We counted the ChatGPT citations in one real run — every link, every domain, and the answers that turned out to be about somebody else.
Asked 12 questions on 15 August 2026, ChatGPT attached 44 links drawn from 25 distinct domains — an average of 3.7 links per answer, with every single answer carrying at least one. Claude, asked the same 12 questions on the same day through the same gateway, attached 0 links, while its transcripts show it had read 277 pages to write those answers. Both numbers are real and they are not measuring the same thing: one engine hands you a list of sources, the other reads and does not show its work. That is the single most useful fact about ChatGPT citations, and it means advice about “getting cited by AI” does not transfer between engines unedited.
Looking for how to cite ChatGPT in APA or MLA for a paper? That is a different question and this is not that page — your university library’s style guide covers it. This page is about which sources ChatGPT attaches to its own answers.
The 44 links, grouped. The grouping is the only judgment on this page, so it is printed rather than summarised:
| Kind of source | Domains | Links |
|---|---|---|
| A business’s own website (practices, clinics, one hospital system) not named here — see the rule below | 16 | 33 |
| Review sites and directories Birdeye, RateMDs, Trustindex, Yahoo Local, toprateddentist.com, emergencydental.com | 6 | 8 |
| Forums | 1 | 1 |
| Manufacturers and professional bodies Invisalign, the American Dental Association | 2 | 2 |
Most of what ChatGPT cited was somebody’s own website. 33 of the 44 links went to a provider’s own domain; the rest came from review platforms, one forum thread, a manufacturer and a professional body. For a local business that ordering is the practical finding: the assistant was mostly reading ordinary business sites, and the review platforms it used are a short, nameable list.
Why the individual domains are not listed. Most of them belong to real small businesses that did not ask to appear in anyone’s marketing, and one of them is a doctor’s own name. We name platforms; we do not name the individual businesses. The full list, unredacted, is in the transcript file the client receives.
The 12 questions were split across four intents — the same mix the paid audit uses, with the wording taken from what a customer would actually type:
| Question type | Questions | Links | Example |
|---|---|---|---|
| discovery | 3 | 13 | “best dentist in Ballard Seattle” |
| comparison | 3 | 12 | “Bayview Dental Studio vs Ballard Family Dental” |
| problem | 3 | 12 | “how much does Invisalign cost in Ballard Seattle” |
| brand | 3 | 7 | “is Bayview Dental Studio good” |
Discovery questions — the ones with no business name in them — produced the most links, which is the good news and the bad news at once: that is where a stranger finds you, and it is also where the assistant is choosing among sources with no help from your brand.
4 of the 12 ChatGPT answers were not about our test business at all. The name we used was invented; the assistant matched it to a similarly-named practice in another country and answered about that one, with confidence.
“is Bayview Dental Studio good”
“Do you mean Bayview Dental
Studio in South Surrey/White Rock, BC, run by Dr. [name]?
If so,
it appears generally reputable but not unanimously reviewed
…”
The doctor’s name is removed — he is a real person. Everything inside the quotation marks is otherwise word for word.
Those 4 answers carried 10 of the 44 links, and the most-cited domain among them was that other practice’s own site. Strip them out and you are left with 34 links across 21 domains. A citation count is only as good as the entity resolution behind it — and a tool that matches business names as strings, as ours did on this run, will hand you a source table that is partly about somebody else.
In this run, 3.7 on average: 44 links across 12 answers, and every answer carried at least one. That is one run of one question set on one day, so treat it as a shape rather than a rate.
Claude answered the same 12 questions the same day and attached 0 links, while its transcripts list 277 pages read. The two engines mean different things by 'source': one publishes the links it used, the other reads and does not surface them. We have not established whether that is Claude's own behaviour or an artefact of the gateway we called it through, and the page says so rather than guessing.
No. There is no ranking to occupy. A citation means the assistant retrieved your page while writing one answer to one question, and the same question asked again can produce a different set.
Ordinary business websites: 33 of the 44 links went to a provider's own domain. Review platforms and directories accounted for 8, one forum thread for one, and a manufacturer and a professional body for the rest.
Yes, with more questions and each one asked three times instead of once. The transcripts, the links, the model version and the timestamps are handed over as files; the audit page describes exactly what the deliverable contains.
raw.jsonl and runs.jsonl for 12 ChatGPT answers and 12 Claude answers, model gpt-5-6-sol and claude-sonnet-5. Every count on this page is computed from those files when the page is built.