Social listening

Where the engine listens

The engine is fed by public social conversation in the two languages most relevant to the product. English comes from Twitter for a broad global signal; Chinese comes from Xiaohongshu, Weibo, and Douyin for the primary China market. The same six analytical dimensions are applied to every platform so findings stay comparable.

Connected data sources

0 posts, 2 languages, 4 platforms

TwitterEnglish

0connectedBroad global flavour and pain-point signal

XiaohongshuChinese

0connected (84.5% positive)Flavour discovery and demand territory

WeiboChinese

0connected (80.4% positive)Open opinion, complaints, competitor talk

DouyinChinese

0connected (79.5% positive)Candid sensory reactions and pain points

Connect a new sourceThe engine is source-agnostic. Only the datasets above are available in this demo build.

How the data is pulled and prepared

Scrape per platformClean, dedupe, relevance filterSentiment and entitiesSegment

Collection is organised around six structured query groups. English is collected with an authenticated Twitter scraper; Chinese is collected with self-hosted crawlers and managed actors across Xiaohongshu, Weibo, and Douyin. Each post is retained with its metadata, author identifiers are hashed so no personal identity is stored, and Chinese posts that mention no protein or beverage term are removed by a relevance filter.

Dual-model sentiment. Sentiment is scored on original-language text with a consensus routed by language: VADER and TextBlob for English, a RoBERTa Chinese model and SnowNLP for Chinese. The final score is the average of the two models, and a confidence value is one minus their absolute difference, so posts where both models agree are weighted more heavily.