How Fantera Recommends Anime and What a Match Score Means

Type Frieren into the Finder and the first card is Delicious in Dungeon at 82 percent, with Wandering Witch: The Journey of Elaina one point behind. Fantera blends two things AniList publishes about every show, the tags its users attach and the recommendations they vote on, into a score for every pair of titles in a catalogue of about 5,000 series and films. It keeps each title’s best 50 picks, all worked out during a scheduled sync. A search reads that stored list, and your seen marks and filters only decide which cards are hidden.

What goes into a match

The first ingredient is overlap. Every title becomes a list of weighted features: each genre counts as 1, and each tag counts for its AniList relevance rank divided by 100, so a tag rated 94 percent relevant weighs 0.94. Spoiler tags are dropped. Cosine similarity then measures how closely two lists point the same way, from 1 for shows built from the same strong tags to 0 for shows with nothing in common.

The second ingredient is other viewers. AniList users attach recommendations to a show and vote on them, and the sync reads each title’s 25 highest-rated ones. Vote totals run from single figures into the thousands, so Fantera takes the logarithm of each total and divides by the logarithm of the anchor’s best-voted recommendation. The gap between 800 and 1,000 votes then counts for far less than the gap between zero and 50. A title nobody recommended from your anchor scores zero here and has to win on tags alone.

The final score is 60 percent overlap plus 40 percent community signal, and Frieren’s list shows the two pulling against each other. Violet Evergarden has more recommendation votes than either pick above it, 1,181 against Delicious in Dungeon’s 761 and Wandering Witch’s 503, and still comes third, which can only mean its tags overlap less. Wandering Witch outranks it on overlap despite carrying the lowest AniList rating of the three: 7.4, against 8.5 for the other two.

Sequels, spin-offs and the sync

Overlap alone would tell every Frieren viewer to watch Frieren Season 2. So before ranking, Fantera links the titles AniList connects as sequel, prequel, parent story, side story, spin-off, alternative version, summary or compilation, and treats each connected group as one franchise. Nothing from your anchor’s franchise is ever recommended, and every other franchise gets a single slot, taken by whichever of its entries scores highest. The three seasons of Mob Psycho 100, 12, 13 and 12 episodes, count as one recommendation.

That slot can go to a later season: Frieren’s ninth pick is Mushoku Tensei: Jobless Reincarnation Season 2 Part 2, because it outscored the rest of its franchise, so check where a story starts. Grouping only sees links inside the catalogue of TV series and films, so two seasons joined only through an OVA can both appear.

Reading the match percentage

The raw score sits between 0 and 1. The percentage on the card is that score stretched and fenced in by one line of code: Math.min(99, Math.max(40, Math.round(45 + 55 * score))). In words, multiply the score by 55, add 45, round to a whole number, then hold the result between 40 and 99. Delicious in Dungeon’s raw score of 0.6644 becomes 45 plus 36.5, which rounds to 82.

The mapping has a floor. The clamp guarantees that no result ever shows below 40 percent, and because only picks with a positive raw score are stored, the lowest figure you can actually meet is 45. A 50 percent match sounds like a coin toss; on this scale it means the two shows have very little in common. It also tops out at 99, which a raw score reaches at about 0.97, so no pick is ever presented as identical to your anchor.

Read the percentage as an order rather than a measurement: picks at 82 and 80 are effectively level. Percentages under different anchors cannot be compared either, because the community half is scaled to each anchor’s own best-voted recommendation. A match is also no forecast of enjoyment; the star rating beside it, AniList’s community score out of 10, is where opinions about quality live.

Where the why line comes from

Under Wandering Witch, the card reads: “Elaina’s episodic wandering gives you that same gentle, philosophical travelogue feeling, just with a livelier young witch at the center.” A language model wrote that sentence, Anthropic’s Claude Sonnet 5, working offline during the sync. For the top 24 picks of every title, it was given each show’s title, year, genres, strongest tags, the opening of its synopsis and the tags the pair shares, and asked for one sentence per pick. Sentences over 40 words, or that use the word “spoiler”, are thrown away. The survivors are stored, and the site serves them as text; it never calls a model when you search.

The model worked from synopses and tag lists, never from the episodes. Its lines are summaries of catalogue data, so they can overstate a resemblance or miss a change in tone. Violet Evergarden’s line calls its heroine a kuudere because that tag is shared, which is fair, and it says nothing about the letters she writes for strangers that make up most of the show.

When no written sentence exists, usually for picks below the top 24, the card falls back to a template built from the shared tags, such as “Shares Travel, Philosophy and Medieval with Frieren: Beyond Journey’s End.” If a pick has no shared tags either, the card shows a line Fantera’s editors wrote for that title, or nothing at all.

What the Finder cannot see

Filters narrow the stored list and never re-rank it. “Short” means 13 episodes or fewer in that single catalogue entry, which says nothing about the seasons that follow; the guide to seasons and episode counts covers that gap. Seen marks live in your browser, and the most recent 500 travel to the server in a cookie so it can hide those titles before paging. The “If you like this” strip on each anime page shows the top four from the same stored list.

The deeper limit is that tags describe ingredients. Two shows tagged Travel and Philosophy can be made with very different skill, and the system cannot know whether you loved Frieren for its dry jokes or its grief. A cult favourite outside AniList’s most popular 5,000 will never surface either. For picks argued one at a time, read anime like Frieren; for the number on a MyAnimeList page, see the MyAnimeList score guide.