How AI engines search and cite: 960 answers measured

Prefer's study of 960 AI answers: how often ChatGPT, Perplexity, Gemini and Claude search the web, what they cite, and how little they agree. Free to cite.

Javed Khatri Javed Khatri Co-founder, Prefer

9 min read Data study

Cover for "How AI engines search and cite: 960 answers measured"

The short answer

How do ChatGPT, Perplexity, Gemini and Claude search and cite?

Prefer's September 2026 study of 960 AI answers asked ChatGPT, Perplexity, Gemini and Claude the same 80 questions about AI search, three times each, on 13 September. All four searched the live web for most answers, from Perplexity's 100% to Claude's 80%. Perplexity cited 19.48 sources per answer and ChatGPT 3.05. Most of the 1,329 sites they cited came from one engine only, but the most-cited sites were shared: 23 of the top 25 were cited by three or four engines.

Key takeaways

  • In Prefer's study, all four engines searched the live web for most answers: Perplexity for 240 of 240, ChatGPT 216 (90%), Gemini 213 (88.8%) and Claude 192 (80%).
  • Perplexity cited 19.48 sources per answer, 6.4 times ChatGPT's 3.05. Gemini cited 8.64 and Claude 4.58.
  • Of 1,329 sites cited across 960 answers, 966 (72.7%) were cited by one engine only and 29 (2.2%) by all four. Among the 25 most-cited sites, 23 were cited by three or four engines.
  • Two runs of the same question shared, on average, 37% of the sites cited across both on ChatGPT and 35% on Gemini, against 96% on Perplexity.
  • Reddit was cited in more answers than any other site (207 of 960), but Claude cited it in none of its 240, and all 113 answers citing YouTube came from Gemini.
  • Google showed an AI Overview for 75 of 80 questions. YouTube was its most-cited site (38 Overviews), while Reddit led its organic first page (61 questions).

This page publishes every number from Prefer's September 2026 AI answer study. We asked four AI engines the same 80 questions about AI search, GEO and AEO, three times each, and recorded whether each answer searched the web and every source it cited. For the same questions we also captured Google's AI Overview, its AI Mode answer and its first page of organic results. The data is free to reuse under CC BY 4.0, and the full dataset is downloadable at the end of the page. This is the September 2026 edition.

The results by engine

Every engine searched the web for most answers, but they cited very different numbers of sources, and they rarely cited the same sites. Each engine answered 240 times (80 questions, 3 runs each).

Engine Searched the web (of 240) Searched on all 3 runs (of 80) Sources per answer Different sites cited Site overlap between runs
ChatGPT 216 (90%) 70 3.05 218 37%
Perplexity 240 (100%) 80 19.48 748 96%
Gemini 213 (88.8%) 63 8.64 580 35%
Claude 192 (80%) 58 4.58 333 68%

We asked for web search on every call. Each engine still decides whether to use it, and the API reports whether it did. Perplexity searched for all 240 of its answers, ChatGPT for 216 (90%), Gemini for 213 (88.8%) and Claude for 192 (80%). Across the four engines, 861 of 960 answers (89.7%) came from a live search. None of the other 99 answers cited a single source.

Bar chart of the share of answers where each engine searched the live web, September 2026, 240 answers per engine: Perplexity 100% (240 of 240), ChatGPT 90% (216 of 240), Gemini 89% (213 of 240), Claude 80% (192 of 240). Side panels contrast answers built from a live search with answers from memory.
Prefer AI answer study, September 2026: 80 questions about AI search and GEO, 3 runs per engine, web search requested on every call. Gemini's rate is 88.8%, shown rounded.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/geo-live-search-rate.png" alt="Bar chart of the share of answers where each engine searched the live web, September 2026, 240 answers per engine: Perplexity 100% (240 of 240), ChatGPT 90% (216 of 240), Gemini 89% (213 of 240), Claude 80% (192 of 240). Side panels contrast answers built from a live search with answers from memory." width="780" height="502" style="max-width:100%;height:auto" />
  <figcaption>AI engine live search rate chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

In the "What does GEO mean?" topic, mostly definition questions, ChatGPT searched for 11 of its 24 answers and Claude for 13, their lowest rates in any topic. Gemini searched for 16 of 24 there and least in the "LLM SEO" topic (15 of 24). In two topics, "Best AEO and GEO tools" and "AI search statistics", all four engines searched for all 192 answers.

