Netflix puts a billion dollars a year on recommendations
On 28 December 2015 Carlos Gomez-Uribe and Neil Hunt of Netflix described the company's recommender system in ACM Transactions on Management Information Systems: it influences the choice of about 80 per cent of hours streamed, with search accounting for the other 20. The authors estimated that personalisation and recommendations together save the company more than a billion dollars a year by reducing subscriber churn.
Why it matters
The company that deployed the system put money on it: not a gain in accuracy but retention of subscribers, measured by A/B tests on millions of people over months. It is the first monetary figure in the atlas for a recommender system, and it is the company's own estimate.
The billion is introduced with 'We think', that is, as an estimate. At the time of writing: more than 65 million members and more than 100 million hours streamed a day; consumer research the authors cite says a member loses interest after 60 to 90 seconds of choosing, having looked at 10 to 20 titles. An A/B test usually runs two to six months, and detecting a retention difference of 0.1 per cent at 50 per cent needs about 2 million members per cell. Netflix's annual report for 2015 (10-K, 28 January 2016) gives over 75 million members in over 190 countries and more than 125 million hours a day, a later measure of the same quantity; neither the 80 per cent nor the billion appears in it.