Every time you search your industry on ChatGPT, the same three brands always come up — and yours is never one of them. You can't simply add yourself in: that list has limited slots, and to get in you have to take the place of someone who's already there. It's not random: it depends on who has built authority on the sources the model considers relevant. Doing more isn't enough — you have to do more in the right places. And those places are identified by analyzing exactly who comes ahead of you.
Ask an AI engine “what are the best tax consulting firms in Milan?” and count the names in the answer. Three, four, five. Never fifteen. Never all the ones that actually exist. If yours isn’t there, it’s not because the AI doesn’t know you: it’s because those slots are already taken by someone else. And to get in, you have to take the place of one of them.
This isn’t an opinion about how the AI market works. It’s how text generation works when the temperature parameter is low, and in commercial applications it’s almost always low. The slots are limited by the very mechanics of the system, not by an editorial choice.
Why the slots are limited: the mechanics of temperature
I covered this in depth in the article on temperature and sampling, but here I need to recall the central point. When an LLM generates an answer, it doesn’t “decide” which brands to recommend. It calculates a probability for every possible next token, and then selects one. Temperature is the parameter that regulates how conservative or open this selection is.
Gao et al. formalize it in their 2025 study:
“The randomness is usually introduced via the concept of temperature. The temperature T is a parameter that ranges from 0 to 1, which affects the probabilities generated by the softmax function, making the most likely tokens more influential.”
“Making the most likely tokens more influential” — this is the phrase that matters. At low temperature, high-probability tokens dominate. Tokens with medium or low probability are ignored. Applied to commercial answers: the brands the model has seen most often, on more authoritative sources, in tighter association with that category, get generated. The others don’t.
That’s why it’s 3-5 names, not 15. The model doesn’t produce an exhaustive list. It produces the tokens that clear the probability threshold, and at low temperature that threshold is high. Below it, there’s silence.
The invisible threshold that cuts you out
Modern RAG systems — the ones used by Perplexity, Bing Chat, and increasingly ChatGPT too — add a further filtering mechanism. They don’t just generate from training data: they retrieve documents from the web and then decide which to keep and which to discard.
Chang et al., in their work on multi-agent filtering, describe exactly how it works:
“Documents to the right of the threshold are retained, while those to the left are filtered out.”
A clean threshold. On one side, those who pass; on the other, those who get cut. There’s no gentle gray zone where “almost” is enough. Either the document that talks about you clears the threshold, or it doesn’t make it into the answer. Period.
From this follows an important deduction — and I want to be clear that it’s a deduction, not a direct experimental result: to move a competitor out of one of those 3-5 slots, it’s not enough to “exist” online. You have to beat their probability. You have to build a stronger signal than theirs on those specific queries, so that the model — when it computes the probability distribution — assigns your brand a higher value.
If the competitor has 20 mentions on earned media and you have 3 scattered across weak sources, you won’t beat them by adding another 5 scattered mentions.
Which signals the game is played on
If the challenge is to beat a competitor in probability, the question becomes: probability based on what? The answer comes from a finding that overturns many people’s intuitions.
In 2025, Kaiwen Chen et al. analyzed which types of sources AI engines favor in their answers:
“AI Search exhibit a systematic and overwhelming bias towards Earned media — third-party, authoritative sources.”
Earned media. Not your site. Not your social channels. Third-party, authoritative sources that talk about you without you having written them. Articles in trade publications, citations in qualified directories, mentions in reports, interviews. This is the signal that weighs most when the model decides who to include in the answer.
Which means competitor displacement isn’t played out on your website. It’s played out elsewhere. It’s played out on the same third-party sources where your competitor built their presence — and where you have to build a stronger, more recent, more consistent presence.
Take the 10 most important queries for your business and run them on ChatGPT, Gemini, and Perplexity.
Displacement isn’t theory: it’s reverse engineering
This is where strategy becomes concrete. Competitor displacement works the opposite of what most people think. You don’t start from yourself. You start from who’s already in.
The first step is mapping. Take the 10 most important queries for your business and run them on ChatGPT, Gemini, and Perplexity. For each answer, note who gets cited. After 10 queries you’ll have a clear map: 3-4 brands that always show up, and yours probably isn’t among them.
The second step is understanding why those brands are there. And here you need to look at the sources, not the answers. On Perplexity it’s easy: the citations are visible. On ChatGPT and Gemini you have to reconstruct them. Search each of the cited brands and analyze where they appear: do they have a Wikipedia page? Are they cited on trade media? Do they have detailed listings on vertical directories? Does their founder have a profile with recognized authority?
The third step is finding the gaps. A competitor is rarely strong on every front. Maybe they have excellent media coverage but an inconsistent brand entity across platforms. Maybe they’re well positioned nationally but weak on local geographic authority. Maybe their brand-category association is generic while you can build a hyper-specific one.
The competitor’s gaps are your entry levers. And this isn’t a theoretical exercise — it’s the work I do with clients before any AI visibility strategy. Without this mapping, every action is a shot in the dark.
Concentrate the fire, don’t scatter it
A mistake I often see: trying to build authority “on everything” at once. As many mentions as possible, as many directories as possible, as much content as possible. But displacement doesn’t work that way.
If the competitor has 20 mentions on earned media and you have 3 scattered across weak sources, you won’t beat them by adding another 5 scattered mentions. You beat them by concentrating the effort on a specific front where you can build a clear advantage. A narrower niche, a more precise geography, a more defined specialization.
Go back to the probability mechanism: the model calculates probability conditioned on the context of the query. “Best tax advisor for startups in Milan” is a different query from “best tax advisor.” On the generic query, the established competitor is almost impossible to dislodge. On the specific query, the game is open — because the probability is recalculated on a context where your specialization weighs more.
Concentrating the signal on a precise niche, with targeted earned media, is the most efficient way to clear the threshold on those specific queries. Once you’re in, you expand from there. It’s the same principle by which a brand unknown nationally can dominate AI answers in its city or its vertical: probability conditioned on the specific query is a different playing field from absolute probability.
The check: do you know who you’re playing against?
Displacement starts with awareness. If you don’t know who occupies those 3-5 slots today, you’re building blind.
Run the 10-query test I described above. For each competitor that shows up, create a profile with: sources where they’re cited, type of mentions (earned vs. owned), naming consistency, geographic coverage, category association. Then make the same profile for yourself.
The comparison will tell you two things: how wide the gap is and where the gap is narrowest. The points where the gap is narrow are the ones where effort produces faster results. It’s not an analysis you can wrap up in half an hour — it requires cross-referencing data you normally don’t look at. But it gives you direction.
You’re not trying to be “better in general.” You’re looking for the precise point where you can clear a specific threshold. When you find that point, all the effort goes there — targeted earned media, naming consistency, brand-category association repeated on those sources — until the model generates you in place of someone else. Displacement isn’t a single event. It’s the cumulative result of signals that, query after query, shift the probability in your favor.
