Recommendations and Rambler/Start

Both the media and the service projects of Rambler&Co needed recommendations. An in-house system meant no third-party solutions and no spend on them.

Our work on the search platform for Rambler&Co properties had given the team access to content from across the group's sites. That became one of the prerequisites for recommendations: the Search team could work with fresh material from the group's properties, and the machine-learning department had user models.

According to Yandex.Metrica, 2.47 million users a day spent under a minute on Rambler and viewed only 1.47 pages on average. The 2.22 million users who stayed longer than a minute had a page depth of 6.16.

I interpreted the first group as people who had not found anything interesting enough to continue. My hypothesis was that if we combined the group's fresh content with user models, personalised recommendations of content and services would raise CTR, time on site and page depth, and reduce bounce.

I framed the product hypothesis and direction: an in-house recommendation system, the MVP plan, priorities and the path from API to consumer products. I connected the Search team's capabilities with the machine-learning department's work and took recommendations from the homepage test to their use in Rambler/Start. The machine-learning department built the user models; my team built Rambler/Start.

Develop an in-house recommendation system for Rambler&Co's media and service content and validate it with an A/B test on the Rambler homepage.

Plan the MVP as a feed of news, articles and service widgets, with user-model improvement built in from the start: encouraging sign-in, linking anonymous identifiers to accounts, and retraining the model on user activity and questionnaire answers.

Bring recommendations into Rambler's own consumer products. A browser start page and extension were part of the plan from the beginning.

We built a recommendation API. Articles from Rambler&Co's publications were grouped by similarity into topics and ranked by how interesting each topic was to the individual user.

We tested the hypothesis on the portal's live audience: we put recommendations from the API on the Rambler homepage and ran an A/B test.

The same API was then used in Rambler/Start. Rambler/Start was a Chrome new-tab extension with Rambler search, traffic, weather, currency and mail widgets, shortcuts to Rambler services and frequently visited sites, history, downloads and a personalised feed of content from Rambler&Co properties.

In Q2 2017 we optimised the recommendation API and added more of the group's properties to Start. We built a promo landing page and, with the Platform department, a feedback form for people who removed the extension.

Before launch we had to clear legal review and resolve issues with images in the feed. Chrome, Opera and Firefox were all planned, but Opera stopped supporting new-tab replacement and Firefox required extensive additional work.