Methodology

How WIO measures community attention

WIO analyzes public Reddit posts and comments in monitored communities and publishes two kinds of metrics: what is most discussed, and what is genuinely changing. This page defines the terms used across the viewer.

A trend compares seven complete UTC calendar days with the immediately preceding seven complete UTC calendar days. The anchor is the latest day with successfully analyzed source records that is no later than the day before publication; the publication day itself is excluded because it is incomplete. A stale community remains explicitly stale — windows never slide forward using wall-clock time without new analyzed data.

Prevalence is the share of distinct analyzed posts and comments that mention an entity: the entity's distinct-document count divided by the window's total distinct analyzed records. The trend score is the prevalence change in basis points — 10,000 × (current prevalence − comparison prevalence) — displayed as a signed percentage-point change (for example +1.8 pp for a score of 180). A rising or falling signal is therefore a change in prevalence, never a change in raw mention counts: activity swings alone never count as trending.

A comparison is ready only when each window contains at least 100 analyzed source records. A rising or new signal is eligible only with at least 5 distinct current-window posts or comments behind it; a falling signal needs at least 5 distinct comparison-window records. "New" means the entity is absent from the comparison window — not new to the world or to Reddit.

Entities are reported as new, rising, steady, or falling under those rules. When coverage is insufficient, WIO says so explicitly — insufficient_current, insufficient_comparison, or unavailable — and never fabricates a zero-change signal.

Most discussed

The most-discussed rankings use the current publication lookback for each community and count NER mentions. Mentions count every occurrence, so one post can mention an entity several times; the "Seen in N posts/comments" support counts distinct source records instead.

Frequency is never called trending. If a trend window is insufficient, the community shows its most-discussed view instead, clearly labeled, with its exact analysis range.

Attention over time

The attention chart on an entity page plots the trailing seven-day share of analyzed posts and comments that mention the entity — the entity's distinct-document count divided by the seven-day window's total analyzed records — across 90 daily points over 96 complete UTC coverage days. It is the same prevalence measure the trending signals use, averaged over overlapping rolling windows instead of two fixed ones, so a change is never raw mention volume or subreddit activity.

Each endpoint covers seven complete UTC days and always ends at the latest complete analyzed day — the series never extends into the current incomplete day, and a stale community stays visibly stale with no synthetic zero tail.

A point is comparable only when its seven-day window contains at least 100 analyzed source records. A covered absence inside a comparable window is a valid zero; a window below the coverage floor is labeled "Low coverage" and a missing analyzed day is labeled missing — neither is ever turned into a zero or joined into an interpolated line.

Entities are extracted automatically and may be wrong, and the chart is a derived product metric, not an immutable archive. It makes no causal, statistical-significance, trending, or predictive claim; the canonical trend support on the same page remains the only directional statement.

How entities are extracted

WIO extracts named entities from public Reddit posts and comments with an automated model, then applies correction and ranking passes. Extraction is automatic and may be wrong — entity cards and evidence views carry that caveat.

Evidence views show up to the five most recent dated excerpts of the public posts and comments behind an entity, labeled Post or Comment, with the matching mention highlighted. Excerpts are cleaned for display (links are removed the same way the extraction model saw the text) and rendered as plain text — untrusted content is never injected as HTML or Markdown.

Evidence lookups use exact published keys only. WIO never synthesizes Reddit URLs from stored ids, and no database id is ever derived from a URL.

Freshness

Every community shows the time of its last successful publication, and trend headers carry the exact published windows and analyzed-record totals behind them.

A community whose successful publication is more than 30 complete UTC days old is explicitly marked stale.

Failed or degraded runs never make stale data look fresh: the last known-good snapshot stays published until a full successful run replaces it.

Limitations and privacy

WIO reports directional product signals, not causality, statistical significance, sentiment, or endorsement. Coverage is limited to the monitored communities and is never real-time.

Excerpts are public Reddit content; WIO does not synthesize Reddit URLs from stored ids, and excerpt text is handled under the published data-retention contract — retained only for the documented window and never exported in bulk.