Why I counted
Connections players talk about the colors as if they were fixed personalities: yellow is the freebie, purple is the trick. I wanted to know how true that is, and whether it has stayed true. This site keeps a day-by-day archive of every published puzzle, so I took every archived puzzle from June 12, 2023 (the first in our archive) through September 21, 2026 that had a complete, usable answer key: 1,191 puzzles and 4,764 groups, after dropping 7 days whose saved answers were incomplete (details in the methodology). I sorted each group’s category into one of twelve types based on how its name is worded. Eight of those types are wordplay: the link between the four words is in their letters, their sound, or a second meaning, rather than in what they are. The other four are semantic: the words simply belong to a set. The methodology section explains the rules and what they get wrong.
Purple is a different game: 70% wordplay
The split is not subtle. A purple group’s category is wordplay 70% of the time (833 of 1,191 groups). Yellow’s is wordplay 1.3% of the time, green’s 3.6%, blue’s 8.5%. Blue stands out on a different type: it has the highest share of named-things categories (brands, characters, athletes, titles) at 13%, against 1.4% for yellow. That blue figure is raw classifier output; I hand-checked the yellow and purple named-things matches but not the blue ones (see Limits).
Share of groups whose category is wordplay
What purple categories are made of
Purple bars are wordplay types; gray bars are semantic types. Types under 0.5% (Hidden words, 1 group) are omitted from the chart but included in every total.
Fill-in-the-blank is the single biggest purple pattern at 36%: categories like ___ MAN SUPERHEROES, ___FISH THAT AREN’T FISH, MR. ___. Next come the starts-and-ends categories (12%, e.g. STARTS OF U.S. COINS, ENDING IN COLORS), letter changes (8.9%, e.g. BAND NAMES MINUS NUMBERS), homophones (5.6%) and one-word-many-meanings categories (4.2%, e.g. WHAT “GREEN” MIGHT MEAN). Anagrams, the first thing I personally test when I am stuck, are only 0.6% of purple groups (7 in 1,191).
For comparison, yellow is 92.5% plain set (members of a category). Its most common wordplay types are a tie between starts or ends of longer words and homophones, at 5 groups each out of 1,191 (0.4%).
Purple has gotten more purple every year
This is the finding I did not expect. In 2023, 59% of purple groups were wordplay (119 of 203); in 2026 so far it is 82% (214 of 260). The share has risen every calendar year in between. The semantic purples that used to be common (a plain set that just happened to be obscure, like TRIG FUNCTIONS or GEMSTONE CUTS) have gone from 22% of purple groups to 10%.
Purple groups whose category is wordplay, by year
The mix inside wordplay has shifted too. Fill-in-the-blank categories have declined from 40% of purple groups in 2023 to 29% in 2026, while the starts-and-ends family (STARTS OF EUROPEAN CAPITALS, ENDING IN HORSE GAITS) has grown from 2% to 26%, and add-or-drop-a-letter categories from 4% to 12%.
| Type | 2023203 groups | 2024365 groups | 2025363 groups | 2026260 groups |
|---|---|---|---|---|
| Fill in the blank / complete the phrase | 40% (81) | 39% (143) | 35% (127) | 29% (76) |
| Starts or ends of longer words | 2% (4) | 8% (28) | 13% (46) | 26% (68) |
| Add, drop or change a letter | 4% (8) | 7% (24) | 12% (42) | 12% (32) |
| Homophones | 5% (10) | 6% (21) | 6% (20) | 6% (16) |
| One word, many meanings | 2% (4) | 4% (13) | 5% (18) | 6% (15) |
| Spelling or sound pattern | 5% (11) | 1% (4) | 3% (10) | 2% (4) |
| Plain set (members of a category) | 22% (44) | 21% (77) | 13% (49) | 10% (26) |
| Named things (people, titles, brands, places) | 9% (19) | 5% (19) | 6% (22) | 2% (6) |
Two cautions. 2023 covers only June to December (203 puzzles) and 2026 is incomplete (260 puzzles), so the endpoints are noisier than the middle years. And the classifier reads category names, so part of the rise could be the editors naming the trick more explicitly (“STARTS OF…”) rather than using it more often. The year-over-year direction is consistent enough that I think the trend is real, but that caveat is why I am not putting a decimal on it.