The choice to search was not fixed either. Asked the same question three times, ChatGPT searched on all three runs for 70 of 80 questions and on none of them for 6; on the other 4, its runs disagreed. Perplexity searched on every run. Gemini searched on all three runs for 63 questions and Claude for 58, and each disagreed with itself on 13 questions. Counting "never searched" as a consistent choice too, the same choice held on all three runs for 76 of ChatGPT's questions (95%), 80 of Perplexity's (100%), and 67 of Gemini's and 67 of Claude's (83.8% each). For how ChatGPT decides, see does ChatGPT use live search.

How many sources does each engine cite?

Perplexity cited 19.48 sources per answer, 6.4 times ChatGPT's 3.05. Gemini cited 8.64 and Claude 4.58. These averages include answers that cited nothing. Counting only answers that searched, the figures are 19.48 for Perplexity, 9.74 for Gemini, 5.73 for Claude and 3.38 for ChatGPT.

Across its 240 answers, Perplexity cited 748 different sites, Gemini 580, Claude 333 and ChatGPT 218. Claude gave 48 answers with no sources at all, ChatGPT and Gemini 27 each, and Perplexity none. Of the ChatGPT answers that searched, 3 still cited nothing.

For a page owner, that is the difference between one of about three source slots in a ChatGPT answer and one of about nineteen in a Perplexity answer. Our answer on how many sources ChatGPT cites covers a separate run of buyer questions.

Two bar charts from 240 answers per engine, September 2026. Average sources cited per answer: Perplexity 19.48, Gemini 8.64, Claude 4.58, ChatGPT 3.05. Different sites cited across all 240 answers: Perplexity 748, Gemini 580, Claude 333, ChatGPT 218.
Prefer AI answer study, September 2026: 80 questions, 3 runs per engine. Averages include answers that cited none.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/geo-sources-per-answer.png" alt="Two bar charts from 240 answers per engine, September 2026. Average sources cited per answer: Perplexity 19.48, Gemini 8.64, Claude 4.58, ChatGPT 3.05. Different sites cited across all 240 answers: Perplexity 748, Gemini 580, Claude 333, ChatGPT 218." width="780" height="645" style="max-width:100%;height:auto" />
  <figcaption>Sources per AI answer chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

Do the engines cite the same sites?

It depends on how you count. Most sites were cited by one engine only, but the most-cited sites were cited by three or four. Of 1,329 different sites cited across the 960 answers, 966 (72.7%) were cited by one engine only, 205 (15.4%) by two, 129 (9.7%) by three and 29 (2.2%) by all four.

Many of the single-engine sites were minor. 338 of the 966 were cited in a single answer, so no second engine could have cited them. Of the 959 sites cited in fewer than five answers, 890 came from one engine. By prominence the picture changes: 23 of the 25 most-cited sites were cited by three or four engines, and 5 of the top 10 by all four (Semrush, Search Engine Land, Ahrefs, Profound and arXiv). Of the 170 sites cited in ten or more answers, 116 were cited by three or four engines and 14 by one.

Question by question the overlap is thinner. On 69 of the 80 questions, no site was cited by all four engines. On each of the other 11, exactly one site was. Reddit and YouTube, the first and fourth most-cited sites, are not among the 29: Claude never cited Reddit, and only Gemini cited YouTube. The 29 include ahrefs.com, semrush.com, searchengineland.com, g2.com and wikipedia.org, and the full list is in the download.

