Anime Semantle

Some days the answer falls in three or four guesses. Other days you feed in twenty titles and never break 0.5. This article is about the possibility that it is not luck — that it is the genre of the answer.

We opened the same similarity data the game runs on and measured all 800 anime, genre by genre.

1. What we measured — isolation

For every title we took the similarity of its nearest neighbor, excluding its own franchise. Sequels, seasons and movie versions are all removed. Without that exclusion the nearest neighbor of "Diamond no Ace: Act II Second Season" is just "Diamond no Ace: Act II," which tells you nothing.

The lower this number, the more the title is an island in the pool. No matter how close a guess you think you are making, the score refuses to climb. A high number means the neighborhood is crowded, so hitting the right cluster carries you the rest of the way.

2. Results by genre

Isolation scores grouped by genre, sorted by median. A title carrying several genre tags counts toward each of them. Genres with fewer than 5 titles are omitted.

"Below 0.7" is the share of titles in that genre whose nearest neighbor still fails to reach 0.7. The higher it is, the more often you get that "I guessed something similar and nothing moved" feeling.

3. Sports is hard by a wide margin

Sports sits alone at the bottom. Its median of 0.679 is 0.03 below second-place Boys Love (0.708), and more tellingly 63% of its 62 titles have a nearest neighbor under 0.7 — seven times the rate of the easiest genre, Supernatural (9%).

The reason is not hard to guess. Different sports do not share plots. A baseball series and a basketball series have the same skeleton — students form a team and chase a national tournament — but almost none of the same words in their synopses. The words are what the model reads.

At the other end, Supernatural (0.781) and Romance (0.774) are easiest because their narrative patterns repeat. "An ordinary student is pulled into something occult" or "two people circle each other without confessing" describes dozens of titles apiece, so landing the cluster drags a whole chain of neighbors with it.

4. The most isolated titles in the pool

Looking at individual titles rather than genres, sorted by how far away the nearest non-franchise neighbor is.

"Lycoris Recoil" is the most isolated title in this pool. At 0.362 there is effectively nothing like it. On a day when it is the answer, no other title you guess will comfortably clear 0.4.

That three of these eight are sports titles is the same story the table above tells.

5. How to use this in the game

• If you hit a ceiling in the 0.5s, suspect the genre. When several different clusters all stall at 0.4–0.5, the answer may be one of the isolated titles above. Rather than hunting for something "even more similar," switch the sport or the subject matter outright.
• With sports, you have to name the sport. Knowing "it is a sports anime" barely helps when baseball, basketball, volleyball and figure skating all sit around 0.5 from each other.
• If supernatural or romance lands, push. Both clusters are dense, so once you see a 0.7 you can usually walk the neighborhood to the answer.
• When titles fail, switch to characters. Even isolated titles usually sit above 0.9 from their own cast. The numbers are in When to Guess a Character.

6. Limits of this measurement

Isolation is unrelated to how much synopsis text exists. It seemed like it should be, so we measured: the 44 titles under 200 characters have a median isolation of 0.742, slightly above the 695 titles at the 500-char cap (0.736). The correlation with length is +0.033.

The real constraint is elsewhere. Synopses are truncated to 500 characters at collection time and 87% of the 800 titles hit that cap, so the tail of a long synopsis never enters the calculation. Isolation is also measured only within the 800-title answer pool — a close relative outside the pool does not show up in these numbers.

Cases where the genre tags themselves disagree with similarity are collected separately in When Similarity Betrays the Tags. The scoring method and the evidence behind the 70:30 weighting are in About the Game.

Further reading

The same measurement runs on finer axes too. A Sharper Axis Than Genre uses the 34 theme tags and spreads from 0.643 to 0.853, while Guessing Shounen Because It Feels Shounen checks whether the shounen / seinen / shoujo labels form real clusters.

Every number here comes from the data the game actually runs on. The main strategy guide has the similarity band reference table, and Isekai & Fantasy and Romance cover splitting a cluster into finer subgroups. Each day's answer and its nearest neighbors accumulate in the past answers archive.