A content library goes stale for a particular audience once that audience has seen its items recently enough to recognise them on sight, usually long before the library runs out. The argument here is that perceived freshness depends on exposure history more than on library size, that research on forgetting and spacing explains why, and that rotating content is therefore a cheaper and stronger retention lever than adding more of it. The sections below show how to state a rotation policy formally, how to measure perceived novelty instead of item count, and where sustaining interest turns into exploiting it.
Content staleness as an exposure-history problem
The usual response to a stale library is to make it bigger. That treats freshness as a stock problem: items get used up, so you author more. It fails in a specific way. Cost grows linearly with the number of items authored while perceived freshness levels off, because in any one session a user meets only a small sample of the library.
Stated more precisely, a library holds N items. For a group g at time t, the exposure history is a pair per item: Δt_i, the time since the group last encountered item i, and c_i, the number of times it has encountered it. Perceived novelty depends on that pair, not on the item itself. Define retrieval strength r(Δt_i, c_i) ∈ [0, 1] as the probability that the group recognises the item as previously seen, and let the effective fresh library be the sum of what the group does not yet recognise:
N_eff(t) = Σ_i [ 1 − r(Δt_i, c_i) ]
r(Δt, c) = exp( −Δt / s(c) ), s(c) increasing in c
The second line is only a convenient assumption. It says recognition decays with elapsed time, at a rate set by a stability parameter s that grows with prior exposures. The useful conclusion does not depend on the exact parameters. N is authored once and stays fixed. N_eff changes continuously with usage, and each session samples from N_eff, so N_eff is the number that predicts whether the next session feels new.
A worked example shows the saturation. Assume a library of 1,080 items, a session that draws 40, and uniform sampling across sessions. After 50 sessions the group has drawn roughly 2,000 times, and the expected share of the library it has met at least once is 1 − (1 − 1/1,080)^2000, about 84 per cent. We picked those assumptions for illustration, but the point holds under reasonable substitutions: a heavy group runs out of unrecognised items long before the library runs out of items. Changing which items are eligible costs far less than changing how many exist.
The forgetting curve and the spacing effect in content scheduling
The decay term comes from one of the oldest quantitative results in experimental psychology. Ebbinghaus (1885) showed that retention of learned material falls off sharply and then flattens, and that the curve is regular enough to model. Read the other way round, a forgetting curve is a novelty curve. The process that erodes deliberate learning also erodes the incidental memory of a content item, and an item last seen long enough ago behaves much like one never seen.
Research on spacing turns this into something a scheduler can use. Work on distributed practice (Cepeda et al.) shows that the interval between exposures, and not only their number, determines what a repetition does. Bjork's work on desirable difficulties adds that effortful retrieval produces more durable learning than easy retrieval, and that conditions which feel harder at the time often work better. We cite both for the general idea and take no specific interval from either.
In a product, this becomes a cooldown window. If recognition decays, an item does not have to be retired to become interesting again. It only has to be ineligible for long enough. That runs against the usual live-operations instinct. Retiring an item throws away the work that went into it for good, while a cooldown parks it and brings it back with most of its novelty restored. The weekly rotation in GUZZL is built on this idea. Every Monday the active set changes across a library of more than 1,080 handcrafted cards. Cards leave the pool and come back, and in any given week a group meets only a fraction of what has been written. How much distinct play a finite deck can actually produce is a separate question, covered in replay entropy for finite card sets.
One asymmetry follows from the model. Because s grows with exposure count, items a group has seen many times need a longer cooldown before they feel new again. A single uniform cooldown therefore rests rare items too long and favourites too briefly, and it is the favourites that players notice.
Novelty, arousal, and the inverted-U relationship
More novelty is not always better. Berlyne's work on conflict, arousal and curiosity found that interest rises with novelty only up to a point. Too little variation bores people, too much confuses or repels them, and the useful range lies in between. Csikszentmihalyi's account of flow has the same shape. Engagement depends on challenge matching capability, and boredom and anxiety are two ways of missing that match rather than opposite ends of one scale.
Shannon (1948) supplies the quantity that changes along this axis. If a group implicitly assigns probability p_i to meeting item i next, the surprise carried by that encounter is −log p_i, and the expected surprise of a session is the entropy of the distribution the group is effectively sampling from. Entropy is useful here because a rotation policy controls it directly. A static library pushes the distribution toward the recognised set and expected surprise toward zero. Replacing everything each week maximises entropy, but it also wipes out the expectations that surprise is measured against. A rule the group has never met cannot break an expectation, and when nothing in a library is familiar there is no baseline for anything to feel new against.
Rotating part of the pool keeps entropy in the useful middle. A stable, familiar core sets up the expectations and a rotating minority breaks them. In GUZZL the core is the structure (14 distinct rule types, boss rounds every 10 cards, 15 lives, duel and spotlight mechanics), and only the card content inside that structure rotates. Localisation complicates this. The six supported languages are culturally rewritten instead of translated, so a card's surprise does not carry over from one language to another, and each rotation set has to be checked per language. Our piece on localisation and cultural rewriting explains why that check cannot be skipped.
