ChatGPT doesn't always give the same answer: ask the same question fifty times and you'll get fifty slightly different answers — and your brand shows up in forty-two of them, disappears in eight. It's not an error on your site: it's how the model works internally. The problem is that this variability is worth money: every time you don't show up, a potential customer doesn't see your name. There are precise techniques to stabilize your presence and drastically reduce the answers in which you disappear.
I ran the same query 50 times on ChatGPT. 42 times the first brand cited was the same. 8 times it changed. Temperature and sampling cause these variations — and you have little control.
The query was “best producers of Carnaroli DOP” and I was testing it for a historic rice mill in Vercelli that had asked me why, in AI answers, it sometimes appeared and sometimes didn’t. The answer lies in the model’s internal behavior — and in the way ChatGPT decides whether to cite you.
In this article I explain why, with ChatGPT, it barely makes sense to aim for a link in the sources: you have to aim to get into the text of the answer as a cited name. And I’ll show you the 50-query test I use to figure out whether a brand is in the “stable zone” of AI answers or in the volatile zone.
How ChatGPT’s sampling really works
Before talking about citations, we need to understand a piece of the mechanics. ChatGPT doesn’t always produce the same answer to the same question. It’s not a bug: it’s how token-by-token generation works.
In a study by Renze and Guven (2024) this mechanic is described in black and white in an ACL paper dedicated precisely to the effect of sampling temperature on LLM performance.
In addition, we aim to provide a systematic study with empirical results to add to the growing body of knowledge used to create LLM and prompt-engineering best practices. Sampling temperature is a hyperparameter of an LLM used in a temperature-based sampling process.
Translated for your business: temperature is the knob that decides how “creative” ChatGPT is in choosing the next word. At a high temperature it introduces more variation, at a low temperature it tends to always pick the most probable token.
The operational consequence is simple. If your brand is the most probable token for a given query, you get cited almost every time. If you’re the third or fourth most probable, you get cited sometimes yes and sometimes no. The work of visibility in AI answers is all here: moving yourself from the volatile zone to the stable zone.
Why ChatGPT almost never cites you with a link
If you’ve already tried asking ChatGPT “who are the best producers of X?”, you’ll have noticed one thing: brands are named inside the text, often without any reference link. If you insist with “give me the sources”, the model adds a list — but the quality of the URLs is often low, sometimes the links are made up or point to generic pages.
It’s a different behavior from Perplexity, which works in retrieval-augmented mode with explicit citations for every sentence. If you want to understand the difference between the two systems, I covered it in the previous articles in this series on AI platforms.
This changes the strategy. It makes no sense to optimize for “being linked by ChatGPT as a source”: it’s something that happens rarely and with little impact. Instead, it makes sense to work so that your name is part of the answer. When someone asks “best producers of Carnaroli DOP in Italy”, the goal is for ChatGPT to write “Riseria X of Vercelli” inside the paragraph, not to put it in a footnote at the end.
It’s the same principle underlying concepts like implicit reference weight and backlinks as a citation proxy: the nominal citation, without a link, weighs more than the link itself in ChatGPT’s logic.
If the brand isn’t named in the body of the article (because “after all, we’re the author, it’s obvious”), the model doesn’t build the association during training.
The 50-repetition test I use to measure stability
The test is trivial to run and tells you one precise thing: whether your brand, for a relevant query, is in the stable zone or the volatile one of ChatGPT’s answers.
The method I used for the Vercelli rice mill, on the query “best Italian producers of Carnaroli DOP”:
- I opened 50 new ChatGPT chats (one at a time, with history turned off)
- Same query, identical word for word
- I recorded the first 3 brands named in each answer
- I counted how many times each brand appeared at least once in the top 3 positions
Result for my client: it appeared in 42 answers out of 50. In the other 8 cases, two competitors in Vercelli rice took its place. The brand already had solid authority work behind it — studies on the Carnaroli variety, a presence on Wikidata, editorial coverage in food publications. It was in the stable zone, but not yet in the cluster of the always-cited.
