Ask ChatGPT the questions your buyers ask, logged out and in a new chat, and note two things in every answer: whether your brand is named and which pages are cited. Repeat each question twice and on a second engine. In 240 answers we read on 26 October 2026, 36% of citations went to "best X" lists.
That is the check in one paragraph. The rest of this guide is what makes it reliable: the difference between a mention and a citation, the questions to ask, what six AI engines actually cited when we ran the same check across 48 questions, and how to repeat it with our free AI visibility audit or with one API call on a schedule.
Mentioned or cited: two different results
An answer can name your brand without linking you, and it can link a page of yours without naming you. The two mean different things, so record them separately.
- Mentioned: your brand appears in the text of the answer. The model knows you and thinks you fit the question.
- Cited: one of your pages is among the sources the engine linked. The engine found that page while searching, and it is the page a reader clicks.
A brand that is mentioned but never cited has no page the engine could find for that question. A brand that is cited but not mentioned has a page that answers well, but the answer does not recommend it. For every answer, write down these fields:
| Field | What it tells you |
|---|---|
| Question, engine, date, signed in or not | Answers change by engine, by day and by account, so a result without these is not comparable |
| Brand named (yes or no) | Whether the model recommends you for this question |
| Your pages cited (full URL) | Which page of yours the engine found, and its position among the sources |
| Other domains cited | Who wins the question where you are absent: a list, a forum thread, a vendor page |
| Kind of each cited page | List, homepage, product page, guide, forum, video: it tells you what to publish or where to be listed |
How to run the check by hand
- Write ten questions in three groups. Generic ("best invoicing software for freelancers"), specific (a task your product does, phrased the way a buyer types it), and brand ("is Example legit?", "Example reviews"). The three groups are won by different kinds of pages, as the data below shows.
- Open ChatGPT logged out, one new chat per question. History and memory steer the answers of a signed-in account; a fresh, logged-out chat is closest to what a new buyer sees.
- Open every citation. OpenAI's help center says to select a citation to open its source, and a Sources button lists them when it is available. In our run the logged-out app showed no separate Sources panel, so we read each URL from the citations themselves. Record the full URL, not only the domain.
- Ask each question twice. Answers vary between runs. In our panel, differences of one or two citations between two runs of the same question were noise; a page that appears in both runs is a real result.
- Repeat on Perplexity, Google AI Mode and Grok. Each engine reads different sources, as the next section shows. Note when you are signed in: signed-in engines can personalize.
Before reading anything into a zero, make sure the engines can reach you at all. OpenAI's crawler documentation says that sites opted out of OAI-SearchBot are not shown in ChatGPT search answers. The AI crawler access checker reads your robots.txt for each AI user agent, and the AI crawler user-agent list lists them.
What 240 AI answers cited on 26 October 2026
We ran this check on our own market: 48 questions about proxies, scraping APIs and brand reputation, in English plus Italian, German, Spanish and French. Six engines answered:
| Engine | How we asked | Answers |
|---|---|---|
| ChatGPT (chatgpt.com) | Browser, logged out, English interface | 48 |
| OpenAI API with web search | gpt-4.1 with the web search tool, through our AI visibility tool | 87 |
| Google top 10 handed to DeepSeek | Our SERP API for the results, then the model writes the answer | 39 |
| Perplexity | Browser, free plan | 26 |
| Grok | Browser, signed in, Fast mode | 24 |
| Google AI Mode | Browser, signed in | 16 |
The 240 answers carried 1,821 citations (for Grok we counted only the numbered citations in the text). We sorted every cited URL by the kind of page it is:
| Kind of page cited | Share of the 1,821 citations |
|---|---|
| "Best X" and "cheap X" lists | 36%: 19% written by companies that sell in the category, 17% by third parties |
| Vendor homepages | 13% |
| Vendor product pages | 9% |
| Profiles and reviews on third-party sites | 5% |
| Reddit threads | 4% |
| Country pages, use cases, pricing pages, YouTube, benchmarks, documentation | 1% to 3% each |
Lists written by a vendor were cited as if they were neutral, even when the vendor had put itself first. On generic questions, the page to be on is a list.
Every engine reads different sources
The same question sent to six engines did not produce six versions of one answer. Each engine leaned on its own kind of source:
- ChatGPT on chatgpt.com gave 41% of its citations to vendor homepages, one card per vendor with a link to the home page. It did not cite Reddit once in 48 answers.
- The OpenAI API with web search cited Reddit in 32 of its 87 answers. Same company, different product, different sources: a check run through the API approximates ChatGPT, it does not replace it.
- Google AI Mode cited YouTube in 13 of 16 answers and Reddit in 8.
- Grok read far more than it cited: its sources panel listed 1,136 search results across 24 answers, 22 to 73 per answer, but the text carried 57 numbered citations.
- Perplexity, in a second run of 32 answers, searched the question word for word in 30 of them, with one search each. A page whose title is the question is the page it finds.
