I agree, intuitively that makes much more sense. At a minimum, you should get some portion of accessories vs replacement items in your recommendations. I'd be

I agree, intuitively that makes much more sense. At a minimum, you should get some portion of accessories vs replacement items in your recommendations. I'd be blown away with surprise to learn that AMZN hasn't optimized for revenue in the calculation though, so it seems likely that we're wrong. Of course we'd prefer AMZN to optimize for customer happiness/benefit instead. To me, the "smart" approach would be to recommend accessories until and unless the customer shows some signal indicating that they are looking for another TV. Then switch to full replacement strategy (or even return-minimization strategy) mode. That or similar logical approach may have been abandoned after getting better results (by some metric) from the ML-optimized approach. Assuming there's a human involved at all.