Bar chart of 1,329 sites cited across 960 answers from ChatGPT, Perplexity, Gemini and Claude, grouped by how many engines cited each one: 966 by one engine only (72.7%), 205 by two engines (15.4%), 129 by three engines (9.7%), and 29 by all four engines (2.2%). On 69 of the 80 questions, no site was cited by all four engines. In the July 2026 study, 5 of 713 domains were cited by all four engines across 47 buyer questions.
Prefer AI answer study, September 2026: 80 questions, 4 engines, 3 runs, 960 answers. Each site is counted once across the whole study.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/geo-engine-overlap.png" alt="Bar chart of 1,329 sites cited across 960 answers from ChatGPT, Perplexity, Gemini and Claude, grouped by how many engines cited each one: 966 by one engine only (72.7%), 205 by two engines (15.4%), 129 by three engines (9.7%), and 29 by all four engines (2.2%). On 69 of the 80 questions, no site was cited by all four engines. In the July 2026 study, 5 of 713 domains were cited by all four engines across 47 buyer questions." width="780" height="451" style="max-width:100%;height:auto" />
  <figcaption>AI engine source overlap chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

Our July 2026 study of 47 buyer questions showed the same count: 5 of 713 domains were cited by all four engines. The practical reading: a site cited widely tends to be cited by several engines, but a site that turns up in a handful of answers usually turns up on one engine only.

Does the same question get the same sources twice?

Usually not, except on Perplexity. On ChatGPT, two runs of the same question shared, on average, 37% of the sites cited across both. On Gemini the figure was 35%. On Claude it was 68% and on Perplexity 96%. Perplexity cited exactly the same sites in 199 of 240 pairs of runs, Claude in 55 of 179, ChatGPT in 27 of 206 and Gemini in 2 of 200.

How we measured it: for each pair of runs of the same question on the same engine, we divided the sites both runs cited by all the sites cited across both runs, then averaged across pairs. It is not the share of one run's sites that turn up again in the next run. For example, ChatGPT's answers to "Do AI Overviews reduce clicks to websites?" in runs 1 and 2 cited 5 different sites between them and shared 2 (ahrefs.com and pewresearch.org), an overlap of 40%. Counting by company would not change this pair. Pairs where one run cited nothing are left out. If they count as sharing no sites, the averages fall to 34% for ChatGPT, 31% for Gemini and 59% for Claude, with Perplexity unchanged at 96%.

Bar chart of how many of the sites cited across two runs of the same question both runs cited, averaged over 80 questions with 3 runs each, September 2026: Perplexity 96% (199 of 240 run pairs cited exactly the same sites), Claude 68% (55 of 179 identical), ChatGPT 37% (27 of 206 identical), Gemini 35% (2 of 200 identical). Worked example: ChatGPT, "Do AI Overviews reduce clicks to websites?", runs 1 and 2: shared ahrefs.com, pewresearch.org; only run 1 blog.google, papers.ssrn.com; only run 2 arxiv.org; 2 of 5 sites, 40%.
Prefer AI answer study, September 2026: sites both runs cited divided by all sites cited across both runs, averaged over pairs of runs that both cited at least one site.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/study-repeat-overlap.png" alt="Bar chart of how many of the sites cited across two runs of the same question both runs cited, averaged over 80 questions with 3 runs each, September 2026: Perplexity 96% (199 of 240 run pairs cited exactly the same sites), Claude 68% (55 of 179 identical), ChatGPT 37% (27 of 206 identical), Gemini 35% (2 of 200 identical). Worked example: ChatGPT, &quot;Do AI Overviews reduce clicks to websites?&quot;, runs 1 and 2: shared ahrefs.com, pewresearch.org; only run 1 blog.google, papers.ssrn.com; only run 2 arxiv.org; 2 of 5 sites, 40%." width="780" height="751" style="max-width:100%;height:auto" />
  <figcaption>AI answer source overlap between repeat runs chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

Each run was a fresh API call with no chat history, and the three answers to each question came back within 23 seconds of each other. So these numbers show how much answers vary from one moment to the next, not how they change over weeks.

Which sites were cited most?

