Content: how fast does engagement fade?

Content decay is the newest member of this family of disciplines, and — measured against the older fields' validation standards — the least rigorously proven in public.

The core idea

A post's engagement doesn't arrive evenly over its lifetime — it spikes shortly after publication and then fades, usually following something close to an exponential or power-law decay curve. Researchers have studied this "half-life" — the time it takes a piece of content to accumulate half of the total engagement it will ever receive — across several platforms and metrics, and the results vary dramatically depending on exactly what is being measured.

What's been published

Link and tweet half-life studies

A 2011 bit.ly study of link clicks found an overall mean half-life of about 3 hours, broken down by referrer: Twitter 2.8 hours, Facebook 3.2 hours, direct/email/IM 3.4 hours — with individual links ranging as widely as 5 minutes to 11 hours depending on the type of content. Half-life here is defined operationally as the time to accumulate half of a link's total eventual clicks.

A more recent and methodologically explicit study, Pfeffer, Matter & Sargsyan's "The Half-Life of a Tweet" (2023), used Twitter's impression_count metric and found that peak impressions-per-second occurs about 72 seconds after posting, with a median tweet half-life of roughly 80 minutes — and that about 95% of tweets show negligible further impressions after 24 hours. This number is much shorter than the older bit.ly figure, and the discrepancy itself is instructive: half-life estimates vary enormously depending on the metric definition (impressions vs. clicks) and the platform/API era being measured.

Popularity prediction — the Szabo & Huberman baseline

The classic and still-cited baseline for predicting content popularity is Szabo & Huberman (2010), which fits a log-linear regression between early and late cumulative popularity:

log(1 + v_tt) = log(1 + X_tr) · θ

fit by minimizing the squared L2 norm on log-transformed view/vote counts. On YouTube, fitting a slope-1 line between log(views at day 30) and log(views at around day 10) produced a Pearson correlation of R = 0.77 — meaning views need to be tracked for roughly 10 days to predict 30-day popularity at that correlation. On Digg, measuring just the first 2 hours of activity predicted 30-day popularity at R = 0.90 — Digg's dynamics stabilize much faster than YouTube's. A related study found log-transformed daily view counts correlate at r = 0.84 between day-2 and day-90 views, and r = 0.92 between day-7 and day-30 views — again showing that prediction accuracy improves the longer the early observation window is extended.

Modern machine-learning benchmarks

Later ML-era work on YouTube view-count prediction reports a mean absolute percentage error (MAPE) as low as 0.2773 using models that add a long-term-decay parameter to early-stage prediction. General findings across this literature: log-transformed view counts roughly follow a power-law-with-exponential-decay shape, and ensemble/ML models (random forests, boosted GAMs, neural networks) generally outperform simple log-linear regression, evaluated with MAE, RMSE, MAPE, and R². One Instagram virality-prediction study reports a notably high MAPE of 0.15 (roughly 85% "accuracy") and R² = 0.924 — worth flagging as an outlier relative to the field's more typical 0.7–0.9 correlation range.

Predicting timing, not just magnitude

One cascade-timing study — the closest published academic analog to "predicting a decay/death timestamp with a stated tolerance" — reports being able to predict the time for a social cascade to reach a given viral target size with prediction error bounded within 30%, using only 20% of the cascade's data. This result structurally resembles the α-λ percent-accuracy convention used in predictive maintenance (see the maintenance page) more than it resembles a survival C-index — but it is a single published study, not a field-wide benchmark.

The gap this field hasn't closed yet: no paper found in this survey reports a calibrated hours-to-death metric analogous to reliability engineering's confidence bounds, survival analysis's C-index plus calibration slope, or predictive maintenance's α-λ accuracy cone. Almost all published accuracy claims in this space are about predicting eventual magnitude (total views at a fixed future day) from an early observation window — a fundamentally different and easier task than predicting when engagement crosses a death or decay threshold. When reading a claim about "content prediction accuracy," it's worth checking which of these two tasks is actually being scored — the literature frequently conflates them by describing a magnitude-prediction result as "predicting popularity."

Why content decay is the newest member of the family

Reliability engineering, survival analysis, and actuarial science are each roughly a century old or more; predictive maintenance and epidemic Rt-estimation are decades old. Rigorous, publicly reported content-decay research is barely 15 years old, and — as the summary above shows — has not yet converged on a shared validation standard the way the older fields have. That gap, and how it might close, is explored further on the methodology page.