What a probability on a forecast actually means
A number attached to a prediction reads as precision. Most of the time it is a statement about how much evidence exists, not about how certain anyone is, and the difference decides whether you should act on it.
It is a rate, not a promise
A forecast published at 70% is not a strong opinion. It is a claim that calls made on this much evidence land about seven times in ten, and it is only true if the record bears that out over many calls. A single 70% forecast that does not happen has told you nothing about the method. Thirty of them that all fail has told you everything.
This is why AI Visibility Forecasting publishes the number rather than a word. "Likely" cannot be graded. Seventy percent can, and it can be shown to be wrong.
The number and the label are answering different questions
Every forecast carries both a probability and a confidence label, and readers routinely collapse them. They are not the same axis.
The probability is how often this call is expected to land. The label is what kind of evidence sits behind it: a confirmed change that already happened, an observable trend that is measurable and underway, a reasonable forecast supported by a scheduled event or a strong analog, or an early signal that may not develop at all.
A confirmed change at 90% and an early signal at 90% are not the same object. The first is near-arithmetic. The second is a weak indicator whose history happens to be good, and it will move a great deal more as evidence arrives. Reading only the number loses that entirely, which is why the label is never dropped from a published call.
What the number cannot carry
It does not describe size. A 90% chance of a change that moves nothing is not worth acting on, and a 40% chance of something that reshapes discovery for a category is. Probability and impact are published separately so they cannot be quietly multiplied into one figure that hides both.
It does not describe timing beyond its window. A probability is always attached to a stated horizon. The same call over a longer window is a different call with a different number, and lifting a percentage out of its window is the most common way a forecast gets misquoted.
It is not a measurement. It is an estimate with a method behind it, made where a measurement is not available. Where something can be measured, it is measured and reported as a fact rather than dressed as a forecast.
The number that should make you suspicious
A forecast at 95% or above is usually either a confirmed change wearing a forecast's clothes, or a method that has learned to publish only what it cannot lose on. Both are worth questioning, and the second is worse, because a record full of near-certainties looks excellent and has told nobody anything they could not already see.
The useful forecasts live between roughly 40% and 80%. That is the range where a call is genuinely uncertain, genuinely actionable, and genuinely capable of being graded wrong.
What we hold, and what we do not
Every published call keeps its probability, its window, its impact and its confidence label together, and none of the four is dropped when a forecast is summarised or syndicated. They are graded against what happened, and the record is public, because a probability with no track record behind it is a decoration.
What we do not do is publish a probability we cannot justify from evidence, adjust one after the fact, or attach a number to a guess to make it look measured. Where the evidence does not support a rate, the call is not published as a forecast at all.
This page covers how to read the number. Why a forecast misses covers what happens when it does not land, accuracy holds the record, and both sit under AI Visibility Forecasting.