How to read the numbers, for my own use:
- Above 40/50: you’re in the stable zone. Work on marginal increments.
- Between 20 and 40/50: volatile zone. There’s a partial but not consolidated citation.
- Below 20/50: absent zone. The brand isn’t in the model’s mental shortlist for that query.
Let me state the limits of this test up front: it’s 50 repetitions, not 500. It’s an indicative test, not an academic study. For a real analysis you need professional AI-search monitoring tools with much larger volumes and more engines in parallel. But as a first step to understand where you stand, it works.
Identify 3 relevant queries for your sector.
Why temperature explains the swings (but not everything)
Someone might think: “I’ll lower the temperature and always get the same answer”. The Renze & Guven (2024) paper says something interesting on exactly this point.
“Despite anecdotal reports to the contrary, our empirical results indicate that changes in temperature from 0.0 to 1.0 do not have a statistically significant impact on LLM performance for problem-solving tasks.”
Translated: despite many telling it differently, changing the temperature from 0 to 1 does not significantly change LLM performance on problem-solving in statistical terms.
From this follows something important for your brand. The swings in citations aren’t just “temperature noise”: they reflect how strong the link is between your name and the query’s concepts within the model’s representations. If you’re close to the center of the “Carnaroli DOP” concept, you show up stably even at high temperature. If you’re on the periphery, the variation makes you come in and go out.
Working on the stable zone means working on the name-concept association — and it’s work done on content, not on ChatGPT’s parameters.
The mistakes I see most often
When a company discovers it doesn’t show up in ChatGPT’s answers, the typical reactions are almost always the same. Here’s a summary.
Trying to get linked as a source. Investments in the link-building profile, thinking that “if ChatGPT sees the link, it cites me”. That’s not how it works: ChatGPT rarely adds links, and when it does they’re often wrong. Better to work on nominal citability.
Signing content with the full legal company name. “Riseria Fratelli X S.r.l. Società Agricola” never makes it into an answer. ChatGPT cites short forms: “Riseria X” or “X of Vercelli”. If your brand name isn’t citable in 2-4 words, you’re at a disadvantage from the start.
Writing articles where the brand only appears in the footer. If the brand isn’t named in the body of the article (because “after all, we’re the author, it’s obvious”), the model doesn’t build the association during training. You need content where the phrase “According to [Brand]’s experience…” appears naturally in the text.
Expecting results from a single test. One query, one answer: it means nothing. Without at least 30-50 repetitions you don’t know whether you’re in the stable zone or reading a random variant.
A 30-minute audit to figure out where you stand
Here are three operational steps to run a first check yourself. It doesn’t replace a professional analysis, but it gives you a concrete idea.
- Identify 3 relevant queries for your sector. For the Vercelli rice mill they were “best producers of Carnaroli DOP”, “historic rice mills Piedmont”, “authentic Carnaroli Italy”. No self-referential queries: you have to use the language of your end customer.
- Repeat each query 30-50 times on ChatGPT, with new chats and history turned off. Record the brands cited in the top 3 positions.
- Compare with the 3-5 competitors the AI cites most consistently. If your brand appears fewer than 20 times out of 50, work on the stable zone: content that associates you with the key concepts, not just with your name.
To understand whether the key concepts are recognized as autonomous entities by the system, I also recommend going through Named Entity Recognition and entry into the Google Knowledge Graph: these are two pieces that sit upstream of nominal citation on ChatGPT.
The thread on AI visibility
All this work serves one purpose only: bringing your brand from the volatile zone to the stable zone in ChatGPT’s answers. It’s not a magic factor, it’s not enough on its own, and it isn’t solved with a trick. You need to work on nominal citability in content, on the name-concept association, and on consistent editorial coverage over time.
In the next articles in this series on AI platforms I’ll take you inside the differences in behavior between Claude, Gemini and Perplexity — because every engine has its own citation logic and the strategy changes depending on where you want to be visible.
