Scientists from Stanford, Pandora and Netflix have analyzed the effects of ads on listeners. They have used a broad dataset from Pandora, representing the real behaviour of 35M users during 21 months. All other parameters remaining constant, what happens when the amount of ads is varied?
The results are presented in the paper available at this page: http://www.davidreiley.com/papers/PandoraListenerDemandCurve.html
TL;DR; They have shown that:
- broadcasting more ads makes people consume the media in shorter sessions, and less often. Broadcasting 12 ads per hour instead of 3 makes people listen 5% less (in total hours). At the root of this reduction, 18% of people listen less every day, 41% listen less often, 41% do not listen at all anymore.
- between 3 and 12 ads per hour, the economics show that broadcasting more ads is marginally profitable. Most of the additional money comes from ad revenue, not from people switching to a paid, ad-free plan. To avoid ads, older people are more prone to pay for an ad-free content than younger people, that most often leave the service.
The average duration of ads was not given in the paper. David Reily kindly answered my question: it is approximately 25 seconds, on average.
Interesting quotes from the article:
More than 80% of listening now takes place on mobile devices.
The coefficients show us that one additional ad per hour results in mean listening time decreasing by 2.075% ± 0.226%, and the number of active listening days decreasing by 1.897% ± 0.129%.
Had we run an experiment for just a month or two, we could have
underestimated the true long-run effects by a factor of 3.
This tells us that approximately 18% of the decline in the hours in the final month is due to a decline in the hours per active day, 41% is due to a decline in the days per active listener, and 41% is due to a decline in the number of listeners active at all on Pandora in the final month. We find it interesting that all three of these margins see statistically significant reductions, though the vast majority of the effect involves fewer listening sessions rather than a reduction in the number of hours per session.
For each additional one ad per hour during the experiment, we see a 0.14 percentage-point increase in the probability of being a paid subscriber at the end of the experiment. Multiplying by the subscription fee of $5 per month, we see that Pandora picks up additional monthly subscription revenue of approximately 0.75 cents per listener for each increase of one ad per hour (minus payment-processing costs). This turns out to be considerably smaller than the effects on advertising revenue implied by the demand-curve estimates above.
For each listener converted to a subscription by increased ad load, three more listeners leave Pandora entirely.
We see that listeners over 55 years old are twice as likely as listeners between the ages of 13 and 24 (marginal impact of 0.21% versus 0.09% or less) to react to an increase in ad load by paying for the ad-free service.
We find weak evidence to suggest that redistributing advertising across more interruptions, conditional on the number of ads per hour being held constant, may have a small impact on customers’ propensity to listen to Pandora.
I have discovered this article thanks to a post in a Hacker News discussion: