Data updated 2026-10-23 10:07 UTC+8 MCP · 7 tools 中文
Sandbox demo Change anything: this copy of the demo data is yours alone, invisible to others and cleared when idle. Alert delivery and user management are off; words not in the data can be searched on the web live (3 per visitor).
· Self-host it for your brand →
Estimated Sentiment here is the lexicon baseline estimate: the dashboard never fans out one LLM call per mention per refresh. LLM-scored sentiment lives in briefs and MCP, labeled as such.
Demo tour

First time here? Start like this

  1. Search a word — try Claude OpenAI MCP or use the search box at the top. It first looks through collected conversation; for a word that is not there, the results page offers Search the web now (3 per visitor).
  2. See whether it matters — the big verdict on this page (calm / watch / rising / alert) answers "do we need to act now?"; below it are the stories driving it and the rules that fired.
  3. Triage mentions — on the Mentions page, mark each one to do / done / ignored, like an inbox.
  4. Add your own subscription — on the Subscriptions page, enter a brand name and its usual spellings; verdicts and mentions then filter by it. For a new name, run Search the web now once on the mentions page to get data; a self-hosted install collects on a schedule.

Your changes are visible only to you; press Reset in the yellow bar at the top to start over at any time.

How to read this page: verdict first → then the stories driving it → then the caliber (Zone 7, every threshold printed)

01Verdict · All keywords · 30d

Calm

Nothing needs action right now.

Evidence quality · high75/100

Evidence quality = how hard the evidence is (sample, sources, baseline, freshness), not how severe things are. The verdict and the evidence score are two different things.

  • google-news, hn, bluesky, lemmy, gdelt-news, reddit are newly connected (824 mentions) with no history, so they are left out of the volume ratio.
  • Volume 0.0× baseline, 12% negative: both under the watch line.
Triage inbox

03Volume & sentiment

824 mentions in 30d, no baseline yet

Measured · Estimated

Daily mentions

Daily mentions009-2400009-2700009-3000010-0300010-0600010-0900010-1200010-1500010-180020410-2146016010-23

Sentiment split

Daily mean sentiment (−1 to +1; days without mentions break the line, never interpolated)

Daily mean sentiment-10+109-2410-23

04Stories driving it

Story clusters ranked by mentions

Estimated
  1. 1 Trend
  2. 2 Trend
  3. 3
    OpenAI models are coming to Google Cloud 8 mentions · Negative 0% · first seen google-news · 3 Source
    Trend
  4. 4
    OpenAI Expands ChatGPT Ads to 26 New Countries - PYMNTS.com 7 mentions · Negative 0% · first seen google-news · 2 Source
    Trend
  5. 5
    Uber gives more than 30 , 000 staff OpenAI ChatGPT Work agent 7 mentions · Negative 0% · first seen google-news · 2 Source
    Trend
  6. 6
    Anthropic posts $80bn loss in first half of 2026: WSJ 4 mentions · Negative 100% · first seen lemmy · 3 Source
    Trend

05Source × sentiment

Source

SourcenShareNegativeMean
google-news 31438% 19+0.07
hn 13416% 14+0.02
bluesky 13016% 10+0.02
lemmy 11414% 33-0.04
gdelt-news 8911% 10+0.04
reddit 435% 10+0.07

06Data health

Data sources

SourceStatusLast storedIn windowTotal
google-news Fresh just now314314
hn Fresh just now134134
bluesky Fresh just now130130
lemmy Fresh just now114114
gdelt-news Fresh just now8989
reddit Fresh just now4343

07Caliber & assumptions

Verdict rules · Evidence-quality breakdown

ConditionResult
n < 5Insufficient data
Mentions ≥ 3.0× Baseline & Negative share ≥ 30%Alert
≥ 2.0× Baseline | Negative share ≥ 35%Elevated
≥ 1.5× Baseline | Negative share ≥ 20%Watch
—Calm
Baseline = mean daily mentions over the 28 days before the window × window days. Sources with no history before the window are excluded from the ratio (counted in every other number).
PartScore
Sample size (50+ mentions = full)30 / 30
Source spread (3+ sources = full)20 / 20
Comparable baseline (share of mentions from sources with history)0 / 25
Sentiment backend (lexicon baseline)5 / 5
Freshness (newest mention ≤6h = full)20 / 20
Evidence quality75 / 100