Anime Semantle

One common approach is to start with highly rated shows on the theory that "today's answer is probably a classic." The game displays ratings, so it is a natural instinct.

It does not work. There are two reasons, and both show up in the data.

1. The pool barely varies in rating

The answer pool is the top 800 titles by MAL global popularity. Which means they are all highly rated to begin with.

The entire pool fits inside a 1.28-point band, and 80% of it is squeezed into the 0.65-point range from 8.02 to 8.67.

For "start with the highly rated ones" to work, the filter has to eliminate candidates. When nearly everything is an 8-point-something, nothing gets eliminated. Titles above 9 are a handful, and the only 7s sit right at the 7.97 floor.

2. Similarity and rating are unrelated

So what about narrowing to "titles rated similarly to the one that just scored 0.7"? That fails too.

Across 599,852 pairs — same-franchise pairs excluded, both sides having a rating — we computed the correlation between similarity and the gap in the two ratings.

Breaking it down by similarity band gives the same answer. In every band the average rating gap is flat at roughly 0.27.

A pair at 0.85 differs in rating by exactly as much as a pair at 0.25. Which makes sense: similarity is computed from synopsis text and ratings come from viewer votes. The two are measuring completely different things.

3. Very close, very split on score

Pairs that tell nearly the same story yet sit almost a full point apart.

"Majo no Takkyuubin" and "Sousou no Frieren" sit at a respectable 0.750 — both are stories of a magic-using girl travelling and meeting people. Yet their ratings, 8.24 and 9.25, are close to the widest gap this pool allows.

So "it is near Frieren, therefore it must be a 9-rated title" does not follow. There are plenty of 8.2s among Frieren's 0.75 neighbors.

4. What to look at instead

• Plot structure — this is what similarity actually reads. Find titles whose skeleton matches: who does what, to whom, and how.
• Tone, not genre — even genre tags disagree more than you would expect. There are real pairs sharing no tag at all that still hit 0.81. They are collected in When Similarity Betrays the Tags.
• Characters — once a candidate title comes to mind, typing one of its characters is the better move 98% of the time. Numbers in When to Guess a Character.
• Genre difficulty — if the answer is a sports title, its neighbors are far away no matter what you do. See Which Genres Are Hardest.

5. The one time ratings are useful

There is one. After the answer is revealed.

Ratings are useless for narrowing candidates but perfectly good for deciding what to actually watch. Find similar shows by similarity, then pick among them by rating — in that order it works. That method is written up in Similar Anime Recommendations.

6. Method

Ratings are the MAL scores from the source data, unmodified. The correlation is a Pearson sample correlation over anime pairs with the same franchise excluded and a rating present on both sides — 599,852 of them, counted directionally. The 0.9-and-above band held only 2 pairs and was dropped from the table. The scoring method and the evidence behind the 70:30 weighting are in About the Game.

Further reading

Every number here comes from the data the game actually runs on. The main strategy guide has the similarity band reference table and a full walkthrough, and each day's answer and its nearest neighbors accumulate in the past answers archive.