Two more patterns held across engines. On brand questions such as "is it legit?", the answers did not lean on ranking lists: they were built from trust-score sites and review profiles. And new pages enter fast: a third-party page published on 24 October was already cited in 5 answers on 26 October.
Specific questions are won by dedicated pages
The 48 questions above were generic. The same day we ran a second set of 303 specific questions, taken from where people really ask them: Google suggestions, the first questions customers send us (rewritten as generic questions), and the pages AI assistants and search engines send visitors to. The OpenAI API answered all 303; Perplexity and ChatGPT answered 32 each.
| Kind of page cited | Generic questions | Specific questions |
|---|---|---|
| "Best X" lists | 36% | 9% (6% by vendors, 3% by third parties) |
| Vendor product pages | 9% | 31% |
| Country pages | 1% to 3% | 10% |
| Use-case pages | 1% to 3% | 6% |
Guides and how-tos took another 8% on the specific questions. On a specific question, the engines cite the page that answers that question. So test both kinds: a brand that never appears on "best X" questions can still win the specific ones its pages answer, and those are the questions closest to a purchase.
Run the same check with the free AI visibility tool
Our AI visibility audit does one round of this check without an account:
- Paste the URL of the page that should be cited, such as your pricing page or the guide that answers the question.
- Type the question a buyer would ask.
- The tool asks one live answer engine and reports whether your domain was cited, whether your brand was mentioned in the text, and which domains the engine relied on instead.
The same run checks whether an assistant can read the page at all: robots.txt for 24 AI crawlers, a second fetch with GPTBot's user agent, the text that arrives without JavaScript, structured data, and whether the copy opens with an answer. A blocked crawler or an empty HTML shell explains a zero before any content work does; AI crawlers and JavaScript covers the second case. The tool is free, with a few audits a day.
Run it on a schedule with the API
One manual round is a snapshot. To see whether a page you published changes anything, run the same questions every week and keep the rows. POST /v1/scraper/ai-visibility takes the page, up to 10 questions, the engines, your brand name and up to 20 competitor domains, and answers with one row per question and engine.
import os, requests
r = requests.post("https://api.quanticdata.io/v1/scraper/ai-visibility",
headers={"Authorization": f"Bearer {os.environ['QD_API_KEY']}"},
json={
"url": "https://example.com/pricing/",
"brand": "Example",
"queries": [
"best invoicing software for freelancers",
"how do I send recurring invoices in euros?",
"is Example legit?",
],
"engines": ["openai", "deepseek"],
"competitors": ["rival-one.example", "rival-two.example"],
"country": "us",
},
timeout=300)
data = r.json()
if data["type"] != "response":
raise SystemExit(data["message"])
for row in data["payload"]["citations"]["rows"]:
print(row["engine"], "cited" if row["cited"] else "-",
"mentioned" if row["mentioned"] else "-", row["query"])
print(" sources:", ", ".join(row["citedDomains"][:5]))Each row says whether the page was cited, at which rank among the sources, whether the brand was mentioned, and the citedDomains the engine used, so the weekly file shows who wins each question where you are absent. With competitors set, the answer adds share of voice for those domains. Every field is in the API reference for POST /v1/scraper/ai-visibility, and agents in Claude, Cursor or any MCP client get the same call as the ai_visibility tool of the QuanticData MCP server.
The price, from the reference: $0.01 for the page audit, $0.01 for each question and engine that answered, and $0.0009 for each of the two searches that check whether the page can be found. The call above, 3 questions on 2 engines, costs $0.07 plus $0.0018. A full panel of 10 questions on 2 engines costs $0.21 plus $0.0018, so the $2 of free API usage every account gets each month covers nine of them. Engine calls have a daily cap per account; past it the API answers 429 with the limit and the time it resets. The same audit for many pages at once is the SEO and AI audit API.
What to do with the result
| What you see | What it means | What to do |
|---|---|---|
| Named in the text, never cited | The model knows you, but the engine found no page of yours that answers the question | Publish a page whose title is the question and whose first sentence answers it |
| A list wins every generic question | Lists took 36% of citations on generic questions | Get into the independent lists with facts they can check, and publish your own comparison by criteria |
| Cited on specific questions, absent on generic ones | Normal: dedicated pages win specific questions | Cover more specific questions, one page each |
| Nothing cited, and the audit shows a blocker | A crawler is blocked, the page is empty without JavaScript, or it carries noindex or nosnippet | Fix the blocker first; which AI crawlers to allow and which to block explains the choice |
| Brand questions answered from review profiles | On "is it legit?" questions engines read trust and review sites | Keep your profiles there claimed, accurate and current |
| The answer describes a different company | The engine confused your name with a similar one | State who you are: Organization markup with sameAs links, an about page, the same name everywhere |
Then run the same questions again a week after each change. In our panel a third-party page was already cited two days after it went live, so a week is enough to see whether a new page is found.
