Question: Within a channel over time, when it posts more than its own baseline, do its videos earn more, fewer, or about the same views, and does its total reach rise or stay flat?
"Post more" is advice you hear constantly, and the common counter-worry is that extra uploads just slice the same audience across more videos ("spreading thin"). This study takes each channel as given and asks how its views move with its own upload cadence, holding the channel itself fixed.
The cohort is English-language business and creator-economy YouTube channels with at least 20 qualifying videos (public + organic + long-form, published at least 90 days before measurement), held in the study's frozen data store. 1,191,937 videos across 16,267 channels. For each video, local cadence is the channel's count of qualifying uploads in a 60-day rolling window ending just before that video's publish date (a lag). Views are a single recent snapshot, looked at per video (per day, to hold age constant) and as the channel's total per quarter.
Within-channel (channel fixed effects) and net of platform-wide era (calendar-quarter fixed effects),
the age-controlled slope of log(views per day) on log(1 + local cadence), on the channels that
varied their cadence, reported as % change in views per day per doubling of cadence, with
channel-clustered inference, computed on the holdout once.
See design-spec.md and preregistration.md. The
pre-reg was tagged (003-v1-preregistration) before any frequency-vs-views query ran.
A small, precisely-estimated dip per video: about −10% change in views per day per doubling of cadence (95% CI roughly −12% to −9%), held-out. The interval excludes zero, so this is not a null; it is a real negative association that is practically trivial, sitting entirely inside the ±15% "worth acting on" bar we set in advance. Within a channel, videos published when it was posting more earn about the same views per day, far too small a move to be the "post more, grow faster" lift the advice promises.
Total reach still rises: because the per-video mean stays flat while the count grows, total views scale roughly one-for-one with output ("bigger pie, same slices"). That total-views view is a post-hoc follow-on, labeled as such; the per-video slope is the pre-registered, sealed-holdout headline.
Read the report.md, the methodology.md, and reproduce the
headline yourself from the public dataset, rebuilding the cadence window from scratch, with
dataset/replicate.qmd.
Published. Pre-registration tagged 003-v1-preregistration; the headline was read once off the
sealed holdout, per the labs protocol. Everything here is a correlation, with the direction unknown;
nothing was randomized.