VantageGet started

Vantage research · September 2026

What AI answer engines cite

We ask ChatGPT, Gemini and Perplexity the same 200 everyday questions every month and record which websites their answers cite. Here is what September 2026 looked like.

Updated monthlyFirst editionNext edition 1 October 2026 Download the data (JSON)

598
AI answers measured
200 questions, 3 engines
4.4%
of cited sites are community sites
Reddit, YouTube, Quora and similar
18.7%
of questions where all three agree
on at least one source
reddit.com
cited most across all engines
in 94 answers

What stood out

  • Perplexity cites the most sites: 9.9 per answer, against 2.9 on ChatGPT.
  • reddit.com is Perplexity's favourite source, cited in 87 of its 199 answers. ChatGPT cites it in 0 answers, Gemini in 7.
  • Gemini answers 46% of questions without citing a single site, from what the model already knows.
  • The engines rarely agree: all three cite a common site for only 18.7% of the questions they all sourced.

How the engines compare

ChatGPTGeminiPerplexity

Sites cited per answer

when any are cited

ChatGPT2.9
Gemini3.4
Perplexity9.9

Community share

of the sites each engine cites

ChatGPT1.2%
Gemini4.1%
Perplexity5.3%

Answers with no sources

answered from memory

ChatGPT1%
Gemini46%
Perplexity0%

Answers with a table

of all answers

ChatGPT14.6%
Gemini14.5%
Perplexity6%

Each bar is scaled to the highest value in its own panel; the number beside it is the actual value.

Where each engine gets its answers

The sites each engine cites most, by the number of answers citing them. All three panels share one scale.

ChatGPT

199 answers

  1. wired.com17
  2. tomsguide.com13
  3. rtings.com12
  4. cdc.gov11
  5. techradar.com10
  6. nhs.uk9
  7. consumerfinance.gov8
  8. healthychildren.org8

Gemini

200 answers

  1. google.com12
  2. pcmag.com12
  3. play.google.com8
  4. reddit.comcommunity7
  5. cnet.com5
  6. coursera.org5
  7. zapier.com5
  8. forbes.com4

Perplexity

199 answers

  1. reddit.comcommunity87
  2. forbes.com55
  3. nytimes.com44
  4. pcmag.com36
  5. healthline.com30
  6. goodhousekeeping.com25
  7. tomsguide.com25
  8. cnbc.com23

Do the engines agree?

Each dot is one question that all three engines answered with sources. A filled dot means all three cited at least one of the same sites: 20 of 107. Hover a dot to see the question.

all three share a source no site in common

Community share by topic

Share of each engine's cited sites that are community sites, per topic. Darker is higher.

Community share of cited sites, by topic and engine
TopicChatGPTGeminiPerplexityMost cited site
careers6.8%6.7%8.5%coursera.org
software0%3.4%7.5%pcmag.com
tech0%5%6.1%pcmag.com
travel3.3%1.7%5.8%reddit.com
food0%4.2%5.5%foodnetwork.com
family0%6.9%4.5%healthychildren.org
home0%3.1%5%consumerreports.org
wellbeing0%4%4.5%healthline.com
finance0%2.9%3%cnbc.com
fitness0%5%3%healthline.com

All 200 questions

Each question has its own page: which sites every engine cited, where they overlap, and what changed since last month.

Careers 20
Family 20
Finance 20
Fitness 20
Food 20
Home 20
Software 20
Tech 20
Travel 20
Wellbeing 20

How does your site do?

Run the same checks on your own domain from your AI agent: which questions cite you, in ChatGPT, Gemini and Perplexity, and what changed since last time.

See the tools Try one keyword free

How this is measured

Each question is asked once per engine, in English, as from the United States. ChatGPT and Gemini are read the way a person sees them in those apps; Perplexity through its own sonar API with web search on. For every answer we keep only measurements: which sites it cites, whether it opens with a list, whether it has a table, and how long its opening is. Never the answer text. "Community sites" are sites whose content is posted by their users (Reddit, YouTube, X, Quora, LinkedIn, Medium, review sites and similar). A site counts once per answer however many of its pages are cited. Answers change from run to run, so treat single rows as a snapshot and the totals as the finding. 2 of 600 requests failed and are left out. Measured on 2026-09-23; the same questions run again on 1 October 2026.