Token Probability
The likelihood a model assigns to each candidate next token; low probabilities flag where the model is uncertain about your brand or topic.
When an artificial intelligence writes text, it is not just pulling words out of thin air. Behind every single word is a calculation called token probability. This is the likelihood the model assigns to each possible next word as it generates a sentence.
Every time the AI outputs a word, it is making a choice from a massive vocabulary, and the math behind that choice reveals exactly how confident or uncertain the model is at that exact moment. If the probabilities are low and spread out among many different options, the model is essentially hesitating. These moments of high uncertainty can highlight weak spots in how the AI understands or describes something, like your brand.
By analyzing these probabilities, we can measure word rarity and uncertainty. This is the raw signal that powers advanced tools and helps us reshape how the model behaves using controls like temperature. Ultimately, tracking these probabilities turns the seemingly mysterious inner workings of AI into a measurable, actionable target.
Token probability is the likelihood a model assigns to each possible next token as it generates text. Every word it outputs is a pick from a probability distribution over its whole vocabulary, and those probabilities reveal how confident — or uncertain — the model is at each step.
This is the raw signal behind our Tree Walker: word rarity comes from how common a token is, and word uncertainty comes from how flat the probability distribution was when the model chose. Where probabilities are low and spread out, the model hesitated — a weak spot in how it describes your brand.
These probabilities are also what sampling controls like temperature and top-p sampling reshape, and they trace back to the tokenization scores baked in during training. Reading them turns AI perception into a measurable target.