Anime Semantle

Anime Semantle's similarity is not computed from genre tags. It comes from running each title's synopsis and structure through a sentence embedding model. Which means it disagrees with the tags more often than you'd expect.

And that disagreement is the most useful signal in the game. If you keep guessing the same genre and the score won't move, or a completely unrelated title suddenly returns 0.8, that isn't a bug — it means the model is looking at something else.

Here are those pairs, pulled from the actual data.

1. Zero shared tags, similarity above 0.76

Pairs that share not a single genre tag yet score 0.76 or higher. Same-franchise entries and sequels are excluded throughout.

Why they cluster

The Fable ↔ Great Pretender (0.813) — by tag they look opposite: "comedy" versus "action." But both are stories about con artists running an elaborate scheme. The characters' occupations and the narrative structure are nearly identical. The model reads that structure, not the genre column.

Classroom of the Elite ↔ Assassination Classroom (0.785) — one is filed as psychological suspense, the other as action comedy. What they share is "a classroom governed by an unusual set of rules." Their synopses open with almost the same vocabulary, so the embeddings land close together.

Spirited Away ↔ Maid Dragon (0.770) — a Ghibli classic and a slice-of-life comedy. Yet both are stories about a human living alongside beings from another world. Completely different in tone, same skeleton.

2. Three shared tags, similarity below 0.35

The reverse case: pairs sharing three or more genre tags that still land under 0.35.

Fullmetal Alchemist: Brotherhood ↔ Shangri-La Frontier (0.260) is the starkest. They share action, adventure and fantasy, and 0.26 still means effectively unrelated. One is an epic about two brothers atoning for a mistake amid a national conspiracy; the other is about a player speedrunning deliberately terrible games. Tags record what appears in a show. Similarity reads how the story is told.

3. How to use this in the game

When you get 0.7+ from a completely different genre — don't switch genres. What the model caught is narrative structure, not genre. Recall the plot of the title you just entered and look for other shows built on the same skeleton: "con artists," "a strange classroom," "living with something otherworldly."

When the same genre never breaks 0.4 — the genre may already be right. As the table above shows, three shared tags can still return 0.26. Drop the tags and switch to tone: serious or light, sprawling epic or short emotional piece.

When you stall at 0.5–0.65 — the broad family is right and the sub-family is wrong. The isekai & fantasy guide and the romance guide break that band down genre by genre.

4. A correction — "short synopsis" was not the reason

This article originally said that titles with short synopses score low because the model has little to read. It sounds plausible, but it was a guess written without measuring, and measuring it proved it wrong.

The briefly described group is in fact slightly higher, and the correlation between length and isolation is +0.033 — effectively zero. The Chinese animation (donghua) we singled out actually sits above the overall median: "Douluo Dalu 2" at 0.763, "Cang Yuan Tu 2" at 0.812, "Battle Through the Heavens Special" at 0.773.

The reason is not hard to see. Similarity does not read the synopsis alone: structural text (title, series name, tags) carries 70% and the synopsis 30%. A thin synopsis still leaves the series name and tags holding the title in place.

A separate limit is real, though — synopses are truncated to 500 characters at collection time, and 87% of the 800 titles hit that cap, so the tail of any long synopsis never enters the calculation at all. The scoring method and the evidence behind the 70:30 split are on the about page.

Further reading

Every number here comes from the data the game actually runs on. Similar Anime Recommendations applies the same data to "what should I watch next," and the main strategy guide has the similarity band reference table and a full walkthrough. Each day's answer and its nearest neighbors accumulate in the past answers archive.

The same data supports a few other angles. Which Genres Are Hardest measures isolation genre by genre, When to Guess a Character explains why a character usually beats a title, and Ratings Are Not a Clue covers why narrowing candidates by score fails.