Vertical integration of platforms and product prominence

Abstract 

Meta-search platforms, which enable consumers to compare product prices between different sales channels, are sometimes integrated with certain channels. A case in point is the online hotel booking industry where the major online travel agencies are integrated with meta-search platforms. Studying web-scraped data from the meta-search platform Kayak, we find indications that, for a given hotel, the offers of affiliated online travel agents (like Booking.com) are more visible than those of other online travel agents with the same price. Moreover, hotels appear to be less prominent in Kayak’s search results when the rival online travel agent Expedia has the lowest price.

Appendices

Appendix A: Google Trends data

In the Kayak data, we identify 22 online travel agents. For each online travel agent, we download the associated relative search volume on Google (Google Trends) for the “search term” in France per month between 2014 and 2018.

On Google Trends, it is possible to look at searches of these keywords related to the “search term” in general, the respective website, or any other category deemed relevant by Google. As we do not observe the category “website” for all online travel agents, we use the more general query “search term” and adapt the request when necessary (for instance, “Tablet hotels” instead of Tablet).

Besides these 22 online travel agents, we distinguish between the hotel’s direct channel and large hotel chains and websites of independent hotels. We collect data from Google Trends for the nine biggest hotel chains and normalize the value to zero for websites of small independent hotels. For each reservation date, we compute a popularity index (up to 100) by sales channel defined as the current Google Trends value divided by the maximum Google Trends value among the available sales channels for the request. The average popularity is higher for online travel agents than for hotel chains (Table 10), which is mostly driven by some very popular websites like Booking.com and voyages-sncf.fr, the French national railroad ticket booking website.

Table 10 Sales channels’ popularity index

The Google Trends index is subject to considerable variation during the observation period. Figure 4 displays the Google Trends Index for the three main online travel agents. Besides seasonal patterns, there is a general upward trend suggesting main online travel agents are becoming increasingly popular. This is consistent with previous work by Hunold et al. (2018) showing that Booking.com gained in popularity in Europe from 2014 to 2017.

Fig. 4

Variation of Google Trends index for main online travel agents

Appendix B: Robustness checks: Chains

Table 11 Hotel chain affiliation

Table 12 Visibility and Position Leadership by chain affiliation

Table 13 Hotel ranking by chain affiliation

Appendix C: Robustness checks: Sales channel

Table 14 Price vs Position leader by sales channel

Table 15 Visibility and Position Leadership at the sales channel-level

Table 16 Hotel ranking at the sales channel level

Appendix D: Google Hotels data

For the analysis conducted in Section 5.4, we obtained additional data from Google Hotels. It covers a period from April, 15, 2015, to September, 22, 2015 (while the data set of Kayak goes from October 2014 to September 2017). In the following paragraphs, we describe the preparation of this data set and how we combined it with the data from Kayak, and we also discuss potential discrepancies.

Data collection

The data from Google Hotels was obtained by a similar procedure as the Kayak data in Larrieu (2019). Search results were collected every day from 6 to 8am using a web scraping program from a Windows desktop. IP addresses were randomized in each iteration using a list of French IPs located in the Paris region. For each iteration, our browser cache was cleared of all cookies and historical searches so as to appear as a new user without any personalization that may affect the Google Hotels ranking algorithm. This lead to 961 search results for 156 distinct reservation dates for one night for two people with different time horizons (mainly 4, 14, 30, and 180 days before arrival).⁵⁴ In total, data for 2,260 different hotel names was collected, which decreased to about 1,100 unique hotels due to different naming across time.

Data merge

Hotels across the two platforms Kayak and Google Hotels were matched by their name similarity and manually. As a result, 1,093 distinct hotels could be matched. Thereby, we cover 61% of the initial sample (1,784 hotels). However, taking into account that we only consider the period from April, 15, 2015, to September, 22, 2015, the coverage increases. Out of the 1,215 hotels present on Kayak for the relevant observation period, we identify 850 hotels also on Google Hotels, yielding a coverage of 69%. This overlap is comparable to the overlap of about 70% of hotels on Booking.com and Expedia in Europe as found in Hunold et al. (2020). The data was then further amended with data from TripAdvisor as before.

Matched vs. non-matched hotels

One might wonder whether the subsample of hotels for our robustness analysis differs from the overall Kayak sample used in the main body of the article. In the following, we provide comparison tables discussing these points. First, we report the statistics of the whole Kayak sample (Panel A) versus the subset of hotels present on Kayak during the observation period of our robustness check (Panel B) in Table 17. Here, we note that while the average number of stars is slightly higher in the “shorter” period, the hotels appear to be smaller in terms of rooms and belong less often to a chain. Second, within the “shorter” time period, we can compare hotels on Kayak which we observe on Google Hotels as well (847, Panel C) and those we do not (365 + 3, Panel D).⁵⁵ It appears that hotels which are present only on Kayak and not on Google Hotels have slightly less stars (3.22 vs. 3.25), a lower number of rooms (49 vs. 62) and are less often part of a hotel chain (21 vs. 24%).

Table 17 Characteristics of hotels in the Kayak data set

These 847 distinct hotels correspond in the matched data to 300 requests, i.e., search results of hotels on both meta-search platforms, reflected in 262,811 hotel positions. The majority of requests (73%) were made for a lead time of 4, 14, 30, or 180 days. In the analyses, we will restrict the sample to the i) hotels which are also available on Kayak, ii) the same dates, iii) the same booking horizon, and iv) the period when all forms of price parity clauses were still allowed, that is, until July 1, 2015.

Effects of data filtering

Finally, we verify that we get qualitatively the same results with the data from Kayak even when restricting the sample in the same dimensions as the Google data set. Table 18 shows the result of the baseline model for the Kayak data until July 1, 2015, using all available hotels, travel dates, and booking horizons. Subsequently, we restrict the sample to the same observation period (column 2), the same set of hotels in both data sets (column 3), and apply both restrictions in column (4). Finally, in column (5) we restrict the Kayak sample further to hotel-travel date observations with the same booking horizon as available in the Google data. One can see that even when imposing these restrictions, results remain qualitatively the same.

Table 18 Kayak hotel ranking - different filters