Jun – Dec 2025 · Beijing

ByteDance · TikTok

AI Search Product Manager

Used foundation-model capabilities to reshape TikTok search, improving AI answer cards across intent detection, visual relevance, blended ranking, and citation interactions. I defined where AI cards added value, built the evaluation framework, and co-designed strategies with algorithm teams from serving and ranking through citations. I also structured large volumes of verbatim user feedback into actionable product and strategy directions, helping launch multiple improvements across global markets.

90%+Intent precision & recall+11%Blended-card CTR+300%Citation CTR

My role

Defined the appropriate scope for AI search cards, built evaluation frameworks, and co-developed algorithm strategies across the end-to-end experience from serving and ranking to citations.

Intent detection

  • Used random-query annotation to identify scenarios where AI cards added value, created a detailed intent taxonomy and test set, and designed the intent evaluation framework.
  • For encyclopedic entities, combined external encyclopedias, foundation-model knowledge, and on-platform video content.
  • Raised both intent precision and recall above 90%.

Visual relevance & ranking

  • Defined image and video use cases, aligned visual evaluation standards, and systematically improved image prompts, raising the end-to-end perfect-score rate to 70% while lowering false positives to 10%.
  • Designed a lightweight two-column blended layout and helped define its evaluation framework. The strategy launched globally, increasing AI-card coverage by 6% and card CTR by 11%.

Citation experience

  • Redesigned citations through a large-scale survey, verbatim user feedback, and competitive analysis—surfacing authoritative sources and domains, distinguishing on-platform from external sources, and simplifying the citation panel and path.
  • Increased citation CTR by 300%.

What I learned

Search is ultimately about efficiency and trust. Foundation models can make search faster, but that efficiency only matters when users can see the sources and trust the answer.