Zeigarnik's work on finished and unfinished tasks points to another lever that needs no new content. People recall interrupted material more readily than completed material. By extension, a set that rotates out before the group has used it up leaves some anticipation behind, and a fully used set leaves none. Rotating out an unfinished pool is cheaper than writing a new one, and it may also work better.
Reinforcement schedules and the ethical boundary
Skinner's work on schedules of reinforcement showed that variable-ratio delivery, where reinforcement comes after an unpredictable number of responses, sustains behaviour longer than fixed-ratio delivery and holds up better against extinction. Applied to novelty instead of reward, this suggests that freshness arriving unpredictably holds engagement better than freshness on a timetable. If a group knows that new content arrives on a fixed cadence in a fixed amount, it writes off the sessions in between before they happen.
This needs a clear line, because the same finding sits behind the least defensible patterns in the industry. Variable-ratio schedules are the mechanism behind slot machines and loot boxes, and their pull comes from separating effort from outcome in a way that people do not tire of. What matters is what the schedule is applied to. If it governs which content a group meets, it varies the experience people came for, and a group that stops playing loses nothing. Using it to decide whether a promised thing arrives is a different matter. With a paid draw, a progression unlock or a competitive advantage, the uncertainty becomes a cost the user carries, and charging for that uncertainty sets the product against its own audience. So rotation can be as unpredictable as you like in which content appears, but access to anything a user was promised has to stay fully predictable.
Rotation policy design
A rotation policy comes down to four decisions, and under this model they do more to keep a group playing for dozens of sessions than writing more content does. Each has a reasonable default.
Cohort partitioning. The eligible pool should depend on the group rather than the calendar, because each group has its own exposure history. A group at session 5 and a group at session 200 have very different N_eff and should not be served the same pool. Partitioning by cumulative exposure lets one authored library serve both. Tenure is a poor proxy for exposure and should not be used in its place.
Cooldown windows. After a group sees an item, the item becomes ineligible for that group for a period, and the period grows with the number of times that group has seen it. This is the spacing effect put into practice, and it lets a finite library behave like a larger one.
Cadence. A weekly boundary is a compromise between two constraints. It is short enough to give a lapsed group a reason to come back while the habit is still there, and long enough to write, localise and check the rotating set. We picked the Monday cadence in GUZZL for the second reason as much as the first. Six culturally rewritten languages cannot be turned around any faster without the quality dropping.
Seeded randomness. When a group plays in one room, the shared-attention constraints described in participation architecture in party games apply, and per-user randomness becomes a fairness problem. Two players drawing from different pools cannot share a reference, and one of them will think the other was favoured. If the pool is derived from a seed shared by the group and the week, nobody can predict the selection in advance, everyone present gets the same one, and support can reproduce it afterwards.
Metrics for perceived novelty in a rotating content library
The metric most teams report is library size, and it is the one least connected to the outcome. The model above calls for measurements taken per group and per session.
| Metric | Definition | What it detects |
|---|---|---|
| Novel share | Fraction of a session's items the group has never encountered | Onboarding freshness, but it saturates early and misleads if used alone |
| Effective library | N_eff at session start, from the recognition model | Whether the pool is exhausted for this group specifically |
| Repeat interval | Distribution of Δt at the moment of re-exposure | Cooldown windows that are too short, visible as a left tail |
| Expected surprise | Entropy of the pool the group is sampling from | Both failure directions: collapse toward zero, or unbounded growth |
| Recognition rate | Share of re-exposures the group treats as familiar | Direct estimate of r, and the only way to calibrate s(c) |
Two cautions apply to all five. First, novelty is a means. What finally counts is whether sessions continue and groups come back, and novelty metrics are only worth tracking if they move before those do. Second, every quantity here is inferred from behaviour rather than observed directly, so recognition needs checking against something external. A group skipping an item quickly tells you little about recognition, since it could mean several other things.
Limitations
Nothing here shows that rotation causes retention. We present no controlled comparison, and we have none from our own work. The GUZZL figures above describe the product's design. They are not experimental results, and we claim no retention outcome for them.
The recognition model is a functional form chosen because it is easy to work with. The forgetting and spacing research behind it studied deliberate learning of verbal material in the lab, and applying it to incidental recognition of entertainment content is an analogy that research does not itself support. The worked example assumes uniform sampling and independent sessions, and neither holds in practice: favourites come up more often, and what a group draws depends on what it drew before. The entropy measure describes the distribution a policy generates, which is not the same as the distribution a group holds in its head, and we have not measured the gap. Last, all of this concerns perceived freshness. It says nothing about whether the content is good, and if the content is weak, rotation will not help.
References
- Berlyne, D. E. (1960). Conflict, Arousal, and Curiosity.
- Bjork, R. A. (1994). Memory and Metamemory Considerations in the Training of Human Beings.
- Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed Practice in Verbal Recall Tasks.
- Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience.
- Ebbinghaus, H. (1885). Memory: A Contribution to Experimental Psychology.
- Shannon, C. E. (1948). A Mathematical Theory of Communication.
- Skinner, B. F. (1938). The Behavior of Organisms.
- Zeigarnik, B. (1927). On Finished and Unfinished Tasks.