The AI-native games glossary
34 terms from publishing deals, Steam launches and LLM character engineering, each defined in a sentence or two and linked to the page that goes deeper.
A
Adaptive director
A system that observes how a player or party plays and adjusts pacing, difficulty or content in response. AI-native directors reason about play style, not just numbers.
AI-native game
A game whose core experience depends on a model running at play time. Remove the model and it stops being the same game. Contrast with AI-assisted development.
ARPDAU
Average revenue per daily active user. Multiplied by expected active days per install, it gives a simple LTV.
Authored fallback
A hand-written line or behaviour used when generated output fails a check, the model is slow or the player is offline. Keeps failures in-world.
B
Boxleiter method
Estimating a Steam game's unit sales by multiplying its review count by a factor, commonly 30–80.
C
Cache hit rate
The share of model calls whose prompt prefix is served from cache. Caching saves money when it exceeds (w−1)/(w−r) for write multiplier w and read multiplier r.
Context window
The maximum number of tokens a model reads in one call. Everything a character knows for this turn must fit in it.
Cost per player-hour
Inference spend for one hour of play: calls per hour × tokens per call × price per token. The core planning number for AI-native games.
D
D1 / D7 / D30 retention
The share of an install cohort active 1, 7 and 30 days after install. Often modelled as a power law r(n) = a·nᵇ.
Distillation
Training a smaller model on a larger model's outputs so it imitates it in a narrow domain, such as one character's voice, at lower cost.
E
Erlang C
A queueing formula giving the probability a request must wait when c servers handle Poisson arrivals. Used to size inference capacity for a latency target.
Eval suite
A fixed set of scenarios whose generated outputs are scored against rubrics (in character, on lore, within rating) on every prompt or model change.
G
Generative world
Levels, quests or regions created at runtime inside hand-designed constraints that guarantee fairness and solvability.
Guardrails
The layered checks that keep generated content within rules: input screening, character constraints, output classification and fallbacks.
I
Inference
Running a trained model to produce output. In games, the per-player runtime cost of AI characters and systems.
IP reversion
A contract clause returning publishing rights to the studio if the publisher stops supporting the game or sales fall below a floor.
L
Latency budget
The allocation of acceptable delay across network, queueing, prefill, first token and voice so total response time stays conversational.
Lifetime multiplier
Lifetime units sold divided by first-week units. Varies strongly by genre; the key assumption in wishlist forecasts.
Live-generated content
Content created by a model while the game runs, as opposed to pre-generated content made during development. Stores ask for guardrail descriptions for it.
LTV
Lifetime value: expected revenue per player or install. For AI games, compare it with lifetime inference cost.
M
Model routing
Sending each request to a small or large model based on a cheap classifier. Pays off when the routed share × the price gap exceeds the classifier's cost.
N
Net receipts
Platform payouts after store fees, refunds and taxes. The base most publishing splits are calculated on.
P
Prefill
The phase where a model processes the input prompt before generating. Long prompts lengthen it; caching shortens it.
Prompt caching
Reusing a provider's processed form of an unchanged prompt prefix. Cheaper and faster on hits, slightly more expensive on writes.
R
Rating envelope
A written spec of the most intense content a game can generate per rating category, enforced in code. The basis for rating games with live content.
Recoup
The publisher earning back its recoupable investment from the game's revenue before the long-term split applies.
Retrieval score
A ranking of stored memories for inclusion in a prompt, commonly a weighted sum of recency, importance and relevance.
Revenue-based financing
An advance repaid as a fixed multiple from a share of future revenue. Annualised cost ≈ M^(1/t) − 1.
S
Square-root staffing
The queueing result that capacity for a fixed service level grows as a + β√a, so larger pools need proportionally less headroom.
Summarise-and-flush
Compressing the last k exchanges into a summary and dropping them from context. Optimal k* = √(2(s+q+ρs)/m).
T
Tone guard
An output check that keeps a character's lines in voice and within the rating envelope before the player sees them.
V
Vertical slice
A short, near-final-quality section of a game that demonstrates the full experience. The usual stage for strong publishing terms.
W
Wishlist conversion
First-week sales divided by launch wishlists on Steam. GameDiscoverCo reported a median near 0.105 for late-2024 releases with 5k+ wishlists.
Got a game that talks back?
Send a build, a deck or two minutes of raw footage. A producer plays every pitch and answers within 10 business days.