When President Trump tweeted “Buy L.L. Bean,” on January 12th, 2017, people searching for “L.L. Bean” soared, with the President’s supporters heeding the President’s call while many long-time fans — and critics of the President — deciding to boycott the store. This tarnished L.L. Bean’s stellar apolitical brand with its CEO Stephen Smith invoking the Chinese proverb “May we live in interesting times” to describe the fallout from the tweet. Interestingly, when President Trump tweeted that Nordstorm, the upmarket department store, unfairly removed his daughter Ivanka Trump’s products from their stores, a similar surge in search traffic but with the opposite reaction was observed — the President’s supporters vowed to boycott the store while his critics proudly displayed their wares bought from the store. Nordstorm came out relatively unscathed, with Co-President Peter Nordstrom remarking that the impact on sales was “negligible,” and that it was “not really discernible one way or the other” suggesting the President’s critics and supporters have had about equal impact.
President Trump’s popularity (or infamy depending on your political persuasion) and the current political climate can explain such extreme trends in search traffic patterns. However, there are hundreds to thousands of smaller trending “events” that prompts customers to search and seek products. Advances in web analytics have made it possible to track and link customer search behavior to product Sales.
Along with several colleagues, I have been working with several online retailers to see if search traffic can better predict sales. Our research tells us that incorporating search data can improve forecasts by about 3-8%.
A Case Study
To investigate if Google searchers can be used to predict product sales better, we use four products from an online purveyor of fine food and cookware. They gave us access to the sales of these products over time (from the beginning of 2012 to the end of 2016). The products we choose typify this retailer. Product A is in the appetizer category; Product B is a specialty meat product; Product C is a Wine/Cheese gift package, popular as gifts during the holidays; and Product D is a hot-selling cookware product, that has gained popularity in the last four years.
The following chart shows the sales (in log) of the four products along with search terms that typically describe the product. For example, for Product A, the chart also plots the values of “indices” (in log) for the search terms ‘Antipasto’ and ‘Hors d’oeuvres.’ Google Trends was used to gather the intensity of search activity. Google Trends is a publicly available service that assigns an index to search terms. The index is reported on a 0-100 scale, where a value of 100 indicates peak popularity of a search term whereas 50 indicates that it is half of its peak value. A zero value indicates that there is not enough data for the search term (see Note 1). On January 11, 2017, the popularity for the search term “L.L Bean” was at a value of 9; but with President Trump’s “Buy L.L. Bean” tweet on the 12th of January, the popularity of “L.L. Bean” surged to a 100, its peak. This translates to about a 10-fold increase in people searching for “L.L. Bean.”
The chart shows how the sales of Product A matches up with the popularity of these two search terms. For example, sales for Product A peaks around the Thanksgiving season, about the same time when the searches for the terms ‘Antipasto’ and ‘Hors d’oeuvres’ peak. It takes a bit of trial and error to figure out what customers may be searching for when they are looking for your product — we talked with the product managers and zeroed in what we thought might be most relevant. For Product B, it was “Summer sausage”; for Product C, it was “Gifts for colleagues”; and for Product D, it was “William Sonoma coupon.” It is interesting to note here that the patterns for people searching for these terms closely match with the sales of the corresponding products.
Can this search information be used to better predict sales? To test, we built two models (see Note 2) to predict Sales — one that does not use search data (“baseline” model) and one that uses it (“trend” model). We built these models on data from the start of 2012 to the end of 2015. We wanted to see how these models performed for the “out of sample” 2016 year. The following chart shows the results of our models for Product A for 2016.
The chart shows the Actual versus the estimated values for Sales for both models for 2016. It also shows the average error for each week for both the baseline model and the trend model. The chart uses a bar for the baseline-model error and a line for the trend-model error to show the contrast — from the chart, the trend model seems to be performing better in many of the weeks than the baseline model. In fact, the prediction was 3.65% better (i.e., the error was less when search terms were included). We are able to show that for Product B, prediction is improved by 3.76%; Product C predictions improved by 5.33%; and Product D’s prediction improved by 7.66%!
Implications for Managers
Businesses have traditionally relied on historical sales and economic data to forecast demand for products. These forecasts serve as the basis for planning supply chain activities — sourcing, making, and distributing products to customers. Errors in forecasts are a source of considerable risk to profits. For example, overestimating sales leads to excessive markdowns and underestimating sales can result in lost revenues.
Search data has been used successfully to predict broad economic sentiments, disease outbreaks, or brand health. But my research shows that incorporating freely available search data into forecasting models improves product forecasts, potentially making it cheaper to buffer against market risks. Better forecasts also help utilize resources better — a better forecast would improve inventory and fulfillment — and has the potential to increase profits.
Google Trends allows you to narrow the search intensity by time and geography. We narrowed the search intensity to match the time frame of sales. We also restricted the geography to all the countries this particular retailer ships to.
For those readers interested in the technical details, see Boone, T., Ganeshan, R., Hicks, R.L., Sanders, N.R. (2018). Can Google Trends improve your sales forecast? Production and Operations Management, 27(10), 1770-1774.