How We Tested What AI Engines Cite
432 creator questions through ChatGPT, Claude, Gemini and Perplexity on one day in September 2026. The method, the results in full, and what the test cannot show.
By Alex Nazarov
On this page
On September 9, 2026 we sent 432 creator questions to ChatGPT, Claude, Gemini and Perplexity with web search on, and recorded every source each one attached. This page is the method and the full result.
Four articles on this blog quote numbers from this run. Rather than make you take them on trust, here is what we did, what came back, and the places where the test does not support a broad claim.
Why we ran it
We wanted a baseline for how visible iSonic was in AI answers, on the questions our own audience asks. That meant sending a fixed set of creator questions to four engines and reading what came back.
The source counts on this page are a byproduct of that run. We kept every citation each engine attached, which turned out to be more interesting than the thing we set out to measure. It is worth saying plainly that this was our own prompt set, run for our own purposes, rather than an independent study of the engines.
How the test ran
| What | Detail |
|---|---|
| Date run | September 9, 2026 |
| Questions | 432, sent verbatim, fixed the day before the run |
| Engines | ChatGPT, Claude, Gemini and Perplexity, each pinned to one model version |
| How they were sent | DataForSEO AI Optimization API, LLM Responses live endpoint, with web search enabled |
| Answers captured | 1,728 of 1,728 planned |
| Length cap | 1,024 output tokens per answer |
| What we recorded | Every source each engine attached to its answer, plus whether it searched at all |
Each engine was pinned to one model version: ChatGPT gpt-5.5, Claude claude-sonnet-5, Gemini gemini-3.8-flash and Perplexity sonar. That matters, because a figure like “ChatGPT searched 35% of the time” describes one model on one day, not ChatGPT in general.
Every question went to every engine, worded identically, with no follow up and no system prompt beyond the API default.
What each engine did
| Engine | Answers | Questions it searched the web for | Sources attached per answer |
|---|---|---|---|
ChatGPT gpt-5.5 | 432 | 153 (35%) | 0.9 |
Claude claude-sonnet-5 | 432 | 225 (52%) | 2.2 |
Gemini gemini-3.8-flash | 432 | 214 (50%) | 3.4 |
Perplexity sonar | 432 | 432 (100%) | 16.4 |
The first column is the one most people guess wrong. An engine decides for itself whether a question needs a live search. ChatGPT decided it did not, for about two thirds of these questions, and answered from training with nothing attached.
The spread in the last column is just as wide. Perplexity attached roughly eighteen times as many sources per answer as ChatGPT. Both were answering the same 432 questions on the same day.
What the engines cited most
Across all 1,728 answers, these were the fifteen domains that appeared most often.
| Rank | Domain | Answers citing it | Share of all 1,728 answers |
|---|---|---|---|
| 1 | reddit.com | 284 | 16.4% |
| 2 | linkedin.com | 223 | 12.9% |
| 3 | medium.com | 123 | 7.1% |
| 4 | youtube.com | 119 | 6.9% |
| 5 | searchengineland.com | 73 | 4.2% |
| 6 | semrush.com | 67 | 3.9% |
| 7 | isonic.ai | 58 | 3.4% |
| 8 | arxiv.org | 51 | 3.0% |
| 9 | developers.google.com | 51 | 3.0% |
| 10 | ahrefs.com | 49 | 2.8% |
| 11 | amicited.com | 45 | 2.6% |
| 12 | forbes.com | 44 | 2.5% |
| 13 | digiday.com | 44 | 2.5% |
| 14 | searchenginejournal.com | 43 | 2.5% |
| 15 | isonicinc.com | 43 | 2.5% |
Two notes on that table. Thirty eight of the 432 questions name a brand, which is why isonic.ai appears. And isonicinc.com is an unrelated ultrasonic cleaning company that shares our name, which the engines reached for often enough to make the top fifteen on its own.
The shape of the rest is the finding. Four of the top five are places where people write about other people.
Where each engine went
The averages hide how differently the four engines read the web. These are each engine's ten most cited domains across its own 432 answers.
