A tale of two site planning strategies: Dunkin' Donuts versus Starbucks

Amid growing demand for quick service restaurant (QSR) properties, private and 1031 investors have started supplementing QSR evaluation metrics, like past sales performance, with new measures related to past site planning decisions to gain competitive edge.

The attraction of QSRs is that investors consider them to be “ecommerce resistant,” or, as Randy Blankstein, president of The Boulder Group, describes it, “no one has perfected how to get a freshly made donut, taco or burger and fries through the internet.”

Popular measure with which traders, sell-side analysts, and market researchers evaluate QSR assets include:

• Strong per site sales performance
• Track record of success whether in select market(s) or across footprint
• Willingness to find and fix gaps in market coverage
• Business strategy that can be replicated and scaled in different markets

Location components exist in each of these metrics, but silo views make it difficult to discern underlying patterns among various measures. And this situation only gets more complicated when it comes to trading competitor brands in the same market.

Recently we helped a client consider ways that comparative spending data might be used to compare the go-to-market strategies of two competitors in the same market. Lacking access to internal sales data, suitable proxy values at comparable scales were needed that could answer two questions:

1. Where did locations for each brand improve the most over time?
2. What site planning strategies can be discerned based on which brand has the more valuable locations?

We were able to build a comparative model that answered these questions (and more) using Mastercard spend scores. Let’s take a closer look at what steps we took in building this model and the various inferences that can be drawn from historical location data.

Establish parameters and proxies with location data

For this use case we’ve used a Points Of Interest (POI) dataset to map Dunkin’ Donuts and Starbucks locations across New York City. We’ve prepared our POI data in Jupyter notebooks because CARTOframes allows us to run analysis and visualization in our notebook helping simplify the extract, transform, and load (ETL) process. And, later on, we’ll push the final dataset (containing POI data joined with Mastercard spend scores) to our CARTO account using the SQL API.

The map below shows Dunkin’ Donuts (gold) and Starbucks (green) locations as of February 2018:

Next, to identify which brand has improved the most in its site selection decisions, another layer of data is needed with insights on sales activity. Focusing on sales performance of locations, not sales performance of stores, enables working with Mastercard spend scores to evaluate past performance of different areas. The model will include all five spend scores, but in determining location improvement overtime the following two scores are used:

1. Sales score: rank based on average monthly sales for the previous year up to and including the current month
2. Transaction score: rank based on average number of transactions for previous year up to and including the current month

Spend scores are available at various geographic levels. Since this analysis is looking at site selection across New York City, we’ll be interested in granular levels (census tract, census block group, census block) while comparing rankings of eating establishments in the same metropolitan area.

The give and take of granularity

Mastercard spend scores are derived from actual transactions but anonymized and aggregated to protect customer privacy in compliance with GDPR regulations. This matters when deciding the granularity at which to view spend scores because scores may not be available at more granular levels, like census block, where minimum threshold counts ensuring merchant anonymity have not been met.

Let’s take a look at available scores at different aggregations.

The benefit of aggregating at the census tract is that scores are available for a large coverage area:

When aggregating at the census block group fewer scores are available:

And at the census block, which is about one New York City block, there is another drop in available scores.

Essentially, the level of granularity should match the analysis and the desired results. For macro analysis–regional market share, coverage gaps, competitor analysis, etc.–census tracts may be the best choice. Because this analysis centers on competitor brands in the same market we’ll be selecting census block scores.

Next, with a spatial join query, we connected the POI data with census block spend scores.

Now, in our Jupyter notebook, different types of analysis visualizations can be generated.

The query above returns the median transaction score for each brand with Starbucks ranking first with a score of 844 and Dunkin’ Donuts in second with 710. As such, we can infer that Starbucks are located in high traffic areas based on the frequency of transactions.

At this point we’re ready to push this data from our Jupyter notebook to CARTO to build our model.

Mapping site planning strategies with CARTO VL

There’s not a drastic price difference between Starbucks and Dunkin’ Donuts products so we’re going to style our map proportional circles sized in relation to each location’s median transaction score. The map below displays Starbucks (green) and Dunkin’ Donuts (gold) store locations across New York City:

With CARTO VL we can style by zoom so that proportional symbology, in this case gradient circles, makes differences between store location and median transaction scores visible at different levels.

In addition, when a specific store location is selected census blocks within a 200 x 200 meter radius are styled by sales scores from February 2018. This comparative framework helps answer questions such as:

1. What is commercial activity like in nearby locations?
2. How does a specific store's sales score compare to industry performance in surrounding area?
3. Where are nearby competitors located?

In answering these questions we’ll be better equipped to assess which brand has more valuable locations.

In the Brooklyn neighborhood of Williamsburg one block has both a Starbucks and Dunkin’ Donuts location next door to one another. The image below shows the spend scores for the shared census block, but notice the difference between the relatively high sales and transaction scores and the lower than expected ticket size score, which ranks the average transaction total for the past year. Since historical spend scores date back to 2012 we can examine view performance over time for this specific block to better understand what may be causing this discrepancy.

Starting in 2014 and lasting well into 2017 there is a sharp decline in ticket size despite steadily increasing sales and transaction scores. It was late 2013 when Dunkin’ Donuts first opened its Bedford Ave. location with Starbucks opening its second Brooklyn location soon thereafter leaving none other than the Daily News to ask if this marked the official end of Williamsburg? As many reported, while Dunkin’ Donuts held a stronger foothold in the outer boroughs and Starbucks concentrated in and around Manhattan, both faced stiffer than expected competition from local Brooklyn coffee chains in Williamsburg. This may have contributed to low ticket size scores with customers perhaps sticking to coffee and not the increasing menu offerings each brand has debuted in recent years.

This kind of historical analysis can be performed for each block of interest in the original map to better assess which store accumulated the most valuable locations over time.

Next steps…

With new location data streams, like Mastercard spend scores, we were able to equip our clients with a comparative framework that included location as an indicator when making investment decisions. Further analysis can be run suited to more specific questions, but here we’ve been able to build a model for looking at individual store locations that can be explored with spend scores from 2012 to the present that reveal the sales performance of each store’s location.