For the reader who wants the AI / ML pipeline named.
Each Czech brand homepage (title, meta description, first headings, first non-promo paragraphs, capped at 600 characters per brand) is read by a multilingual AI language model — intfloat/multilingual-e5-base — which turns the text into a 768-dimensional vector capturing meaning across languages. The same model reads the top rising Czech queries (pytrends + hand-curated fallback). The vectors are projected to 2D for the map via UMAP and grouped into neighborhoods via KMeans (run on the original 768-dim AI vectors, not on the 2D projection). Neighborhood labels are rolled from sector composition. The opportunity score for each trend is velocity × (1 − mean similarity to the three closest brands in the AI vector space) — a high score means the trend lives in an empty zone of the map.
This is a text-mode AI pipeline. The sister projects in the series use vision AI (Color Fingerprint) and audio AI (Sound Fingerprint); the Spot Scorer combines all three into one multimodal score for any video URL.