Elizabeth Windram

Google web search established the default search results page as a list of N text-based results. YouTube adopted this standard at some point before my time, so the default search result was a video thumbnail, accompanied by title, description and other textual information, and this was shown in a list of twenty results. Having worked on Google web search and knowing next to nothing about video search, I was curious to better understand what people were searching for on the world's largest search engine and if there were ways we could or should be thinking about video results differently than standard web results.

To answer the question above, a quantitative researcher and a PM kicked off an investigation to analyze our queries. We started with a quick and dirty approach of grabbing the 1000 top queries in a day and hand-categorizing them. We then followed this up by looking at a random sample of queries that were categorized by trained raters.

We first split queries into two main buckets: navigational and browsy (navigational being when there was only right answer, browsy being there were a subset of videos that could answer your query.) We then split these into finer categories that included navigational meme (keyboard cat), navigational music (bad romance), broad newsy (japan tsunami), broad informational (how to tie a tie), and others. As a quick way to get feedback, we launched a 1% experiment with multiple columns of results. This wasn't perfect, but was intended to be a starting point to gather data. We intentionally wanted to keep as much metadata to most closely mimic the existing results. CTR between the top three results definitely broke it's "normal" pattern, CTR on the right column went way up and pagination through results increased.

While we were gathering the above information, I started organizing a brainstorming session to get folks together from different disciplines (front-end eng, search quality, product, design, and research) and from different offices (Tokyo, San Bruno, Mountain View). We spent two hours listening to folks present data we had on video search (from Google and from YouTube), as well as demos people had built. I then split people up into teams and each team spent the afternoon generating ideas for we could best handle different types of queries. e.g., one team focused on navigational queries, another on newsy queries, another on broad informational queries, and so on.

We designed, built, and launched two of these concepts shortly after. For navigational searches (e.g., lady gaga born this way), we launched an enlarged thumbnail for the navigational result. We chose to be conservative in our triggering and opted for the most conservative of the UIs I developed, so that we could get an experiment out the door quickly. I was concerned people would skip over the large thumbnail, assuming it was an ad, but the numbers were resoundingly positive and we pushed it out the door. We also launched the artist onebox (a result block that listed out the most popular tracks from an artist's discography) as a first step to better answering broad music queries (e.g., lady gaga). We then turned our focus to the idea of using concepts as building blocks to help people build better queries. For more on this, check out YouTube topics.

The core YouTube search team is about ten people. For the projects mentioned here, I worked mostly with a PM, the engineering team, and a quantitative researcher. The brainstorming session was a much larger interdisciplinary group and included folks from many related teams. I later worked with another designer who helped with ad formats, oneboxen, and additional explorations for browsy queries.