What Makes a Good Word Search
6 July 2026
When we launched Word Search in April, the puzzles it produced were perfectly decent. Since then we have rebuilt the part of the game that lays out the grid, and the puzzles you get today are noticeably better. The reason why turns out to be a nice little lesson in how a small change of rule can change the character of a whole thing.
The obvious way to build a word search is to take each word in turn, pick a random empty stretch of grid that fits it, drop it in, and once every word is placed, fill the leftover squares with random letters. That works. It also produces a scatter: each word sits by itself in its own patch of the grid, surrounded by filler. And a word sitting by itself is easy to spot, because real words and random letters look different at a glance. You end up scanning for the tidy island in a sea of noise rather than actually reading the grid.
So we changed the rule. Now, before placing a word, the game looks at every letter already on the board and works out every possible way this new word could be threaded through one of them, sharing that square. It strongly prefers one of those crossings, and it gives extra weight to crossings that land on the first or last letter of a word already placed, because building outward from the tips of existing words spreads the structure across the grid instead of knotting it in the middle. Only when there is genuinely no shared letter to work with, as with the very first word, does it fall back to dropping the word somewhere at random.
The effect is that the puzzle becomes woven rather than sprinkled. A single square can belong to two different answers at once, so finding a letter no longer tells you which word you have found. There are fewer dead regions of pure filler, the words are distributed evenly instead of clustering wherever there happened to be room, and the whole grid reads as one object. It is the difference between a page of stickers and a crossword.
The second change was quantity. We had a suspicion that our bigger grids felt emptier than the small ones, so rather than argue about it we measured. We wrote a benchmark that generates three hundred boards for each configuration and reports the numbers: how many of the requested words actually got placed, their average length, what fraction of the squares hold a real word letter rather than filler, how many squares are shared between two words, and how evenly the eight directions get used. The small board turned out to be the one we had tuned by instinct and got right, at roughly four squares in ten carrying real letters. So we set the word counts for the larger sizes to hold that same density: ten words on a small grid, twenty on a medium, twenty-eight on a large, thirty-four on a huge one, with a deliberate ceiling so that the biggest board stays a satisfying afternoon rather than a chore.
None of this is the sort of thing anyone will notice and comment on. Nobody finishes a word search and thinks about placement strategy. But the value we claimed for this game when we launched it, the visual attention it exercises and the vocabulary it puts in front of a young reader, is only real if the puzzle is genuinely worth solving. A grid where the answers stand out like islands trains a child to spot islands. A grid that is properly woven makes her actually read it. We could have left the easy version in place and nobody would have written to complain, which is exactly the kind of thing we would rather not do.