Reddit was cited in more answers than any other site: 207 of the 960 (21.6%). The table counts answers that cited a site at least once, split by engine.

Site Answers citing it (of 960) ChatGPT Perplexity Gemini Claude
Reddit reddit.com 207 16 98 93 0
Semrush semrush.com 130 10 66 39 15
Search Engine Land searchengineland.com 125 2 87 24 12
YouTube youtube.com 113 0 0 113 0
Ahrefs ahrefs.com 103 14 52 21 16
Profound tryprofound.com 103 8 57 12 26
Google Search Central developers.google.com 96 58 15 23 0
LinkedIn linkedin.com 75 0 73 0 2
Geoptie geoptie.com 73 0 33 28 12
arXiv arxiv.org 66 44 6 7 9

Each engine's most-cited site was different. ChatGPT cited Google Search Central docs most (58 answers), Perplexity cited Reddit (98), Gemini YouTube (113) and Claude Profound (26). ChatGPT's next three were arXiv (44), TechRadar (27) and OpenAI's help center (23). Claude's next two were SE Ranking (17) and Ahrefs (16).

Profound, Geoptie, Semrush, Ahrefs and SE Ranking sell SEO or AI visibility software, as Prefer does. We report their counts as measured.

Bar chart of the 10 sites cited in the most answers, of 960, with the engines that cited each one: reddit.com 207 (ChatGPT, Perplexity, Gemini; not Claude); semrush.com 130 (all four engines); searchengineland.com 125 (all four engines); youtube.com 113 (Gemini; not ChatGPT, Perplexity or Claude); ahrefs.com 103 (all four engines); tryprofound.com 103 (all four engines); developers.google.com 96 (ChatGPT, Perplexity, Gemini; not Claude); linkedin.com 75 (Perplexity, Claude; not ChatGPT or Gemini); geoptie.com 73 (Perplexity, Gemini, Claude; not ChatGPT); arxiv.org 66 (all four engines). Neither Reddit nor YouTube was cited by all four engines.
Prefer AI answer study, September 2026: answers citing each site at least once, of 960. Subdomains of reddit.com, youtube.com, linkedin.com and wikipedia.org are merged. developers.google.com is Google Search Central.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/study-top-sites.png" alt="Bar chart of the 10 sites cited in the most answers, of 960, with the engines that cited each one: reddit.com 207 (ChatGPT, Perplexity, Gemini; not Claude); semrush.com 130 (all four engines); searchengineland.com 125 (all four engines); youtube.com 113 (Gemini; not ChatGPT, Perplexity or Claude); ahrefs.com 103 (all four engines); tryprofound.com 103 (all four engines); developers.google.com 96 (ChatGPT, Perplexity, Gemini; not Claude); linkedin.com 75 (Perplexity, Claude; not ChatGPT or Gemini); geoptie.com 73 (Perplexity, Gemini, Claude; not ChatGPT); arxiv.org 66 (all four engines). Neither Reddit nor YouTube was cited by all four engines." width="780" height="731" style="max-width:100%;height:auto" />
  <figcaption>Most-cited sites in AI answers chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

Two choices move this ranking. The first is how sites are counted. We merge subdomains only for Reddit, YouTube, LinkedIn and Wikipedia, whose subdomains are language, mobile or regional versions of one platform. Other subdomains stay separate because they are often different properties: developers.google.com is Google's developer documentation and support.google.com its help center. Counted by company instead, SE Ranking (seranking.com plus visible.seranking.com) would have 110 answers and rank sixth, and it would pass Profound on Claude (29 answers against 26). google.com would have 112 and HubSpot 79.

The second is the question set. 8 of the 80 questions name Profound or Peec AI, and 29 of Profound's 103 answers came from those 8. In our August 2026 study of 25 B2B software buyer queries, measured with a different method, Reddit was ChatGPT's top source on every query. Here, on questions about AI search, ChatGPT cited Reddit in 16 of 240 answers.

