“we use machine learning to target platinum uh to people that we thought based on data based on previous experiments based on profile we have about 900 kind of data points that we use for that people that we thought had a higher propensity to subscribe to platinum It doubled the percentage of new users that subscribing to platinum But he did that without losing a single gold subscriber”

Use ML to find net-new premium buyers, not to cannibalize existing ones

ML-targeted upsell to a higher tier isn't about pushing existing buyers up — it's about routing the small slice of users with platinum-shaped preferences directly to platinum (instead of the safe gold default they would have settled for). Done right, you double the higher-tier mix with zero cannibalization of the lower tier. The signal is whether your ML lifts higher tiers while leaving the lower tiers flat in absolute terms.