ChatGPT vs Perplexity vs Gemini: How Each Cites Sources
All three engines search the web, then attach links to the answer, but they pick different sources. Perplexity shows numbered citations on every answer 3. ChatGPT and Gemini show links only when their search tools run. In one study of 15,000 prompts, 28.6% of Perplexity's cited links ranked in Google's top 10, against about 8% for ChatGPT 1.
- Perplexity shows numbered citations on every answer. ChatGPT and Gemini cite when their search tools fire, so the same question can return different sources.
- Overlap with Google's top 10 was 28.6% for Perplexity and about 8% for ChatGPT in one 15,000-prompt study.
- Engines disagree: in a 2026 local-search study, all four AI engines named the same top business in only 4% of cases.
- No single study covers all three engines on one prompt set. Treat every share as a snapshot, then test your own prompts.
How do ChatGPT, Perplexity and Gemini differ in citing sources?
They differ in when they cite, which sources they favour and how closely they follow Google. The common pattern is the same: the model runs web searches, reads the results, writes an answer, and links the pages it used.
Perplexity is built around citations. Its help centre says each answer includes numbered citations that link to the original sources 3. ChatGPT and Gemini are general assistants, so they cite when a web search tool is used. OpenAI's developer guide says the tool adds inline citations for URLs found in the results 2, and Google's Gemini API documents the same idea for grounding with Google Search 4.
This page is a review of third-party studies. Chalk Labs has not run its own citation tests for it. The studies in this page show a second difference. Each engine leans on a different mix of Reddit, Wikipedia, LinkedIn, directories and brand websites. So the question is not which engine cites best. It is which engines your buyers use, and which sources each one trusts for your topic.
How does each engine find and show sources?
Each engine runs searches first and cites second. The details below come from the vendors' own documentation, so they describe design, not ranking rules. None of the three publishes how it ranks sources.
| Question | ChatGPT | Perplexity | Gemini |
|---|---|---|---|
| When does it cite? | When the web search tool runs. Inline citations are on by default for URLs found in results 2. | Every answer includes numbered citations 3. | When grounding with Google Search is used. The model decides if a search can improve the answer 4. |
| How many searches? | Depends on mode: a simple lookup, agentic search that manages itself, or deep research across hundreds of sources 2. | Normal answers search in real time. Deep Research runs dozens of searches and reads hundreds of sources 3. | The model may generate one or several queries and run them 4. |
| Are sources and citations the same? | No. The sources list is often longer than the citations list 2. | The help page does not say. | Citations link answer spans to URLs. The developer page lists the search queries run 4. |
| Which crawler lets a site appear? | OAI-SearchBot. Blocking it keeps a site out of ChatGPT search answers, though navigational links can still appear 5. | Perplexity publishes its own search crawler. Check its bot documentation for the current name and rules. | Grounding uses Google Search, so a page that Google can index is the starting point (our inference) 4. |
| What can you not see? | Why one page was chosen over another. | How it ranks sources. The help page does not explain it. | How it ranks sources beyond using Google Search. |
Do the engines just copy Google's top results?
No, and Perplexity comes closest. Ahrefs ran 15,000 long-tail prompts through Google and Bing, then through AI assistants. About 28.6% of Perplexity's cited links appeared in Google's top 10, against 8.0% for ChatGPT's in-text citations and 6.1% for its bottom-of-response references 1.
Ahrefs also reported an average of about 12% across assistants. That figure is an unweighted average of platform values, and the article does not label Gemini's value separately 1. So do not quote Gemini's overlap from this study.
The study has limits. It used long-tail prompts only, ChatGPT data was collected in early July 2025, and assistants may search variations of your prompt instead of the original wording 1. The practical reading is that ranking in Google helps you most on Perplexity, and helps much less as a predictor on ChatGPT.
Which kinds of sites does each engine cite most?
It depends on the study, and the studies disagree. Semrush tracked about 230,000 prompts for 13 weeks in 2025. Reddit and Wikipedia were the two most-cited domains in ChatGPT search, Google AI Mode favoured LinkedIn, YouTube and Reddit, and Perplexity favoured Reddit, LinkedIn, NIH, Microsoft and Google 6.
Those shares moved fast. In the same Semrush data, Reddit appeared in close to 60% of ChatGPT responses in early August, then fell to about 10% by mid-September 6. A ranking of cited sites is a weather report, not a law.
Profound analysed about 11.84 billion citations across eight models from April to July 2026. Brand-owned sites made up 47% of ChatGPT's citations and 69% of Gemini's 7. Profound is a vendor, and its figures describe its tracked categories, not the whole web.
Local questions look different again. Yext analysed answers from four AI engines for local queries and found directories supplied 40.9% of OpenAI's cited sources, against 28.8% for Gemini and 32.5% for Perplexity 8. Yext sells listings software, so read its framing with that in mind.
Will the three engines even recommend the same business?
Usually not. Yext compared 491,972 local query cases across Gemini, OpenAI, Perplexity and Anthropic between February and August 2026. All four named the same top business in 4% of cases. All four named different top businesses in 48% 8.
At least three engines agreed in 18.3% of cases 8. These were unbranded local questions, such as the top 10 businesses of a type near a neighbourhood, so B2B results may differ.