Each engine leans on a different platform

Gemini cited YouTube in 113 of its 240 answers (47.1%). ChatGPT, Perplexity and Claude cited YouTube in none. Perplexity cited LinkedIn in 73 answers (30.4%), 55 of them LinkedIn Pulse articles, while ChatGPT and Gemini cited LinkedIn in none and Claude in 2. Reddit appeared in 98 Perplexity answers and 93 Gemini answers, against 16 for ChatGPT and 0 for Claude.

Platform ChatGPT Perplexity Gemini Claude
Reddit 16 (6.7%)98 (40.8%)93 (38.8%)0 (0%)
YouTube 0 (0%)0 (0%)113 (47.1%)0 (0%)
LinkedIn 0 (0%)73 (30.4%)0 (0%)2 (0.8%)
Wikipedia 6 (2.5%)17 (7.1%)22 (9.2%)8 (3.3%)
Grid of answers citing each platform, of 240 per engine, September 2026. Reddit: ChatGPT 16, Perplexity 98, Gemini 93, Claude 0. YouTube: ChatGPT 0, Perplexity 0, Gemini 113, Claude 0. LinkedIn: ChatGPT 0, Perplexity 73, Gemini 0, Claude 2. Wikipedia: ChatGPT 6, Perplexity 17, Gemini 22, Claude 8. Only Gemini cited YouTube, and almost only Perplexity cited LinkedIn.
Prefer AI answer study, September 2026: 240 answers per engine. Counts are answers citing the platform at least once.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/study-platform-by-engine.png" alt="Grid of answers citing each platform, of 240 per engine, September 2026. Reddit: ChatGPT 16, Perplexity 98, Gemini 93, Claude 0. YouTube: ChatGPT 0, Perplexity 0, Gemini 113, Claude 0. LinkedIn: ChatGPT 0, Perplexity 73, Gemini 0, Claude 2. Wikipedia: ChatGPT 6, Perplexity 17, Gemini 22, Claude 8. Only Gemini cited YouTube, and almost only Perplexity cited LinkedIn." width="780" height="779" style="max-width:100%;height:auto" />
  <figcaption>Platforms cited by each AI engine chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

These splits are not fixed. In our July 2026 study of 47 buyer questions, with the same Perplexity model (sonar-pro), Perplexity cited YouTube 28 times; here it cited YouTube in none of its 240 answers.

What this means for your measurement

If you track your own brand in AI answers, three things follow from these numbers. Prefer tracks your buyer prompts on ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode and reports each engine separately, with the sources behind each answer.

  1. Run each question at least three times on each engine. Two ChatGPT runs of the same question shared, on average, 37% of the sites cited across both, and two Gemini runs 35%, so one run shows only part of the sources an engine uses for that question. Done when every tracked question has three runs per engine on record.
  2. Record whether each run searched the web. All 99 answers that skipped search cited nothing, so a run with no citations can mean the engine answered from memory, not that you lost a citation. The three runs disagreed about searching on 4 questions for ChatGPT , and on 13 each for Gemini and Claude . Done when each run's record has a searched yes or no next to its sources.
  3. Report each engine separately, with the platform it leans on. Gemini cited YouTube in 113 of 240 answers. Perplexity cited Reddit in 98 and LinkedIn in 73. ChatGPT cited Google Search Central most (58 answers), and Claude never cited Reddit. Done when your report shows one line per engine instead of one blended number.

Our guide to measuring AI visibility sets out the full method.

What do Google AI Overviews and AI Mode cite?

Google showed an AI Overview for 75 of the 80 questions (93.8%). Each Overview cited 8.64 sources on average. AI Mode returned an answer for 76 of the 80 questions; the other 4 captures came back empty. Those 76 answers cited 5.80 sources on average, and 2 of them cited none.

YouTube was the site Google's AI answers cited most: in 38 of 75 AI Overviews and 51 of 76 AI Mode answers. The organic results for the same questions ranked differently. Reddit appeared on the first page for 61 of 80 questions and YouTube for 14. LinkedIn was third on both AI surfaces and second in the organic results.

