For particularly popular, lucrative search terms on Amazon, organically moving up towards the top of the first page of results is a long-term challenge, and buying ads on those terms can be prohibitively expensive – perhaps even unprofitable. This has led many sellers into the “magic keyword” trap – attaching an incredibly large number of keywords to their product listing, putting some budget behind ads to get a top placement if and when they are searched for, and hoping to find a diamond in the rough. The problem? The amount of possible search terms on Amazon is essentially limitless, but an analysis showed that most, at current bid levels, are likely to receive less than one click per day on average. By spreading out your budget across so many keywords with little to no volume, gathering enough data to intelligently shift budget towards the terms that actually will result in meaningful sales becomes nearly impossible. Across Amazon, less than 1% of search terms generate an average of three or more clicks per day.
This data came to light thanks to Alin Constandache, a colleague of mine and lead researcher at Teikametrics, who kept running into issues while attempting to train conversion rate models for our network of Amazon advertisers. Briefly stated, Alin tests and trains new models by selecting a random sample of keyword and click data, which is split at random into a training set and a holdout set, for out-of-sample testing.
This approach had worked well for Alin in other applications, but in April 2019 the conversion rate models he built all seemed to perform equally poorly on the holdout data, all in exactly the same ways. Digging in, Alin analyzed all keyword and associated performance data for the first four months of 2019 – this amounted to roughly 6.5 million keywords across several thousand advertisers. The result was clear, as seen below.