When you play Wordle, you choose words based on some abstract concept of maximizing information. A computer's approach is not dissimilar; at first, the full range of valid input words is possible. After each guess, this list of words is pruned, based on the feedback received. Since eventually, we want to get to only one possible word, an intuitive goal for a greedy algorithm is to maximally reduce the size of the resulting possible word list. However, because we want to find the optimal guess before we know the resulting square colors, it is not as simple as minimizing a single list. For each word, there exists a number of square color permutations (238 including the all-correct case and excluding the 5 impossible "GGGGY" cases to be exact), each with their resulting list of remaining words.

​

This gives motivation to find some way to quantify the performance of a word based on how well it cuts down the lists of resulting words. This is where dealing with the probability space of a particular guess becomes useful. For example, at the start of the game there are 12966 valid words. It turns out that if we choose aeros as our first guess, and all of the resulting squares are grey, this cuts down the list of possible answers to 577 words, or in other words, this event has a probability of 577/12966 (all grey squares still provides information about which letters are not in the word). Alternatively, if the first two squares were green, that actually leaves only one possible answer: aecia. This event has a probability of 1/12966. It's no use to simply sum up all of the probabilities for a particular word, because that would simply equal 1 with the way this probability space is partitioned. We instead want to use expected value.