Why ChatGPT Is Not Citing Your Website: I Measured 47 AI Answers

ChatGPT is not citing your website because you are not in the vendor set it already carries for your category, and you have never measured which vendors it does carry. You find out by asking the engines the questions your buyers ask, recording every brand named in the answers, and counting how often yours appears. I ran that on my own category across three engines and 47 answers. My site was named 3 times.

Two months ago I ran five buyer queries on Perplexity and got cited zero times out of five. That was a small sample and an ugly result. This is the same experiment done properly: 16 questions, three engines, 48 attempts, 47 usable answers, run on 2026-08-13.

The result did not get better. It got more precise, which is more useful.

The short answer (the method)

To find out why an answer engine is not citing you, stop auditing your page and start auditing the answer. Your page is one input. The answer is the outcome. Measuring the page tells you whether you followed best practice. Measuring the answer tells you whether it worked.

The method is four steps, and it needs no special tooling:

1. Write down the questions a buyer types before they buy. Not keywords. Questions.
2. Ask every engine each question, in a clean session with no personalization.
3. Record every brand named in each answer, including your competitors and yourself.
4. Count. Share of voice is mentions divided by answers.

That is the whole instrument. The value is not in the sophistication. It is in the fact that almost nobody runs it, so almost nobody knows their real number.

What the 47 answers showed (the data)

Sixteen buyer questions, asked on Claude, ChatGPT, and Perplexity on 2026-08-13. One attempt failed and is excluded, leaving 47 usable answers. Here is who the three engines named:

Brand	Answers naming it	Share of voice
Semrush	24	51%
Profound	23	49%
Otterly	23	49%
Peec AI	19	40%
Ahrefs	18	38%
Scrunch AI	12	26%
Athena	4	9%
citability.dev (mine)	3	6%
Similarweb	2	4%
Conductor	2	4%
Evertune	2	4%
Writesonic	2	4%

Split by engine, my own three mentions land in one place:

Engine	Answers	Times it named me
Claude	15	3
ChatGPT	16	0
Perplexity	16	0

Two of the three engines have never heard of me in a buying context. The third names me only on questions about measuring citations and checking crawler access, never on the plain question “what are the best tools for this.”

That distinction matters more than the totals. Being named on a narrow technical question means the model knows what you do. Being absent from the general recommendation means it does not consider you a vendor. Those are different problems with different fixes, and you cannot tell them apart without running the questions separately.

The finding I did not expect

I went in looking for my own number. The useful result was a different column: how often the engines named nobody at all.

Engine	Answers	Answers naming no vendor
Claude	15	1
ChatGPT	16	6
Perplexity	16	9

Perplexity declined to name a single vendor in more than half its answers. ChatGPT in over a third. Those questions have no incumbent. There is no brand to displace, because no brand is there.

One question came back empty on all three engines: “How do I find out why ChatGPT is not citing my website?” Nobody owns that answer anywhere. That is why you are reading this page.

Eight more questions were empty on at least one engine, including how much a visibility audit costs, how to check whether AI crawlers can reach your site, and which consultants do this work. Those are buying questions with real intent and no answer.

What this changes about the work

The instinct when you see a table like the first one is to try to beat Semrush. That is the wrong read, and it is expensive.

Semrush appears in half the answers because it is a large established brand with a decade of corpus behind it. You are not going to displace that with better structured data. The competition on those questions is settled for now.

The unowned questions are a different game entirely. There is no incumbent, the model has nothing to pull from, and the first genuinely useful answer tends to become the answer. Winning an empty question costs a good page. Winning a crowded one costs years.

So the order of work inverts. Find the empty questions first, answer them properly, then worry about the crowded ones.

Three things make an answer likely to be the one an engine pulls, and all three showed up in the pages my competitors won with:

• The question is the heading, and the answer is directly underneath it. Models lift a clean claim near a matching heading. They do not dig.
• The claim is specific and checkable. “47 answers, 3 mentions, measured 2026-08-13” survives extraction. “Most sites struggle with visibility” does not.
• The numbers are yours. Original measurement is the only thing on your page that is not already somewhere else in the corpus. It is also the only reason to cite you rather than the source you copied.

That last one is why I published my bad number instead of a good estimate. A real 3-out-of-47 is worth more than an invented benchmark, and it is the only version I can defend when someone checks.

Run this on your own category

You do not need my tooling. You need an afternoon.

1. Write 15 to 20 buyer questions. The ones a customer types the week they are ready to spend. Include pricing questions and “who does this” questions. Skip anything that sounds like a keyword.
2. List every competitor you would lose a deal to, plus every big generalist brand that might crowd the category. Fifteen names is plenty.
3. Ask each engine each question. Use a clean session. Personalized results will flatter you.
4. Record every brand named per answer, then divide by the number of answers to get share of voice.
5. Sort the questions by how many engines named nobody. That column is your roadmap.

Two warnings from running it. Match brand names, not just domains, because a model says “Profound” far more often than it says the web address, and matching only the address will make everyone including you look absent. And do not average your engines together as one score. Mine disagreed sharply, and the disagreement was the most actionable thing in the dataset.

If you run this on your category and find a gap, the AI visibility audit page covers how to read the result and decide whether it is worth fixing, and the AI Discovery Review is the same measurement done for you.

What I am doing about my own number

I am not going to chase the crowded questions. I am answering the empty ones, starting with this page, and I am rerunning the same 16 questions on a schedule so the next number is comparable to this one.

If it works, share of voice moves. If it does not, I will publish that too.

If you would rather not run it by hand, that measurement is what I built citability.dev to do: the same questions, the same counting, on a schedule, with the competitor set filled in for you.

Method note: 16 questions, 3 engines, 48 attempts, 47 usable answers, collected 2026-08-13. One Claude attempt failed and was excluded rather than counted as an absence, because a failed request and a genuine non-mention are not the same thing and averaging them together understates every brand in the table. Brands were matched on name and common aliases as well as domain.