| Engine | Its ten most cited domains, with the number of answers citing each |
|---|---|
| ChatGPT | help.openai.com (24), developers.google.com (23), blog.google (13), isonic.ai (12), openai.com (11), app.isonic.ai (8), creativecommons.org (7), arxiv.org (7), copyright.gov (6), isonicinc.com (6) |
| Claude | semrush.com (12), searchengineland.com (12), isonicinc.com (11), arxiv.org (8), medium.com (7), otterly.ai (7), trysight.ai (7), matthewcanabarro.com (6), digiday.com (6), llmpulse.ai (6) |
| Gemini | youtube.com (102), reddit.com (51), medium.com (37), isonic.ai (20), facebook.com (17), searchengineland.com (17), google.com (16), searchenginejournal.com (13), forbes.com (12), isonicinc.com (12) |
| Perplexity | reddit.com (230), linkedin.com (216), medium.com (79), searchengineland.com (40), semrush.com (40), amicited.com (40), ahrefs.com (31), forbes.com (29), arxiv.org (27), developers.google.com (26) |
ChatGPT went to documentation. Six of its top ten are official docs or help centers from OpenAI, Google, Creative Commons and the US Copyright Office. Perplexity behaved like a forum reader. Gemini, which shares a parent company with YouTube, cited YouTube more than twice as often as its next source. Claude spread itself thin across industry blogs, with no domain above twelve answers.
How often social platforms were cited
This is the table that matters most if your work lives on social. Each number is the count of answers, out of that engine's 432, that cited the platform.
| Platform | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| YouTube | 0 | 0 | 102 | 16 |
| 0 | 0 | 17 | 6 | |
| 0 | 2 | 0 | 3 | |
| TikTok | 1 | 0 | 0 | 0 |
Read the zeros carefully. They do not mean ChatGPT cannot cite YouTube. They mean that across 432 creator questions on one day, it did not.
One caveat on this table specifically. Our per answer records keep the first ten domains an engine attached. For ChatGPT, Claude and Gemini that captures everything, because they averaged well under ten sources. Perplexity averaged 16.4, so its numbers here are floors rather than exact counts. Its real YouTube and Instagram totals are at least what this table shows and may be higher.
What the questions were about
The 432 questions were grouped into twenty topics before the run. The mix explains a lot about which sources came back.
| What the questions were about | How many |
|---|---|
| Citation discovery | 39 |
| Technical AEO foundations | 36 |
| Platform specific visibility | 33 |
| Creator vertical applications | 31 |
| AI invisibility diagnosis | 30 |
| Category and tool comparison | 27 |
| Question and answer libraries and repurposing | 27 |
| Trust, legitimacy, privacy and safety | 24 |
| Traffic loss from AI search | 23 |
| Consent, scraping and control | 23 |
| Branded questions about iSonic | 20 |
| Entity and authority building | 19 |
| Formats and multi platform publishing | 18 |
| Getting started and objections | 16 |
| Licensing and compensation | 14 |
| Measurement and proof | 12 |
| AI shopping and recommendations | 12 |
| Branded comparisons | 11 |
| The wider shift | 10 |
| Ownership, copyright and legal | 7 |
These are questions about AI, search and creator work. That is the single biggest limit on the results below.
What this test cannot show
Everything above describes one day, one prompt set and four pinned models. Here is where it stops.
- One topic. Every question was about AI, search or creator work. A cooking question or a gear question would pull a different set of sources, and almost certainly a different set of platforms.
- One day. September 9, 2026. These systems change their search behavior and their partners without announcing it. Treat the numbers as a dated snapshot.
- One model each. A different version, a different account tier or a consumer app rather than an API can all behave differently from what we measured.
- Answers were capped at 1,024 output tokens, so a source named deep in a long answer could be missed.
- Correlation, not mechanism. We can say what the engines cited. We cannot say why, and none of them publish a complete ranking formula.
- Our own prompt set. We wrote the questions and ran the test on ourselves. It is evidence, not a peer reviewed study, and we would rather you read it that way.
The articles that use this data
Four pieces on this blog quote the numbers above.
- How Does ChatGPT Choose Sources? uses the search rate and the sources per answer.
- How to Get Cited by ChatGPT as a Creator uses the YouTube and Instagram counts.
- Why Does AI Cite Some Websites and Skip Others? uses the top sources for the citation discovery questions.
- Can AI Search Cite Your Social Media Posts? uses the per engine table in full.
If a number in one of those pieces does not match this page, this page is right.
Quick answers
What exactly did this test measure?
Which sources four AI engines attached to their answers when asked 432 questions about AI and creator work, and whether each engine chose to search the web at all. It does not measure answer quality, accuracy or ranking.
Why did ChatGPT cite so few sources?
Because it searched for only 153 of the 432 questions. When an engine answers from training rather than live search, there is no source to attach and no creator to credit. That is the whole reason the number matters.
Can I reproduce this?
In principle, yes. Send a fixed question set to each engine's API with web search enabled and record the sources attached to each answer. You will not get our exact numbers, because the engines have changed since September 9, 2026 and will keep changing.
How current are these numbers?
They describe September 9, 2026 and nothing else. We date every figure for that reason. When we run the test again we will publish the new numbers here with the new date rather than quietly updating these.
Numbers without a method are decoration. If you are going to act on anything on this blog, this is the page that tells you how much weight it will hold.