The takeaway is plain. Winning one engine does not mean winning the others. If your buyers use two engines, test both.
What should you do differently for each engine?
Start with access. Allow the search crawlers of each engine, and keep your answers in plain, server-rendered text. For OpenAI, the search crawler is separate from the training crawler, so you can allow one and block the other 5. Our guide to GPTBot, OAI-SearchBot and ChatGPT-User explains the choice, and the AI crawler checker tests your site.
Then tune by engine. For Perplexity, ordinary search strength carries further, so see how to rank in Perplexity. For ChatGPT, the evidence points to a wider spread of sources, including third-party mentions, so see how to get recommended by ChatGPT. For Gemini and Google's own answers, compare AI Overviews and AI Mode.
Last, remember the searches behind each answer. Every engine here can run several queries per prompt 24, which is why a single page rarely wins alone. Read what query fan-out is to plan pages around the sub-questions.
- Allow OAI-SearchBot and Perplexity's published search crawler (see its bot documentation) unless you have a business reason not to.
- Publish one clear answer per question, in the first two sentences of the page.
- Earn mentions on sites each engine already cites in your field, such as directories, forums or industry bodies.
- Re-test every month, because source mixes shift within weeks.
How do you check which engine cites you?
Run the same 10 buyer questions in each engine and record the cited domains. Do it in a fresh session, because personalisation can change citations for the same prompt 1. Repeat it three times, since answers vary between runs.
A small log is enough to start. The example below is a layout, not data.
- Prompt: best [your category] for [your buyer]
- Date | Engine | Run | Cited domains | We appear? | Competitor named
- 2026-11-01 | ChatGPT | 1 | domain-a.com, domain-b.com | no | Competitor X
- 2026-11-01 | Perplexity | 1 | domain-a.com, domain-c.com | yes (cited) | Competitor X
- 2026-11-01 | Gemini | 1 | domain-d.com, domain-b.com | no | none
Which engine should you optimise for first?
Pick the one your buyers already use, then confirm it in your own referral data. No public study gives a safe universal answer, and traffic shares vary by source and month.
If you cannot tell, build for the shared basics: crawlable pages, direct answers and credible third-party mentions. The quiz below turns this into a short decision.
Frequently asked questions
Which AI engine shows the most citations?
Perplexity shows numbered citations on every answer, according to its help centre 3. ChatGPT and Gemini show them when their web search tools run, so a simple factual question may come back with no links at all.
Does ranking in Google mean ChatGPT will cite me?
Not reliably. In Ahrefs' 15,000-prompt study, only about 8% of ChatGPT's in-text citations were in Google's top 10 for the same prompt 1. Ranking helps, but ChatGPT draws on a wider spread of pages.
Do ChatGPT, Perplexity and Gemini cite the same websites?
Often not. In a 2026 Yext study of local queries, all four AI engines named the same top business in only 4% of cases, and four different businesses in 48% 8. Test each engine you care about.
Should I block AI crawlers to protect my content?
Blocking search crawlers removes you from that engine's answers. OpenAI says a site blocking OAI-SearchBot will not be shown in ChatGPT search answers 5. Training and search crawlers are separate, so you can choose each one.
How often do citation patterns change?
Quickly. Semrush saw Reddit's presence in ChatGPT responses drop from close to 60% to about 10% within weeks in 2025 6. Check your own prompts monthly instead of relying on a published ranking.
Sources
- Ahrefs AI search overlap study (ChatGPT, Gemini, Copilot, Perplexity vs Google and Bing) · AhrefsOpened October 20, 2026. Supports: 15,000 long-tail prompts; Perplexity 28.6% and ChatGPT 8.0% in-text, 6.1% references overlap with Google top 10; 12% average caveat; personalisation and query variations.
- Web search · OpenAI (API docs)Opened October 20, 2026. Supports: Inline url_citation annotations; sources list longer than citations; three search modes including deep research.
- How does Perplexity work? · Perplexity Help CenterOpened October 20, 2026. Supports: Real-time search, numbered citations on each answer, Deep Research runs dozens of searches and reads hundreds of sources.
- Grounding with Google Search · Google AI for DevelopersOpened October 20, 2026. Supports: Model decides whether to search, can run one or more queries, returns url_citation annotations linking answer spans to sources (page updated 2026-10-13).
- Overview of OpenAI Crawlers (Bots) · OpenAIOpened October 20, 2026. Supports: OAI-SearchBot surfaces sites in ChatGPT search; blocking excludes a site from answers; about 24 hours to take effect; separate from GPTBot.
- The Most-Cited Domains in AI: A 3-Month Study · SemrushOpened October 20, 2026. Supports: About 230,000 prompts, 14 July to 12 October 2025; top cited domains per engine; Reddit share in ChatGPT fell from close to 60% to about 10%.
- Where do AI citations come from? · ProfoundOpened October 20, 2026. Supports: 11.84 billion citations, April to July 2026; Eight models covered; brand-owned share of citations 47% for ChatGPT and 69% for Gemini.
- Four AIs, one question, four different answers · Yext ResearchOpened October 20, 2026. Supports: 491,972 local query cases, February to August 2026; 4% same top business, 48% four different, 18.3% three or more agree; directory shares per engine.