The repeat words: BALL has been on 29 boards
Some words keep coming back because they can go anywhere. BALL has appeared in 29 different puzzles out of 1,191, and over those appearances it has been in a yellow group, a green group, a blue group and a purple group at least once each. Of the 40 most-repeated words (each seen in at least 15 puzzles), 35 have landed in every color at least once. Most of them are spread fairly evenly; a few lean one way, and the table shows which.
| Word | Puzzles | Yellow | Green | Blue | Purple | Split |
|---|---|---|---|---|---|---|
| BALL | 29 | 11 | 5 | 9 | 4 | |
| RING | 24 | 11 | 6 | 6 | 1 | |
| BAR | 22 | 3 | 3 | 7 | 9 | |
| FLY | 22 | 6 | 9 | 2 | 5 | |
| CHECK | 21 | 2 | 4 | 11 | 4 | |
| LEAD | 20 | 9 | 8 | 2 | 1 | |
| BLUE | 19 | 5 | 0 | 7 | 7 | |
| JACK | 19 | 4 | 5 | 6 | 4 | |
| WING | 19 | 6 | 6 | 5 | 2 | |
| CAN | 18 | 3 | 6 | 5 | 4 | |
| HEART | 18 | 5 | 4 | 4 | 5 | |
| HOOK | 18 | 2 | 6 | 8 | 2 | |
| SPLIT | 18 | 6 | 6 | 4 | 2 | |
| KING | 17 | 1 | 4 | 6 | 6 | |
| LOVE | 17 | 5 | 1 | 5 | 6 |
The leaners: FISH was purple 9 of 16 times; EGG was purple 7 of 15 times. DIRECT was yellow 8 of 16 times; STICK was yellow 8 of 17 times; RING was yellow 11 of 24 times; LEAD was yellow 9 of 20 times. Those are tendencies, not rules: each of these words has still appeared in at least 3 different colors.
The words themselves do not give the color away
I checked whether the colors differ in the words they use. They do not, on the two things I can measure. The average answer is 5.64 to 5.71 letters long in every color (yellow 5.66, green 5.64, blue 5.71, purple 5.69), and multi-word answers such as TOP HAT make up 5.0% to 6.3% of answers regardless of color. The difference between the shortest and longest color averages is 0.07 letters, under a tenth of a letter.
The category names are another story. Yellow categories average 2.29 words (FOOTWEAR, CONDIMENTS); purple categories average 3.37 words, because a trick takes more words to describe than a set does. You only see the name after you solve the group, so this is confirmation rather than a hint, but it fits the rest of the data: the color labels track the kind of category, not the length of the words on the tiles.
Day of the week: not significant
I expected weekends to lean harder toward wordplay. The data does not show it. The share of puzzles whose purple category is wordplay ranges from 62% (Thursday, 106 of 170) to 78% (Friday, 133 of 170) across the seven days, with 169 to 172 puzzles per weekday. A chi-square test of independence on those counts gives χ² = 11.4 on 6 degrees of freedom, p ≈ 0.08: the difference does not reach significance at the 5% level. That is a statement about purple wordplay share only; I did not test anything else by weekday.
How to use this when you play
- Run the wordplay checks in order of frequency. When four leftover words have no shared meaning, ask first whether they all complete the same phrase (36% of purples), then whether they are the start or end of a longer word (12%), then whether adding or removing one letter makes a set (8.9%). Anagrams and hidden words together are 0.7% of purples; check them last.
- Say the words out loud. Homophones and one-word-many-meanings categories are 9.8% of purples combined, and neither is visible on the page.
- Treat the repeat words as wild cards. If BALL, RING, BAR, FLY, CHECK or another word from the table above is on the board, hold off on assigning it until the other three members of a group are certain. It keeps being chosen because it fits more than one group.