Site AI Overviews (of 75) AI Mode (of 76) Organic first page (of 80)
YouTube 38 51 14
Reddit 29 26 61
LinkedIn 18 14 27
Google AI surfaces for 80 questions, 13 September 2026. An AI Overview was shown for 75 of 80 questions (93.8%), with 8.64 sources per Overview. AI Mode returned an answer for 76 of 80 questions (4 captures came back empty), with 5.80 sources per answer. The organic first page had 6 to 10 results per question. Questions where each site appeared, in AI Overviews (of 75), AI Mode answers (of 76) and the organic first page (of 80): YouTube 38, 51 and 14; Reddit 29, 26 and 61; LinkedIn 18, 14 and 27.
Prefer AI answer study, September 2026: one Google capture per question for the AI Overview with organic results and one for AI Mode, US English desktop.
Reuse this chart

Free to reuse with attribution. Copy the embed code, or save the image.

Download PNG 
<figure>
  <img src="https://tryprefer.com/figures/study-google-surfaces.png" alt="Google AI surfaces for 80 questions, 13 September 2026. An AI Overview was shown for 75 of 80 questions (93.8%), with 8.64 sources per Overview. AI Mode returned an answer for 76 of 80 questions (4 captures came back empty), with 5.80 sources per answer. The organic first page had 6 to 10 results per question. Questions where each site appeared, in AI Overviews (of 75), AI Mode answers (of 76) and the organic first page (of 80): YouTube 38, 51 and 14; Reddit 29, 26 and 61; LinkedIn 18, 14 and 27." width="780" height="720" style="max-width:100%;height:auto" />
  <figcaption>Google AI Overviews and AI Mode sources chart by <a href="https://tryprefer.com/data/ai-answer-study/">Prefer</a></figcaption>
</figure>

Google's AI Overviews cited LinkedIn for 18 questions and AI Mode for 14. Among the four chat engines, Perplexity cited LinkedIn in 73 answers and the other three in 2 between them. If an AI Overview does not appear for your own searches, our answer on why an AI Overview is not showing lists the usual reasons.

How we measured this

Questions. 80 questions about AI search, GEO and AEO, in 10 topics of 8. We chose the topics to feed a series of LinkedIn articles and a review of which topics our own site covers. 8 questions name Profound or Peec AI, and 29 of the 103 answers citing tryprofound.com came from them. No question names Prefer. The full list is below.

Engines and runs. Each question went to ChatGPT (gpt-5.6-luna), Perplexity (sonar-pro), Gemini (gemini-3.8-flash) and Claude (claude-sonnet-5) through the DataForSEO LLM Responses API, with web search requested on every call. Each question ran 3 times per engine, giving 960 answers. These are API answers. What a person sees in the consumer apps can differ.

Timing. All 960 answers were collected on 13 September 2026 between 17:56 and 18:12 UTC. The three answers to each question on each engine came back within 23 seconds of each other. Each run was a separate API call with no chat history.

Google. For each question, one capture of Google's results page with the AI Overview and organic results, and one capture of AI Mode, both US English desktop through DataForSEO, the same day. 4 of the AI Mode captures came back with no answer.

Definitions. An answer "searched the web" when the API reported that it used web search. A source is a unique URL cited in the answer. A site is the host name without "www". Subdomains of reddit.com, youtube.com, linkedin.com and wikipedia.org are merged, because they are language, mobile or regional versions of one platform. Other subdomains count as separate sites (blog.hubspot.com and hubspot.com are two), because they are often different properties. Merging every subdomain gives 1,294 sites instead of 1,329, still 29 cited by all four, and moves the ranking as described above. "Answers citing a site" counts an answer once, however many of that site's pages it cited.

Limits. This is one run date. The questions are about AI search itself, so rates for other topics can differ. We counted what the engines cited; we did not grade whether the answers were right. We also left out a breakdown by source type, because most sources could not be classified automatically.

Our stake. Prefer sells software that tracks AI citations, so we have a stake in how these engines behave. This page leaves out every result about Prefer's own site.

