Customer research on Amazon contributes to the offline path to purchase

By: German Zenetti, Data Scientist

Many advertisers are curious about how Amazon fits within customers’ purchase decisions and seek to use insights related to these topics to inform their own marketing strategies.

Story highlights:

In a recent opt-in survey of 8,000 randomly selected customers in-market for a laptop or smartphone, we found that customer research in Amazon’s store signaled and accurately predicted purchase intent, regardless of whether the final purchase was made on Amazon or elsewhere.

In addition to the "Add to Cart" button, Amazon also offers a large collection of product information, reviews, prices, and other, similar information. Customers often uses these resources to help inform their purchase decisions, regardless of where they may make a purchase. Advertisers and sellers want to understand Amazon’s role as a source of product information in the customer purchase journey, beyond just driving sales in Amazon's store. To help answer this question, we developed a ratio to compare on-Amazon and off-Amazon purchases in a given category.

Top-performing advertisers have on average 40% higher year-on-year sales growth, 50% higher year-on-year growth among customers viewing their product pages on Amazon, and a 30% higher return on their ad spend (ROAS). These advertisers represented 6% of the 7,000 advertisers included in the study.

1. Advertisers can estimate the impact on-Amazon research has on non-Amazon purchases.

Amazon is frequently a part of the customer offline path to purchase. To measure this, we used customers' reported purchases both within and outside of Amazon’s store to calculate the amplifier ratio: for every customer who reported researching and purchasing a product on Amazon, how many customers researched on Amazon and purchased off Amazon?

We found amplifier ratios of 1.6 for laptops and 2.1 for smartphones. That is, for each customer who researched and purchased laptops or smartphones on Amazon, 1.6 and 2.1 customers, respectively, said they researched on Amazon and purchased off Amazon. When predicting the amplifier ratio using our machine learning model, we were able to accurately predict the amplifier ratio of smartphones using laptop data and vice versa.

Amplifier ratio in laptops and smartphones

What if your laptop products is not bought on Amazon?

For every 10 buyers in laptops, there are 16 other users who research on and purchase elsewhere.

What if your smartphone products is not bought on

For every 10 buyers in smartphones, there are 21 other users who researchon and purchase elsewhere.

Our research emphasizes the opportunity for advertisers to help inform these highly active customers when they are researching on Amazon prior to buying the product on Amazon or elsewhere. Advertisers should not limit their online return on investment (ROI) solely to online conversions. If an advertiser relies solely on on-Amazon return on ad spend (ROAS) as a key performance indicator (KPI) for determining budget allocation, the impact of their marketing dollars on customers’ offline purchase path may be misrepresented. Advertisers should consider marketing their products on Amazon as an opportunity to affect not only on-Amazon purchases but also off-Amazon purchases.

2. Customers who perform more research on product pages than average are more likely to make a purchase, either in Amazon’s store or elsewhere.

Detail page views, clicks on product pages, and time spent on product pages all correspond to customer purchases, regardless of where the purchase takes place. Our analysis of survey responses shows that 31% of customers who spent 10 minutes or more on product pages ended up purchasing a product within the category, with 14% of customers purchasing on Amazon and 17% purchasing off Amazon. The likelihood of an on-Amazon purchase rises faster than off-Amazon purchases with increased customer research activity. Therefore, we can expect relatively greater non-Amazon sales in groups where activity levels are low, and a greater share of Amazon purchases in groups where activity levels are high.

Yellow circle key: Off Amazon purchase

Off-Amazon purchase

Turquoise circle key: No/not yet purchase

No/ not yet purchase

Blue circle key: On-Amazon purchase

On-Amazon purchase

Detail Page views activity groups

Detail Page views activity groups. Group 0-2: Off-Amazon purchase: 2%; No/ not yet purchase: 85%; On-Amazon purchase: 13%. Group 3-7: Off-Amazon purchase: 6%; No/ not yet purchase: 77%; On-Amazon purchase: 17%. Group 8+: Off-Amazon purchase: 4%; No/ not yet purchase: 82%; On-Amazon purchase: 18%.

Clicks on product pages per DPV activity groups

Clicks on product pages per DPV activity groups. Group 0-1: Off-Amazon purchase: 4%; No/ not yet purchase: 82%; On-Amazon purchase: 15%. Group 1-2: Off-Amazon purchase: 15%; No/ not yet purchase: 68%; On-Amazon purchase: 17%. Group 2+: Off-Amazon purchase: 22%; No/ not yet purchase: 60%; On-Amazon purchase: 18%.

Minutes on product pages activity groups

Minutes on product pages activity groups. Group 0-2: Off-Amazon purchase: 2%; No/ not yet purchase: 85%; On-Amazon purchase: 13%. Group 2-10: Off-Amazon purchase: 6%; No/ not yet purchase: 79%; On-Amazon purchase: 16%. Group 10+: Off-Amazon purchase: 14%; No/ not yet purchase: 69%; On-Amazon purchase: 17%.


Survey respondent selection and validation: For our analysis, we randomly selected 8,000 laptop or smartphone shoppers and surveyed those who had not purchased in the category in Amazon’s store (between 60 days prior to the survey date and 90 days after). To validate the accuracy of the survey responses, where possible, we compared the survey responses to shopping actions on Amazon. Overall, these tests help us to increase our confidence in the survey instrument and the survey responses.

Modeling approach and calibration: Our model attempts to predict whether the customer purchased a product in the focal category elsewhere (=1) as opposed to their not having bought a product in the category (=0) within a 60-day timeframe. In our predictive model, we excluded customers who made a purchase on Amazon do not require a predictive model. To predict purchases elsewhere, we train a boosted classification tree algorithm to the in-sample training data set. The general intuition is that with boosted classification, the overall model will be able to explain different aspects of variations of the response variable. Due to this reason, the model overall may better predict the response variable compared to, e.g., a model without additional recursive model layers or compared to a bagging approach such as a Random Forest algorithm. Click here to learn more about decision tree models, and here to learn more about boosting in machine learning models.