrections and the next round

  • November 3, 2026: corrected on the day of the first round after re-checking the final data. ChatGPT never searched for 9 questions, not 10.

  • November 6, 2026: YouTube recounted as 73 of 173 AI Overviews, not 72 of 175. The first count missed m.youtube.com links and included 2 searches without an AI Overview.

  • November 12, 2026: Google’s redirect links are now resolved to the real site. The share of AI Overview citations that also rank in Google went from 61% to 63%, and from 50% to 51% on the strict reading. Reddit went from 36 to 37 AI Overviews, and Quora from 13 to 16.

  • November 18, 2026: AI Overviews and Google ranking figures replaced with a browser check of Google on November 16. The earlier figures came from a data service whose Google pages didn’t match what searchers in the US see. Sites cited in AI Overviews that also rank on Google’s first page went from about 6 in 10 to about 8 in 10, YouTube from about 4 in 10 AI Overviews to about 8 in 10, Reddit from about 1 in 5 to almost half, and the most two engines share from 12% to 16%.

The next round runs next week, with the same questions and settings, so the rounds stay comparable. We update this page after each round and log every correction here.