Download the data

The dataset is free to reuse under CC BY 4.0. Cite it as Prefer AI answer study, September 2026, with a link to this page.

  • Full dataset (JSON): the method, all 80 questions, per-engine results, the top sites, cross-engine and run-to-run overlap, Google surfaces and per-topic search rates.
  • Per-engine results (CSV): one row per engine with every number in the first table.

The 80 questions

What does GEO mean? (8 questions)
  1. What does GEO mean?
  2. What is generative engine optimization (GEO)?
  3. What does GEO mean in marketing, and is it different from SEO?
  4. What does GEO stand for in SEO?
  5. GEO meaning in business
  6. What is GEO in digital marketing?
  7. Who came up with the term generative engine optimization?
  8. What is a GEO strategy?
SEO vs AEO vs GEO vs AIO (8 questions)
  1. What is the difference between GEO and SEO?
  2. What is the difference between AEO and SEO?
  3. Are AEO and GEO the same thing?
  4. SEO vs AEO vs GEO vs AIO: what's the difference?
  5. Is GEO replacing SEO?
  6. What does AIO mean in SEO?
  7. What is answer engine optimization (AEO)?
  8. Do I still need SEO if I'm doing GEO?
Best AI visibility tools (8 questions)
  1. What are the best AI search visibility tools in 2026?
  2. What is the best AI visibility tool?
  3. Which AI visibility tools track ChatGPT, Perplexity, Gemini, and Claude?
  4. What is the best AI visibility tool for a small marketing team on a budget?
  5. Is there a free AI visibility checker?
  6. How much do AI search visibility tools cost?
  7. Which AI search optimization tool is easiest to use?
  8. Which AI visibility platforms also include SEO features?
Ranking in Google AI Overviews and AI Mode (8 questions)
  1. How do I rank in Google AI Overviews?
  2. How does Google decide which sites to cite in AI Overviews?
  3. What is Google AI Overview?
  4. Why isn't an AI Overview showing up for my search?
  5. How do I get my website cited in Google AI Mode?
  6. What is the difference between Google AI Overviews and AI Mode?
  7. Do I need to rank on page one to appear in AI Overviews?
  8. Do AI Overviews reduce clicks to websites?
LLM SEO (8 questions)
  1. What is LLM SEO?
  2. What is LLM optimization (LLMO)?
  3. How do I optimize my content for LLMs?
  4. How do I rank in ChatGPT search?
  5. How do I get my website cited by ChatGPT?
  6. How do I get my brand mentioned by Claude and Gemini?
  7. Why isn't my brand showing up in ChatGPT answers?
  8. Which sources does ChatGPT cite most often?
AI SEO and GEO agencies (8 questions)
  1. What are the best AI SEO agencies?
  2. What is the best generative engine optimization agency?
  3. Which agencies offer GEO or AEO services for B2B companies?
  4. How much does an AI SEO agency cost per month?
  5. What does a GEO agency actually do?
  6. How do I choose an AEO agency?
  7. Should I hire an agency or use a tool for AI search visibility?
  8. What are the best LLM SEO companies?
How to measure AI visibility (8 questions)
  1. How do you measure AI visibility for a brand?
  2. How do I check my brand's visibility in ChatGPT and other AI engines?
  3. What is AI share of voice and how do you calculate it?
  4. How do I track citations of my website in AI search?
  5. What are the best LLM tracking tools?
  6. How many prompts should I track for AI visibility?
  7. How can I improve my brand's visibility in AI search?
  8. What should be on an AEO checklist?
Profound vs Peec AI (8 questions)
  1. How much does Profound cost?
  2. What are the best Profound alternatives?
  3. Is Profound worth it for tracking AI search visibility?
  4. How much does Peec AI cost?
  5. What are the best Peec AI alternatives?
  6. Is Peec AI any good?
  7. Peec AI vs Profound: which AI visibility tool is better?
  8. What are the downsides of Profound?
Best AEO and GEO tools (8 questions)
  1. What are the best AEO tools?
  2. Best AEO (answer engine optimization) platforms
  3. What are the best generative engine optimization (GEO) tools for brands in 2026?
  4. What are the best LLM SEO tools?
  5. What tools should an agency use to offer AEO or GEO services to clients?
  6. Which GEO tools help create content, not just track visibility?
  7. What are the best practices for answer engine optimization?
  8. Best tool to track AI citations / share of voice in AI search
AI search statistics (8 questions)
  1. What percentage of Google searches are zero-click?
  2. What are the latest AI search statistics?
  3. How many people use ChatGPT for search?
  4. How much traffic do AI chatbots send to websites?
  5. Is AI search replacing Google?
  6. Which AI engine sends the most traffic to websites?
  7. What share of Google searches show an AI Overview?
  8. How many sources does ChatGPT cite per answer?