- Expect the trick to be the trick. Since 2023, an obscure-but-plain set has become a rarer purple every year. Across the whole dataset, 36% of purple groups (427 of 1,191) complete a phrase, and another 30% (356) have a link in the letters or sounds of the words themselves (anagrams, homophones, hidden words, letter changes, spelling and sound patterns, starts and ends). If the four words you cannot place are ordinary words, those two families cover 66% of purples between them.
- Trivia leans blue, by the classifier’s count. The named-things rules matched 150 blue groups (13%) against 66 purple (6%). Those blue matches are unreviewed classifier output, so treat this as a hint about the data, not a rule for your board.
The general strategy, including what to submit first and how to protect your four mistakes, is in our Connections tips guide.
Methodology
Source
Our own archive of the daily puzzle, saved each day from the public puzzle and stored as one file per date. You can browse the same data on the Connections archive. The dataset runs from June 12, 2023 through September 21, 2026: 1,191 puzzles, 4,764 groups, 19,056 answers. The most recent day is always excluded until it becomes an archive page, so the dataset ends one day before the run date.
Exclusions
8 archive files were excluded: 1 because it was the most recent day and not yet an archive page, and 7 because the saved answers contained the placeholder “UNDEFINED” in place of real words (a capture error on our side, not a puzzle feature). Any file with a missing solution, a malformed grid or a color set other than the four standard colors would also have been skipped; none were.
Classification
Each category name is matched, in a fixed order, against a list of wording rules and assigned the first type that fits. Wordplay types: fill-in-the-blank (a blank, “WORDS BEFORE…”, “FIRST WORDS OF…”), hidden word (letters of one word hidden inside another, as in “BODY PARTS SURROUNDED BY TWO LETTERS”), anagram, homophone (“SOUNDS LIKE”), letter change (“MINUS”, “PLUS”, “WITH A LETTER CHANGED”, “IN THE SINGULAR”), spelling or sound pattern (palindromes, silent letters, rhymes), starts or ends (“STARTS OF”, “ENDING IN”, “PREFIX”), and one word, many meanings (“WHAT ‘X’ MIGHT MEAN”, heteronyms). Semantic types: synonyms and slang (“SYNONYMS FOR”, “…, WITH ‘UP’”, a quoted exclamation), things-that (“THINGS THAT ARE RED”, “ASSOCIATED WITH…”), named things (“BRANDS”, “CHARACTERS”, “MOVIES”, “MAGAZINES”, sports leagues, “…, WITH ‘THE’”), and plain set for everything else. The full rule list is in the open script scripts/connections-study.mjs in the site’s code, which also writes the raw counts behind every percentage on this page.
Limits
- The classifier reads only the category name, and it errs in both directions. A wordplay group with a coy name (“E” THINGS, meaning E-mail and E-scooter) is counted as a plain set; a set whose name happens to contain a trigger word can be counted as wordplay or as named things. I do not know which direction dominates, so the shares above should be read as estimates with an unmeasured error, not as floors or ceilings.
- Synonym groups with a one-word name (PLUNDER, WEARY) cannot be told apart from one-word sets (INSECTS, FONTS) without knowing the words, so both land in plain set. The synonym type only counts categories that say so.
- The named-things rules are keyword matches and misfire (AREAS BETWEEN MOUNTAINS AND HILLS is not trivia; it matched on “mountains”). I read the yellow and purple matches by hand: a minority of the yellow matches and a smaller minority of the purple ones are false positives, and I have not turned that reading into a number I would stand behind. The green and blue named-things figures were not hand-reviewed at all and are unreviewed classifier output.
- The starts-or-ends type includes categories like SPICES BEGINNING WITH “C” (a yellow group), where a single letter is the whole constraint. I kept them as wordplay because the link is in the spelling. The type is still overwhelmingly purple: 146 of its 182 groups are purple, 19 blue, 12 green and 5 yellow, and those few yellow and green cases are part of why neither color is at 0% wordplay.
- No difficulty ratings, solve times or player data are involved. “Purple” here means the group the editors labeled purple, nothing more, and every finding on this page is about how that label relates to the category type, not about how hard anyone found the puzzle.
Corrections welcome through the contact page. The dataset will be re-run periodically; the numbers on this page cover puzzles through September 21, 2026.