A census of 19,776 roleplay cards:
the ecosystem behind 950 million chats

Four headlines from this census

1. Traffic is a pyramid: the top 1% of cards (197 of them) absorb 40.3% of all chats, the top 10% absorb 73.4%. Yet the middle is real — the median card still logs 9,873 chats. This is not a winner-take-all wasteland.
2. 83.0% of characters are male; female characters are 5.5%. It's an audience-of-women universe: the play is “pursue him”, and the top tags — pure-love, jie, position-locked — are community pacts around exactly that.
3. Every tag where demand outruns supply is a structurally complex card: multi-character (lift 4.22), infinite-loop worlds (3.31), simulators (2.37), system cards (2.22), ensembles (1.90). The most-written mood tags — redemption, secret crush, love at first sight, single-character — are all oversupplied.
4. A hit card is an engineering artifact: median greeting 1,232 chars of prose, 90% wrapped in HTML, median beautify code 39,189 chars; 73% embed a BGM player, 52% ship collapsible panels, 49% a memory zone. A card stopped being a persona paragraph — it's a miniature front-end project.

The shape: a pyramid with a living middle

First, the question everyone asks: is this a winner-take-all market? It's a pyramid — and the middle is alive. The top 197 cards (1%) absorb 40.3% of all chats, yet the median card still logs 9,873. Not a market of one star and a crowd of extras.

The chat pyramid: who lives on each floor

Five floors by lifetime chats per card. Left wing = that floor's share of all cards; right wing = its share of all chats. The 102 million-chat cards (0.5%) absorb 32.6%; the 10k–100k middle class takes 27.9% with 42.4% of the cards.

Concentration on another scale: the top 10 cards absorb 12.8%, top 100 take 32.4%, the top 1% (197 cards) 40.3%, and the top 10% take 73.4%.

The creator side is even more top-heavy

4,596 creators, 4.3 cards each on average. 39.6% published exactly one card and never returned; at the other end, the most prolific account carries 123.

Chats concentrate toward top creators even harder than toward top cards: 1% of creators absorb 46.2% of all chats. An author brand is worth more than a hit card.

Who lives here: 83% male characters

83.0% of cards are male characters; female characters are 5.5%. Keep that number in hand and everything else clicks — the pact tags (pure-love, jie, position-locked) and the pursuit-style play are the grammar of a women-audience community.

Gender-code semantics are inferred from the creation form's enum order plus tag co-occurrence (see method); the all-gender and multi-character rows house most ensemble cards.

Character age: a peak at 20, a wall at 18

986 of the top 1,000 declare an age: mode 20, median 21, p90 at 32. Age 18 is the second peak (151 cards); seventeen-and-under is exactly zero — the adulthood line is a hard border here. The small step at 26–28 matches the steady demand for older love interests.

Names doubling as shelf labels, covers built for portrait phones

Median name length is 3 Chinese characters — it reads like a person. But 9.8% of names carry a parenthetical edition note, and 7.3% use ensemble notation like “4+1” or “n+1”. The name field is quietly doing shelf-label duty.

75.2% of covers are ultra-vertical art (aspect < 0.6), plus 21.6% ordinary portrait; square-or-wide is 3.2% combined. The portrait phone is this universe's only picture frame.

The tag universe: 9,995 words, a few dozen in circulation

Of 9,995 unique tags, 6,879 were used exactly once — signatures, private memes, one-off declarations. The real shared vocabulary sits in the top few dozen. 67.2% of cards fill all six tag slots (5.49 per card on average); tags are the search surface, and nobody wastes it.

Top 20 tags by card count, colored by semantic group. Audience-pact words (all-audience, pure-love, jie) own the podium — this universe negotiates terms before genre.

“Jie”: a community pact that grew outside the schema

“Jie” (洁, the exclusivity pact: the character's body and heart have only ever belonged to the player) is declared by 4,692 cards; “not-jie” by 543. That's 8.6 to 1. It isn't a content rating — the platform has no such field, so the community grew one in the tag slots, at a density where staying silent reads as a statement.