For what these numbers mean in practice, read our guide to generative engine optimization, which is built on this study, or see all of our research and data. To see which sites AI engines cite for your own buyers' questions, run a free Prefer AI visibility audit.

People also ask

  • How do ChatGPT, Perplexity, Gemini and Claude search and cite?
  • How many sources does ChatGPT cite per answer?
  • Do AI engines cite the same sources?
  • Do AI answers change when you ask the same question twice?

Frequently asked questions.

Updated 19 September 2026

How often does ChatGPT search the web before it answers?

In Prefer's September 2026 study, ChatGPT searched the web for 216 of 240 answers (90%), with web search requested on every call through the API. Its lowest rate was in the "What does GEO mean?" topic: 11 of 24 answers. None of the 24 ChatGPT answers that skipped search cited a source.

Which AI engine cites the most sources?

Perplexity. In Prefer's September 2026 study of 960 AI answers, it cited 19.48 sources per answer and 748 different sites across 240 answers, and it was the only engine that cited at least one source in every answer. Gemini cited 8.64 per answer, Claude 4.58 and ChatGPT 3.05.

Do ChatGPT, Perplexity, Gemini and Claude cite the same sites?

By count, rarely. In Prefer's September 2026 study, of 1,329 sites cited across 960 answers, 29 were cited by all four engines, and on 69 of 80 questions no site was. By prominence, more often: 23 of the 25 most-cited sites were cited by three or four engines. Reddit, the most-cited site, was never cited by Claude.

Do AI answers cite the same sources if you ask the same question twice?

Usually not, except on Perplexity. In Prefer's September 2026 study, two runs of the same question shared, on average, 96% of the sites cited across both on Perplexity, 68% on Claude, 37% on ChatGPT and 35% on Gemini. Identical sets of sites came back in 199 of 240 Perplexity run pairs, 55 of 179 on Claude, 27 of 206 on ChatGPT and 2 of 200 on Gemini. The three answers came back within 23 seconds of each other, so this is variation from one moment to the next.

How was this study measured?

On 13 September 2026 Prefer sent 80 questions about AI search, GEO and AEO to four engines through the DataForSEO LLM Responses API, with web search requested on every call: ChatGPT (gpt-5.6-luna), Perplexity (sonar-pro), Gemini (gemini-3.8-flash) and Claude (claude-sonnet-5). Each question ran 3 times per engine, giving 960 answers. For each question we also captured Google's results page with any AI Overview, and Google AI Mode, which returned an answer for 76 of the 80. These are API answers, which can differ from the consumer apps.

How do I cite this data?

Cite it as "Prefer AI answer study, September 2026" and link to this page. The data, the JSON and CSV downloads and the charts are free to reuse under CC BY 4.0.

Keep reading

All research
Guide30 min read

What is GEO? The complete guide

 Data6 min read

Reddit is the source AI cites most

 Data7 min read

AI search citation statistics

Get your free AI visibility report
in about 10 minutes.

See how answer engines describe your brand today, and where the openings are to outpace the competition.

Free AI visibility audit

Takes about 5 minutes. No signup required.

Free. No card required.