How a Wordle Solver Works: Our Algorithm, Step by Step

The short answer

Short answer: a Wordle solver is a filter with a ranking on top. It keeps a list of every word that could still be the answer, deletes whatever your green, yellow and grey tiles rule out, and orders the survivors so the most useful next guess sits at the top. None of that needs to be a black box, so this page walks through the exact pipeline our free Wordle Solver runs in your browser, including the duplicate-letter rules where most home-made solvers quietly go wrong.

Step 1: start from the right word list

Every solve begins with a candidate list, and which list matters more than people expect. Our solver first loads a curated list of 1,352 likely answers, the pool the game draws from. A far larger list of 15,783 valid five-letter guesses sits behind an extended toggle and loads on demand, for the rare puzzle whose answer falls outside the curated set. Filtering the smaller pool keeps the results realistic, because obscure dictionary words are legal guesses but poor predictions. Both lists are static files fetched once, and everything that follows happens locally on your device.

Step 2: green tiles pin letters to positions

A green tile is the strongest clue in the game: this exact letter, in this exact position. The solver stores each green as a position-to-letter pair and keeps only the candidates whose letter at that position matches. One character comparison per green per word, no subtlety at all, which is exactly why greens are checked first.

Step 3: yellow tiles are two clues in one

A yellow tile carries two facts at once. The letter is somewhere in the answer, and it is definitely not where you just played it. The solver records both: a surviving word must contain the letter, and it must not have that letter in any square where it came back yellow. Human players routinely use the first fact and forget the second, then burn a guess repeating a yellow letter in the same spot. The filter cannot forget.

Step 4: grey tiles remove letters, usually

A grey tile normally means the letter appears nowhere in the answer, so the solver bans it outright and discards every candidate containing it. Greys are quietly the workhorse clue, since one grey on a common letter can eliminate more words than a green does. But there is an exception the code has to respect. If the same letter is also green or yellow elsewhere in the row, the grey copy does not ban the letter at all. It tells you the answer holds fewer copies than you played, which is the business of the next step.

Step 5: duplicate letters and the two-pass rule

Duplicates are where casual solvers break, so ours reproduces the game’s behaviour rule for rule. To rank guesses the solver must simulate feedback, and its feedback function scores a guess against a possible answer in two passes. It first counts how many copies of each letter the answer contains. Pass one walks the guess and awards greens, and every green immediately claims one copy of its letter from that count. Pass two walks the guess again for the yellows: a letter earns yellow only while unclaimed copies remain, and each yellow claims one more. Once a letter’s count runs dry, any further copies come back grey. The order matters, because a green takes priority even when the yellow copy appears earlier in the word.

The filtering side mirrors the same arithmetic. Coloured copies of a letter within a row set a minimum count for that letter, and a grey copy alongside them upgrades it to an exact count. So the SPEED row above does not just say the answer contains an E somewhere. It says the answer contains exactly one E, and the solver drops every remaining word with two. With 31.7% of answers repeating a letter, that distinction does real work. Our duplicate-letter research digs into how often each letter doubles up.

Step 6: rank what survives

Filtering leaves a list, not advice, so the solver ranks the survivors in two ways. The full results grid is ordered by a simple letter-frequency score: words built from common letters rank higher, with a bonus for using more distinct letters. That heuristic is instant at any scale. The headline suggestions are stronger. Whenever 400 or fewer candidates remain, the solver plays every surviving word against every other survivor, simulates the tile pattern each pairing would produce, and groups the outcomes into buckets. Each candidate guess is then scored by the Shannon entropy of its buckets. High entropy means the guess splits the field into many small groups, so whichever colours actually come back, few candidates survive. The three highest scorers become the suggestions, with the frequency score breaking ties. Above the threshold the entropy pass is skipped, partly because it is quadratic and partly because early on almost any sensible guess is informative. The design trade-offs are unpacked in our solver ranking research and the entropy explainer.

One deliberate choice worth knowing: suggestions are drawn from the surviving candidates themselves, never from outside probe words. Every suggestion could actually be the answer, which keeps the advice legal in hard mode and means you are never asked to burn a guess on a word that cannot win.

A worked example: CRANE, then BOARD

Watch the pipeline run on a real game. The hidden answer is GUARD and the opener is CRANE. Across every possible game, our precomputed opener data rates CRANE at 5.64 bits of expected information, leaving about 47 of the 1,352 answers standing on average. Every tile row below is generated by the exact feedback function described in step 5.

The filter reads that row as four rules. Words containing C, N or E are deleted, because those letters are grey and never coloured elsewhere. Only words with A in the middle position survive the green. And every survivor must contain an R, though not in second position, since that is where the yellow sat. What remains is a small slice of the answer list, comfortably inside the suggester’s threshold, so the entropy ranking is already live. Suppose you follow with BOARD, one of the surviving candidates.

Two more letters join the banned list, B and O. The middle A is confirmed again, which changes nothing but costs nothing. The interesting tiles are the new greens: the yellow R from guess one has found its home in fourth position, satisfying the earlier not-in-second rule, and D is locked in last place. Survivors must now fit the shape “_ _ A R D” while avoiding five banned letters. In the curated answer list that leaves a mere handful of words, GUARD and AWARDamong them. With so few candidates the suggester’s job is trivial: any surviving guess either wins immediately or cuts the field to a single word, so the game ends within two more rows. Here it takes one.