Ask ChatGPT to recommend a company in your industry and the same three or four names always come up — never yours. It's not a matter of quality or size: it's the result of a precise technical mechanism that tends to keep surfacing the same established sources. The brands that made it onto that list stay there — and they collect customers every day without doing anything. The ones off the list never appear, even when they would be the best choice. Understanding how that mechanism works is the first step to getting in.
Run an experiment. Ask ChatGPT “what’s the best email marketing tool?” and write down the answer. Then ask the same question nine more times. The names will be almost identical: Mailchimp, HubSpot, ActiveCampaign, always those, more or less in the same order.
I actually did this a few weeks ago, across 30 different commercial queries — recommendations for tools, services, professionals — repeating each one 10 times on ChatGPT, Gemini and Perplexity. The result: on ChatGPT and Gemini, the brands in the top 3 positions were the same in 85% of cases. On Perplexity the variation was greater — but we’ll get to that shortly, because the reason is illuminating.
This repetitiveness isn’t a flaw. It’s the result of how text generation works, and there’s a specific parameter that governs it.
Temperature: the thermostat of creativity
When an LLM generates a response, for each word it computes a probability distribution over all the possible next tokens. It doesn’t “choose” — it calculates how likely each option is, and then draws one. The parameter that controls this drawing is called temperature.
The literature describes it very precisely. In the study on model safety by Jun Zhuang et al. (2025), the authors explicitly set:
“The temperature was set to 0 for reproducibility.”
(Exploring the Vulnerability of the Content Moderation Guardrail in Large Language Models via Intent Manipulation)
That “for reproducibility” is the key. Temperature at zero means: the model always picks the token with the highest probability. No variation, no creativity, identical results every time. And that’s exactly the setting most business applications use — because when a company integrates AI into customer service or recommendations, it wants predictable answers, not surprises.
The mechanism is easy to grasp even without formulas. At low temperature (0.0-0.3), the model is conservative: it almost always takes the most well-trodden path. At high temperature (0.7-1.0), it spreads the probabilities more evenly, leaving room for less obvious options.
In the survey by Minaee et al. (2025) on LLMs, there’s a passage that explains it with a concrete numerical example:
“Suppose we have 6 tokens (A, B, C, D, E, F) and k=2, and P(A)= 30%, and P(B)= 20%. In top-k sampling, tokens C, D, E, F are disregarded, and the model outputs A 60% of the time, and B, 40% of the time.”
(Large Language Models: A Survey)
Applied to your business: if token A is “Mailchimp” and token B is “HubSpot”, top-k cuts out all the others: tokens C, D, E, F — that is, the other email marketing tools — stay below the threshold of the top k and are never generated. Not because they’re worse, but because their probability doesn’t make it into the top spots. And at low temperature the model concentrates even more on the leading tokens.
Hence the deduction: if you’re not on the list, the thermostat cuts you out
This is the point where the documented mechanism turns into reasoning about your business — and I want to be clear that this is a deduction, not a direct experimental result.
The reasoning is this: a token’s probability depends on how often the model saw that sequence during training. A brand cited heavily on authoritative sources — Wikipedia, national media, industry directories, academic papers — has built up high frequency in the corpus. When the user asks for a recommendation, that brand is the high-probability token.
At low temperature, only the high-probability tokens get through. It follows that if your brand hasn’t accumulated enough frequency on the right sources, it doesn’t appear — regardless of the quality of your product.
And here’s why results vary more on Perplexity: Perplexity uses RAG, searching the web in real time before answering. It doesn’t depend only on the training data. If your site is well structured and indexed, you can show up on Perplexity even if ChatGPT ignores you. It’s a different access window, and for many brands it’s the most accessible one in the short term.
If an agency calls itself “brand strategy specialists” on its website, “full-service creative agency” on LinkedIn, and “digital marketing consultants” on the directory, it’s spreading its probability across three different token sequences.
The vicious circle — and the only way to break it
There’s a structural problem worth facing head-on. The brands the AI recommends today receive more traffic, more mentions, more citations — which increases their probability in future training. It’s a self-reinforcing loop.
I analyzed 10 Italian B2B niches and in each one there were 3-4 brands that appeared in 80%+ of the AI responses, with everyone else under 10%. The gap didn’t correspond to actual market share — it corresponded to presence on high-authority sources.
A communications agency in Bologna — with major clients and solid case studies — never appeared on ChatGPT for the query “communications agency Bologna”. The ones that did appear had two things in common: a Wikipedia page (even just as a source cited in an existing entry) and at least 3-4 citations on local media with the full name associated with the specialization.
You can’t change past training. But you can build the presence for future training:
- Get cited on media in your industry with the brand tied to the specialization
- Be present on Wikipedia as a source cited in relevant entries
- Have a detailed listing on industry directories (not generic ones)
- Keep the site optimized for RAG systems — which are the bypass to low temperature
List the 10 most important queries for your business.
Consistency beats volume
One thing that emerges when you analyze the brands the AI recommends: they aren’t necessarily the ones with the most mentions in absolute terms. They’re the ones with consistent mentions.
If an agency calls itself “brand strategy specialists” on its website, “full-service creative agency” on LinkedIn, and “digital marketing consultants” on the directory, it’s spreading its probability across three different token sequences. None of them reaches critical mass. The AI doesn’t know which association is the right one, and in low-temperature contexts — where only the most probable token gets through — none of the three variants clears the threshold.
Ten consistent mentions on authoritative sources beat a hundred scattered, contradictory ones. Low temperature rewards those who concentrated the signal, not those who dispersed it.
A test you can run right now
Before building any strategy, you need to know where you stand. The “10-query test” is the one I use with clients:
List the 10 most important queries for your business. Run each one on ChatGPT (with browsing off), Gemini and Perplexity. For every answer, note who gets mentioned. Repeat in 30 days.
If your brand doesn’t appear on ChatGPT but does appear on Perplexity, the problem is in the training data — and the solution is to build presence on authoritative sources for the next training cycles. If you don’t appear on any of the three, the problem is broader and probably concerns the very structure of your content — and here is where the brand’s tokenization, the positional encoding and the context window structure come into play.
Temperature is the mechanism that filters. The log-probability is the score that gets filtered. If your score is low, no temperature value will save you.
