Methodology and AI transparency

How we collect, detect mentions, estimate sentiment — and what we do not claim.

Collection

We only collect publicly accessible content: RSS/Atom feeds, web pages, sitemaps, PDF/DOCX documents, official APIs and recordings of events we stream. We respect robots.txt, source terms and rate limits, and never bypass CAPTCHAs, authentication or paywalls.

Mention detection

A hybrid method: exact and normalized matching, aliases and transliterations (Cyrillic–Latin), controlled fuzzy matching, topical keywords and semantic similarity. Every match carries a method, a confidence level and its reasons; teams confirm, reject, correct or reassign mentions.

Sentiment

Sentiment is an automated estimate of the tone of a text fragment, not a measurement of public opinion. The methodology version is stored with every analysis.

Narratives and coordination

Coordination signals are indicators for human review, not accusations; each lists the concrete evidence that triggered it.

AI assistant

Answers only from retrieved documents and cites a source for every claim; without evidence it refuses. It never publishes without human approval, and customer data is not used for model training.