At Twitter, the Brand Strategy team partners with high profile brands to push the boundaries of social media advertising. Our remit has been to leverage insights and creativity in order to create ads that people can feel excited about—a high bar in the online advertising landscape.

As the sole data scientist on the team, I have been involved in countless advertising campaigns for top brand accounts in the US and abroad. While the scope of my work has mostly focused on conversation analysis, projects like Trend Genius have been an opportunity to contribute directly to the production of campaigns on the platform.

Over the course of my tenure at Twitter NEXT, I have often produced research for luxury and fashion brands including Louis Vuitton, Dior, and others. Those research projects gave us great insight into the fashion world online.

Among many particularities, I was always interested in the importance of brand ambassadors in the fashion conversation. For fashion houses to dress celebrities has been common practice for ever. But in the recent past we've seen it systematized, as Louis Vuitton started recruiting a rolling cast of brand ambassadors, including Zendaya, Bradley Cooper, Lebron James, and many others. At Twitter, we saw how this shift impacted the fashion conversation. Viral moments involving fashion brands rarely focused on product launches anymore: they were instead driven by celebrity fandom, and the advertised products became a secondary center of attention.

This change in strategy created an opportunity for Louis Vuitton. While the brand was able to launch ambassador-driven social media campaigns that were timed with their product launches, they were missing out on the many viral moments that their sponsored celebrities were generating organically during the rest of the year. Typical viral moments the brand was interested in included Zendaya being seen on a red carpet, Bradley Cooper cheering for his team at a NFL game, or Lebron James setting a new record on the court.

How could Louis Vuitton capitalize on the excitement that brand ambassadors generate all year long, beyond product launches? I started working on Trend Genius as a way to answer this question.

In a nutshell, the idea behind Trend Genius was to monitor the Twitter mentions of brand ambassadors, and to automatically launch customized brand campaigns when they started trending. This was a simple solution (the kind that makes you wonder why we were not already doing it) with some interesting twists. Because the Twitter Trends API would not work for our purpose, I quickly opted to create our own data pipeline and trend detection model.

I started by exploring how Louis Vuitton ambassadors were discussed on the platform. In order to understand the shape of viral moments, I analyzed conversation volume, impressions, and various meta data. For each of those data points, I looked at the distribution of trend lengths, lead times, and baseline variances. This exercise highlighted the need for our prototype to account for various levels of conversation volume (e.g Zendaya generates 10x more conversation than Tyler the Creator) and seasonality (e.g Lebron James trends 10x more often than Zendaya).

Preliminary research into viral celebrity moments enabled us to quantify the lag between posts and user impressions. This proved helpful when picking an approach to trend detection.

As an initial prototype, I built a simple model that relied on a combination of a standard score (Z-score) as a measure of activity intensity, and an exponential weighted moving average (EWMA) as a measure of steady increases in mentions. We ultimately selected this approach because of its simplicity and precision. In comparison, machine learning methods like gradient-boosted trees offered shorter lead times in trend detection, which wasn't as beneficial since I had found that trending messages took an average of two to six hours before they were seen by to the majority of our users.

We designed a simple UI that provided visual feedback based on selected threshold parameters. Red highlights indicated trending moments detected in the training data.

The interface I prototyped enabled our team to fine-tune activation thresholds based on a trade-off between precision/recall and trend frequency. As trends unfolded, our team of brand strategists were able to update trend rates based on moving advertising budgets.

^ This was an actual question asked during a client meeting. Thankfully we were prepared for it.

The real-time nature of trend detection posed a serious challenge to brand safety. Namely, how could we prevent the launch of an ad campaign, when the associated celebrity trendeded for the wrong reasons? We tackled this issue from two angles, both preventative and real-time. To do this, I designed a process that identified predominant topics discussed in mentions of our brand ambassadors.

Most engaged topics in the Zendaya conversation in 2025. Circles represent tweets, clustered/colored by topic, and sized by engagement. See methodology below.

This analysis relied on manifold learning. I had been using this method since my Twitter Aurora days, except this time, the data input wasn't based on follower graph, but on tweet content annotated by a large language model (LLM). By prompting a LLM, we were able to obtain embedding coordinates for any given tweet, and eventually construct a n-dimensional space that described an entire conversation. In order to visualize the n-dimension output, we applied a manifold learning algorithm (t-SNE), which turned it to two dimensions. Finally, a clustering algorithm (HDBSCAN) and some additional LLM annotations enabled us to identify clusters in the resulting landscape of tweets. At the end of this process, we had effectively generated a visual map of a given conversation, where tweets similar in meaning appeared close to each other. Applying this analysis to the top 1,000 most engaged mentions of a celebrity on Twitter, we were able to quickly identify how that person was being discussed over any given period of time.

This topic analysis was useful in two different ways in the context of Trend Genius. First, we used it during preliminary research as a way to identify potential issues associated with Louis Vuiton ambassadors, and build a stoplist of keywords that would prevent Trend Genius from activating when those came up. Second, the analysis could be used in real time by the client as way to assess how brand-safe a given trend was during activation.

When Bradley Cooper unexpectedly turned up at a Taylor Swift concert in 2023, the conversation exploded on Twitter. Louis Vuitton was ready to seize the moment as Trend Genius launched an activation that targeted Bradley's fans on the platform.

Trend Genius for Louis Vuitton was a resounding success and a demonstration that online advertising doesn't have to be creepy. In fact, it can get the right audience downright excited: Trend Genius ads saw a link-click rate twice higher than concurrent Louis Vuiton ads on Twitter.

Since then, Trend Genius has generated millions of dollars in revenue, and was deployed on behalf of high profile clients including Google, Microsoft, McDonald's, T-Mobile, and many others. It also opened the door to more creative experiments, including the monitoring of in-game plays during sport events, and the launch of customized ads based on weather patterns.

A selection of brand accounts that have used Trend Genius, since the initial campaign for Louis Vuitton.

As of November 2025, Trend Genius is on track to be rolled out as an ad product offered to all brand accounts on Twitter.