media-monitor

Media Monitor

Raw mention feeds are noise; the costly failures are symmetric - missing the forum post that becomes a national story eighteen hours later, and escalating a comms fire drill over three sarcastic tweets. This skill turns mentions into an actionable picture: queries precise enough to trust, sentiment grounded in text, numeric thresholds that separate shift from chatter, and a digest that lets a comms team act within minutes.

Operating procedure

Query design comes first because every downstream number - volume, sentiment, spike detection - inherits the query's precision; a noisy query makes every alert a false one.

Step 1: Gather inputs

1. Entities to track: brand names, products, executives, competitors, campaign hashtags.
2. Known ambiguities: homonyms, generic words in the name, celebrity name collisions.
3. Sources in scope and any platform access limits.
4. The team's alert appetite: who gets paged, and how many false alarms per week they will tolerate (default: tune for under 2).
5. Baseline history if it exists - 30 days of mention volume; without it, run 7 days silent to establish the baseline before alerting.

Step 2: Design the queries

• Include common misspellings, abbreviations, and acronyms for every entity.
• Exclude ambiguous homonyms with negative keywords, and iterate until precision holds.

Bad: Apple - drowns in fruit, recipes, and idioms; every metric downstream is garbage. Good: ("Apple" OR "AAPL" OR "@Apple") AND (iPhone OR Mac OR Cupertino OR earnings) -fruit -pie -recipe -crumble

• Tier the sources: news (wire, national, trade), social (X, Reddit, forums), and owned channels. Weight by reach and credibility - one wire-service story outweighs fifty anonymous forum posts for severity, but the forum is where stories start, so it is watched, not ignored.
• Test each query on a sample of 50 recent hits; require at least 90% of results to be genuinely about the entity before the query goes live. Below that, alerts will be ignored within a week.

Step 3: Classify each mention

Capture for every item: source, date, reach estimate, sentiment, and topic/theme. Sentiment rules:

• Three labels - positive / neutral / negative - plus a one-line rationale quoting the text.
• Ground sentiment in the body text, never the headline alone; headlines are written for clicks, not accuracy.
• Mark low-confidence calls explicitly, especially sarcasm and irony - an over-confident sentiment score is worse than an honest "unclear."

Step 4: Detect narrative shifts with numeric thresholds

A single negative post is noise; a shift is signal. Alert on:

• Volume spike: mentions exceeding the rolling 7-day baseline by 2x or more. Use the rolling baseline, not a fixed number, so growth and seasonality do not fire false alarms.
• Sentiment swing: net sentiment (positive minus negative, as a share of total) moving more than ~20 points period-over-period.
• New framings: keywords or phrases prominent this period that were absent last period - the earliest shift signal, and the one volume metrics miss.
• Source migration: a story jumping tiers, forum to trade press to national. Migration is the strongest predictor that a story is about to get bigger; page on it even when volume is still small.

Step 5: Triage severity

• Critical: legal or safety claim, an executive named, volume accelerating, credible source or migrating upward. Page now.
• Watch: rising volume, mixed sentiment, contained to one platform. Recheck within hours, note in the digest.
• Routine: normal chatter, stable sentiment. Digest only.

Step 6: Report

Deliver the daily digest in this order:

1. Headline numbers: total mentions, net sentiment, both versus prior period.
2. Top 3 stories driving the conversation, each with reach and sentiment.
3. Narrative shifts detected, with the threshold that tripped and why it matters.
4. Recommended action per critical item: respond / monitor / escalate.

When citing any mention, link the primary source and quote verbatim - never paraphrase a quote in a way that changes its meaning.

Worked artifact: daily digest template

MEDIA DIGEST - [FILL: brand] - [FILL: date]

HEADLINE NUMBERS
  Mentions: [FILL]  (7-day baseline: [FILL], ratio: [FILL]x)
  Net sentiment: [FILL] pts  (prior period: [FILL], change: [FILL] pts)

TOP 3 STORIES
  1. [FILL: story] - reach [FILL], sentiment [FILL], link [FILL]
  2. [FILL]
  3. [FILL]

SHIFTS DETECTED
  [FILL: e.g., "volume 2.4x baseline, driven by <framing>; migrated from
   r/<sub> to <trade outlet> at 14:00 - migration threshold tripped"]

ACTIONS
  CRITICAL: [FILL: item] -> [respond / monitor / escalate] - owner [FILL]
  WATCH:    [FILL: item] -> recheck at [FILL: time]

Deliverable

Produce a monitoring package: the query set per entity with negative keywords and the 90% precision test result, the source tiers with weights, the alert thresholds (2x volume, 20-point sentiment, new-framing and migration triggers) tuned to the team's false-alarm tolerance, and the daily digest in the template above with a recommended action per critical item.

Do NOT

• Do not alert off a raw mention count - always ratio against the rolling 7-day baseline, or growth guarantees permanent false alarms.
• Do not score sentiment from headlines; headlines are engineered for clicks and routinely invert the body's tone.
• Do not treat organic and coordinated activity as one signal - flag suspicious clusters (identical phrasing, synchronized timing, young accounts) and report them separately.
• Do not infer causation from a single correlated spike; note the correlation and keep watching.
• Do not paraphrase quotes; verbatim with a link, every time.
• Do not dox individuals - aggregate patterns, respect platform terms and privacy.
• Do not ignore small-volume stories that are migrating up source tiers; migration outranks volume as a warning sign.

Quality bar

• Every live query passed the 90% precision sample test, and the test is recorded.
• Every sentiment label carries a text-grounded rationale, with low-confidence calls marked.
• Every alert in the digest names the numeric threshold that tripped it.
• Every critical item has a recommended action and an owner.
• Bot-suspect clusters are flagged and excluded from the organic numbers.