📡 Competitor Signal Radar — Sentiment & Topic Clustering

The radar chart has one spoke per topic, drawn once per competitor as an overlaid polygon, so its shape answers "which competitor is being talked about, on which topics, right now, and is that positive or negative" — it is built for comparing competitors against each other. The bubble chart collapses the competitor dimension and instead plots every topic on axes of sentiment (x) and volume (y), so its shape answers "which topics are hot and which are toxic across the whole market" — it is built for spotting outliers at a glance. Together they cover the two questions a competitive-intelligence analyst asks most: who, and what.

Each competitor in this simulation is assigned a fixed probability distribution over the five topics — for example Meridian is configured to generate far more Outage/reliability and Customer complaint mentions than the other two, simulating a competitor with a rockier operational track record, while Astra skews toward Pricing and New feature launch mentions, simulating a company that is shipping and marketing aggressively. This mirrors real competitive intelligence, where different companies genuinely do generate very different mixes of press coverage, social chatter, and review content depending on what is actually happening inside each business.

A Poisson process models events that occur independently at a constant average rate, with the time between consecutive events following an exponential distribution. It is the standard model for arrival streams that have no memory of when the last event happened — customer mentions, support tickets, and web requests all behave this way reasonably well in the short term. This simulation draws each inter-arrival gap from an exponential distribution parameterised by the arrival-rate slider, so raising the rate does not just add more events uniformly — it also changes the burstiness pattern in a statistically realistic way.

This simulation never reads any actual text — topics and sentiment are drawn straight from probability distributions, so it cannot capture the things real text analysis lives on: sarcasm, mixed sentiment within a single mention, entity disambiguation (is "Vantek" the company or someone's cat?), source credibility weighting, language and dialect variation, or the fact that a single influential post can be worth more than a thousand ordinary ones. Real systems also have to deduplicate reposts, detect bot amplification, and decide how much to trust an anonymous account versus a named journalist — all judgement calls this simplified generator skips entirely by construction.