A side note: small-town China entered the card universe

Small-town (298 cards), Cantonese (254), Hong-Kong-circle (230), esports (321), band (220) — the aesthetic currents of the recent Chinese internet play out here in real time. Cards are an instant slice of pop culture, with a far shorter reaction arc than long-form fiction.

• 县城 small-town China298
• 粤语 Cantonese254
• 港圈 Hong Kong circle230
• 电竞 esports321
• 乐队 band220

Supply vs demand: the gap hides behind the engineering bar

Which lanes are overwritten, which undersupplied? For 59 tags (≥250 cards each) we computed a crowding index, lift: the tag's share of all chats ÷ its share of all cards. Above 1, demand over-rewards the tag; below 1, supply is grinding itself down.

Bar length is log-scaled (lift 4.22 and 0.33 sit roughly symmetric). ◆ = entries cross-checked under a second recipe (top-1,000 occurrence rate ÷ overall rate); every check agrees in direction.

How to read this board

The amplifiers share one trait: multi-character (4.22), infinite-loop worlds (3.31), simulators (2.37), system cards (2.22), ensembles (1.90) — structurally complex cards that ship a play system, not just a person. The red oceans are the mirror image: redemption, secret crush, love at first sight, single-character — mood-heavy tags with a low writing bar and therefore thick supply.

The demand gap sits behind the engineering bar. That is the most practical sentence on this board: to eat an amplifier's traffic, first accept its production cost.

Low lift doesn't mean “don't write it”. The redemption lane holds 1,804 cards and still has million-chat hits — it's a crowding signal, not a quality verdict.

Red packets are creator-side promotion spend: 13.7% of cards carry one, and those cards absorb 22.1% of chats with a median of 15,018 — 1.5× the universe median. Direction can't be settled from a snapshot: packets may buy heat, or confident authors may be the ones who pay. Recorded as is, no causal claim.

The build of a hit card: half writing, half front-end

Dissect the top 1,000 and “writing a card” turns out to be half prose, half engineering. Median greeting: 1,232 characters of pure text (even p25 is 758), 90% wrapped in HTML. Median beautify code: 39,189 characters; the heaviest single card ships 368,517. Only 8 of 1,000 go bare.

Greeting length in pure text (HTML stripped)

Manual culture: creator notes read like product docs

Across all 19,776 cards, 33.1% of creator notes ship a $-command list, 31.7% discuss model choice, 25.6% carry do-not clauses; median length 191 characters. Players have been trained to read the manual before chatting — your creator notes will be read as product documentation.

The average card: a dossier of a card that doesn't exist

Stitch every median and mode together and you get this “average card”. No such card exists — but every line of the dossier is a real majority.

Six takeaways creators can copy straight into decisions

1. The red-ocean trio: enter with careSingle-character (lift 0.33), secret crush (0.42) and redemption (0.54) are the three most oversupplied lanes. Not unwritable — just the most crowded, thinnest average returns, sharpest differentiation required.Back to the board →
2. Structure is the biggest open slotMulti-character, simulators, system cards and infinite-loop worlds all run 2×+ over supply. The bar is engineering, not prose — which is exactly why the gap persists. A first step: bolt a progressing status panel onto a single-character card.Card studio →
3. Don't open under 800 charactersThe top-1,000 p25 is 758; the median 1,232. The convention is to perform what's happening right now — 48% open with a timestamp, 36% with a location. Preview your first screen before shipping.Greeting preview →
4. Memory design isn't decoration49% of hit cards build in a memory panel — the community's self-rescue from long-chat amnesia. Treat “what should be remembered” as a field you design, not something the model improvises.Card checkup →
5. Tag slots are free search surface67.2% of cards fill all six. Copy that — and keep at least one slot for a mechanic word (pursuit / simulator / sandbox). The board shows mechanic words carry traffic; pure mood words don't.Back to the tags →
6. Write creator notes as product docsA third of creators list commands; three in ten give model advice. Readers expect a manual: recommended model, boundary declarations, command usage — that's the passing grade here.Our card-writing guide →