How AIs decide which brands to cite when someone asks 'what's the best company for X'

A generative AI doesn't consult a table of rankings and pick the top three brands. It generates, token by token, the most plausible continuation for that question — and that plausibility is higher for brands that show up clearly and repeatedly in the sources the model knows or checks. This mechanism, which we'll call apparent consensus here, works in a structurally different way from the PageRank that shaped two decades of SEO.

From vote-counting to apparent consensus

PageRank, the basis of Google's original algorithm, treated each link as a vote: the more (and better) sites pointed to yours, the more authority you accumulated. It was a graph logic — countable, relatively transparent.

The mechanism behind an AI-generated answer is different. There are no votes being counted — there's a probability calculation over which word is most likely to come next, given everything that came before in the answer and given the material the model has available. When that material mentions your brand consistently, in multiple places, clearly tied to the problem in the question, the probability of it showing up in the answer goes up. Not because it "won a vote," but because it is, statistically, the continuation most coherent with what exists.

That's why we prefer the term apparent consensus: the model doesn't verify the truth of a claim the way a fact-checker would. It detects what appears to be widely corroborated within the available material — and treats that as safer to assert. A brand mentioned consistently, by different sources, in different contexts, looks more "established" to that calculation than a brand mentioned once, even if that single mention is technically more detailed.

Two layers of evidence, weighted differently

There are two sources of information available for the AI to answer with, and they don't carry equal weight for every kind of question:

• Trained knowledge — what the model learned during training, up to a cutoff date. It's static: it doesn't update between one version of the model and the next.
• Real-time corroboration — when the AI has access to web search (ChatGPT, Gemini, and Perplexity all operate this way by default today), it checks current pages before answering, and uses that result to confirm, update, or supplement what it already "knew."

For time-sensitive questions — "does this company still exist," "what's the current price," "does it still serve this area" — the model tends to lean more on the real-time corroboration layer. That's actually the best news: it means content published now can change an answer today, without waiting for a new training cycle of the model.

One piece of text that confirms something, and one that confirms nothing (illustrative)

Take two sentences that could sit on the "About" page of competing companies in the same industry:

Company X: "We're a leader in management solutions, with a team dedicated to delivering excellence and innovation for our clients."

Company Y: "We serve dental clinics with up to 15 chairs, with monthly fees between $290 and $690 depending on the module, WhatsApp support during business hours, and integration with the most widely used insurance systems."

Company X's sentence lets a model confirm nothing specific. Company Y's sentence answers nearly every element of a concrete question like "which management system for a small dental clinic costs less than $500 a month." Between the two, when generating an answer, the model effectively has more usable material from Company Y — not out of editorial preference, but because only Y offered something concrete to work with.

Why different categories have different dynamics

This mechanism doesn't work the same in every market. In heavily contested, generic categories — "which is the best bank" — the set of brands already widely corroborated by years of press coverage is so large that the barrier to entry for a new brand is high, almost regardless of content effort.

In narrower, more specific categories — "which management system for a small dental clinic in a rural town" — the volume of corroboration available about any competitor is much smaller. That means two things at once: it's easier to become the most-corroborated option in these categories, because the comparison bar is lower; and it's riskier to be left out, because there are few slots in the answer and each competitor that takes one makes a proportionally bigger difference.

The limits of this mental model

An honest caveat: "apparent consensus" is a mental model built from how these systems behave observably — not a literal description of the internal weights of any specific model, which are largely proprietary and not public. It's meant to guide practical decisions, not to predict with technical precision the exact behavior of a specific version of ChatGPT or Gemini in an isolated case.