ChatGPT, Perplexity, Copilot and Gemini Cite Different Sources: A 120-Answer Study

We asked ChatGPT, Perplexity, Microsoft Copilot and Google Gemini the same 30 questions on the same day and recorded every source each answer cited. The four engines cited 458 distinct domains between them. 349 of those (76%) were cited by one engine only, and for 29 of the 30 questions no single domain was cited by all four. If you measure your brand's visibility in one AI engine, you have learned very little about the other three.

The short version

• The engines barely overlap. Across 120 answers the four engines cited 458 distinct domains: 349 by one engine only, 85 by two, 23 by three, and 1 (aws.amazon.com) by all four.

• Agreement is the exception. For 29 of the 30 questions, no domain was cited by every engine.

• Each engine has its own habits. Perplexity cited Reddit in 10 of its 30 answers and Gemini in 4; ChatGPT and Copilot never did. Copilot cited YouTube in 7 answers; no other engine cited it once. None of the 120 answers cited Wikipedia.

• Citation counts differ by two to one. Perplexity and Copilot cited a median of 10 sources per answer, Gemini 7 and ChatGPT 5. Gemini gave 8 answers with no sources at all, ChatGPT 6.

How we ran it

• 30 questions a person might ask an assistant: 22 "what is the best…" or "which…" questions about products and software, and 8 "how do I…" questions (lowering blood pressure, starting a podcast, negotiating a salary, and so on).

• Each question asked once per engine, in US English (gl=us, hl=en), in each engine's own consumer web interface in a signed-out session. We collected the answers through Searlo's AI answer API, which is our product: read the results with that in mind.

• All on 29 September 2026, between 10:33 and 12:45 UTC. Copilot needed a retry on 9 questions before it answered; in the end every engine answered all 30.

• ChatGPT's answers were collected twice. In the first pass our ChatGPT collector had a fault that could carry one question's prompt into the next. We fixed it, discarded that set and asked ChatGPT all 30 questions again; the numbers below use the second set only.

• A citation is a source the engine linked in its answer. A domain is its hostname without "www.", so uk.pcmag.com and pcmag.com count separately.

• The data: ai-citation-study-2026-09.csv has the engine, question, time, number of citations and cited domains for all 120 answers. It does not include the answer text.

Sources per answer

ChatGPT and Gemini show sources only when they search the web, so an answer written from the model's own knowledge has none. Perplexity and Copilot cited something every time. Time per call is the wall-clock time of each API call, page load included.

The engines barely agree

For each question we compared the set of domains each pair of engines cited. Overlap here is the share of all the domains the two engines cited for a question that both of them cited (Jaccard similarity), averaged over the questions.

The closest pair, Perplexity and Copilot, shared about one domain in five on a typical question. ChatGPT and Gemini shared almost nothing: on the 28 questions where at least one of them cited a source, their overlap averaged 0.01. Put the four answers to one question side by side and they cite a median of 19 different domains (13 at the fewest, 25 at the most).

Each engine has its favourite sources

• Perplexity cited Reddit in 10 of 30 answers. Most cited: techradar.com (in 11 answers), reddit.com (10), nytimes.com (8), pcmag.com (7), rtings.com (6).

• Copilot cited YouTube in 7 of 30 answers. Most cited: youtube.com (7), forbes.com (5), toolradar.com (4), rtings.com (4).

• Gemini cited Reddit in 4 of 30 answers. Most cited: zapier.com (6), reddit.com (4), pcmag.com (3).

• ChatGPT cited neither Reddit nor YouTube. Most cited: wired.com (4), rtings.com (3), tomsguide.com (3), techradar.com (3).

• Wikipedia was cited in none of the 120 answers. Most of our questions were commercial; a set of factual questions would probably look different.

Review publishers (rtings.com, techradar.com, pcmag.com, wired.com) are the closest thing to common ground: each shows up in the top list of at least two engines.

Our own question

"What is the best SERP API for developers?" is in the set because it is our market. All four engines named SerpApi in their answers. None named Searlo. Copilot listed searlo.tech among its 11 sources, but the answer text did not mention us.

That is the gap an AI visibility program has to measure: being cited is not the same as being recommended. A page can be read and linked by an engine and still not make it into the sentence a buyer reads.

What this means if you track AI visibility

1. Track each engine separately. With overlaps between 0.01 and 0.19, one engine is not a proxy for another. A tracker that samples only ChatGPT tells you nothing reliable about Perplexity.

2. Count mentions and citations separately. The SERP API question shows the two can disagree about the same brand in the same answer.

3. Sample each prompt more than once. We asked each question once, and AI answers vary from run to run, so treat single-run numbers like these as a snapshot. Our brand tracking tutorial covers how many samples you need.

4. Know where each engine looks. Reddit threads matter to Perplexity and Gemini, YouTube to Copilot, and review publishers to all four.

Limits of this study

• 30 questions, one run each, one day, in US English, from signed-out sessions. Signed-in users, other countries and other days will see different answers.

• The questions lean commercial ("what is the best…").

• Domains are counted by hostname, so subdomains of one site count separately.

• We sell the API used to collect the data.

Run it yourself

The listing below asks every engine every question, one call at a time, and prints the mean overlap for each pair. The same questions on all four engines cost 22 credits per question at September 2026 prices (Perplexity 2, Gemini 4, ChatGPT 8, Copilot 8), so the full 30-question study is 660 credits: about $0.20 at the Scale pack's $0.30 per 1,000 credits. Failed calls are refunded.

python
# citation_overlap.py: ask each engine the same questions and measure how
# much the domains they cite overlap. Needs SEARLO_API_KEY in the environment.
import itertools
import os
import time

import requests

API = "https://api.searlo.tech/api/v1/search/ai/{engine}"
HEADERS = {"x-api-key": os.environ["SEARLO_API_KEY"]}
ENGINES = ["chatgpt", "perplexity", "copilot", "gemini"]
QUESTIONS = [
    "What is the best CRM for a small business?",
    "How can I reduce my AWS bill?",
    "What is the best password manager?",
    # ... add your own
]


def cited_domains(engine, question):
    r = requests.get(
        API.format(engine=engine),
        params={"q": question, "gl": "us", "hl": "en", "cache": "false"},
        headers=HEADERS,
        timeout=150,
    )
    r.raise_for_status()
    return set(r.json().get("sources") or [])


def overlap(a, b):
    union = a | b
    return len(a & b) / len(union) if union else None


cited = {}
for question in QUESTIONS:
    for engine in ENGINES:
        cited[(engine, question)] = cited_domains(engine, question)
        # One call at a time: identical calls in flight are merged into one run.
        time.sleep(4)

for e1, e2 in itertools.combinations(ENGINES, 2):
    scores = [overlap(cited[(e1, q)], cited[(e2, q)]) for q in QUESTIONS]
    scores = [s for s in scores if s is not None]
    if scores:
        print(f"{e1:>10} and {e2:<10} mean overlap {sum(scores) / len(scores):.2f}")

sources is the list of unique cited domains in each answer; citations has the full URLs and titles if you want page-level overlap instead. See the AI answers docs for the response fields.