Two ways to measure rarity
Curated lists. Krillion, by its own account, builds accepted-answer lists from dictionaries and reference data, cleans them by hand, and attaches a familiarity rating that draws on word-frequency data plus editorial judgment. It then adds a wrinkle: an answer can be the “standard obscure choice” that everyone who wants to look clever reaches for, and that gets its own low-paying tier, Too Clever. The strength of curation is taste; the cost is that every prompt is hand-built.
Corpus frequency. Wordillion measures how often each word appears in a very large sample of written English (books, news, subtitles, the web) and scores on that alone. The strength is consistency across thousands of prompts without a human rating each one; the cost is that the number sometimes disagrees with intuition, which we'll get to.
The Zipf scale
Word frequencies span many orders of magnitude, so they are usually expressed on a log scale. The Zipf value of a word is log10 of its occurrences per billion words. A word with Zipf 6 appears about a million times per billion (once every thousand words); Zipf 3 is once per million words; Zipf 2 is once per ten million. Some anchors from the wordfreq dataset:
| Word | Zipf | Roughly |
|---|---|---|
| ABOUT | 6.4 | once per 400 words |
| HORSE | 5.1 | once per 8,000 |
| ROBIN | 4.3 | once per 50,000 |
| HERON | 3.5 | once per 300,000 |
| GREBE | 2.6 | once per 2.5 million |
| GAUR | 1.9 | once per 13 million |
Turning a number into a tier
Wordillion cuts the scale at fixed points. Common is Zipf 4.6 and above (10 points), Uncommon from 4.0 (20), Rare from 3.4 (40), Epic from 2.8 (55), Legendary from 2.2 (70), Mythic below 2.2 (90). On top of that, whichever valid answer to the prompt has the lowest frequency of all is One in a Wordillion and pays 120, so there is always exactly one best answer to hunt for. The cutoffs are 0.6 apart, which means each tier is about four times rarer than the one above it.
Voting-based rarity (“only 2% of players said this”) rewards whoever plays late in the day with knowledge of what others said, and it collapses when few people play. A frequency scale gives the same answer at 6 am and 11 pm and for prompt number ten thousand.
Where frequency disagrees with intuition
- Proper nouns run low. ODIN or MONET appear in text far less often than their fame suggests, so a frequency scale calls them rarer than players feel they are. Wordillion compensates by requiring every trivia category to contain a floor of well-known answers and by measuring trivia prompts a notch more leniently.
- Homographs run high. SOLE the fish borrows frequency from SOLE the adjective; DRAKE the duck borrows from the musician. These words are undervalued as answers.
- Short words are inflated. Three-letter strings appear as abbreviations and typos. Wordillion's scoring accounts for length; its answer lists also exclude -S plurals, as Wordle does.
- Too clever. Krillion's insight is that a word can be rare in text and still be the obvious “rare” answer (every quiz player says AXOLOTL). A frequency scale cannot see this. Wordillion's answer is structural rather than editorial: the single rarest fit always pays most, so the obvious clever pick is only the best pick if it really is the rarest.
What this means for how you play
Learn which kinds of word score. Scientific and specialist vocabulary (GENUS names, textile terms, nautical words) sits at Zipf 2 to 3 and pays Epic or better. Everyday synonyms sit at 4 to 5 and pay little regardless of how clever they feel. Names of famous people are rarer than you'd guess and often pay well in five letters. The rare short words by category lists are generated straight from these numbers, and how to find rare answers fast turns them into a method.
Questions people ask
The base-10 logarithm of how many times a word appears per billion words of text. Zipf 6 is very common (about once per thousand words), Zipf 2 is very rare (once per ten million).
According to the game and its community guides, from curated answer lists combined with familiarity data and editorial judgment, with a separate low-paying “Too Clever” tier for the predictable obscure answer. It is not live player voting.
By the word's frequency in a large English corpus (the wordfreq dataset), cut into six tiers. The rarest valid answer to each prompt is One in a Wordillion and pays 120.