Price steering is the practice of “changing the order of search results to highlight specific products” and products prices. In this paper, we show an initial investigation to quantify the price steering level in search results shown to different kind of users on Google Shopping. We mimic the category of affluent users. Affluent users visit websites offering expensive services, search for luxury goods and always click on the most costly items results at Google Shopping. The goal is checking if users trained in specific ways get different search results, based on the price of the products in the results. Evaluation is based on well known metrics to measure page results differences and similarities. Experiments are automised, rendering large-scale investigations feasible. Results of our experiments, based on a preliminary experimental setting, show that users trained on some particular topics are not always influenced by previous search and click activities. However, different trained users actually achieve different search results, thus paving the way for further investigation.
Popular e-commerce websites, such as the Amazon Marketplace, offer a window to thousands of merchants, able to advertise their goods and services to millions of potential buyers.
Recently, the traditional advertising approach has moved towards a targeted one: the ad is shown only to online users with a specific profile - location, gender, age, e-shopping history are among the monitored aspects. This way, the merchant pays only for ads shown to users matching the ideal buyer for the merchant products. Targeted advertising is possible since the ads system is able to build a user profile tracking her online behaviour, e.g., on the e-commerce website, plus considering the data inserted by the user on the platform, at registration time.
Although personalised ads have the significant advantage to guide
the customer mostly towards products she likes, concerns were born
since the ads system could 1) hide to the user other potential
interesting products [
Price steering refers to the practice of changing the order of search
results to highlight specific products prices [
Google Shopping4 is a promising platform to study the effect of price personalization. It allows vendors to reach a large number of customers, really interested in specific products, thus showing the right product to the right customer. Google Shopping creates a selling campaign, placing specific products “in front of millions of online shoppers searching on Google.com”5. This is possible since Google can access several information on the user search activity, not only including that on Google Shopping. Actually, Google monitors the circle of websites known as Google Display Network (GDN), a large set of websites publishing Google ads6.
As shown in [
Past work showed that price steering is affected by the user
location, see [
Our work focuses on how the user behaviour, e.g., visiting a website of luxury goods, clicking on expensive products, affects the price of the items shown as the result of future queries over Google Shopping. We emulate the on-line behaviour of an affluent user and we compare her results list with one of a fresh user, which has not searched before on Google Shopping. Preliminary results show that, overall, affluent profiles have been shown different results with respect to those shown to the fresh control user. However, there is no a fixed rule, leading first to the most expensive products shown to the affluent users. The difference in the results list is however worth to be acknowledged, and calls for further investigation, with a more complex experimental settings and a more extensive evaluation.
The rest of the paper is as follows. Next section briefly presents related work in the area. Section 3 describes our methodology. In Section 4, we describe the experiments and we give the results. Finally, Section 5 concludes the paper. 2
This section briefly relates on literature in the area of
personalization of web results, particularly focusing on price steering and price
5 https://services.google.com/fh/files/misc/product_
listing_ads_intro.pdf
6 https://support.google.com/adwords/answer/2404190?
hl=en
discrimination. While price steering denotes the practice of showing
different products with different prices to different users,
discrimination is a similar practice, but related to the same product. Work
in [
Mikians et. al [
In [
The issue of price steering analysed in this paper has close
relationships with the ties between online user behaviours and the search
results (and/or advertisements) presented by a search engine (or a
website) to the same user. Indeed, price steering and
discrimination constitute only one aspect of a wide phenomenon, originally
put in the spotlight by Pariser with his Filter Bubbles [
In this work, we take inspiration from the analysis of online user behaviour to evaluate the effect on that behaviour on product prices shown to the user. 3
Our goal is measuring how the user online behaviour affects price
steering on Google Shopping. We use the approach which is similar
in [
The behaviour of a control user is different, she is idling while the affluent is in action.
Our expectation is that, when users query Google Shopping after a training phase where they behave as described above, affluent (resp., control) ones will likely see the highest (lowest) price products ranked first in their list of results.
In order to identify websites and keywords for luxury goods, we have exploited a tool originally intended for setting up targeting advertisements, the Display Planner tool of Google AdWords. The tool guides the user to find websites and keywords inherent to specific topics and terms8.
It is well known that Google monitors the users’ behavioural activities over the Internet through tools such as Google Analytics, Google Plus, and the Google ads system9. Thus, we train two affluent user profiles according to the specific online behaviours described above. Each profile has a control user profile associated. A control user has the same configuration as the user she is associated to (same browser, same OS). The difference is that control users are not logged into a Google account. Then, we compare the results obtained searching on Google Shopping the same keywords for both the trained users and the control users. In detail, the training and test phase are as follows: • Training step 1: The two affluent users visit a list of websites with topics related to their category. Websites have been chosen using the Google Display Planner. • Training step 2: The two affluent users search on Google Shopping for keywords related to their user category (keywords have been chosen according to the Display Planner, too). Then, they click on the most expensive product results and on those results coming from websites visited at step 1 (when present). • Test: All users (trained plus control) search for new products on Google Shopping. They do not interact with the results.
We let the two affluent users repeat the training phase eight times.
At the end of each training, we run the test. This is for building a
longer behavioral history of the user. Indeed, Google itself states that
it uses browsing and search histories to personalise the results and
enhance the user experience10. This is why we have repeated the
training phase eight times, always on the same profile, to
emphasize possible personalization aspects. One evidence of the efficacy of
this modality is in [
In our work, AdFisher runs browser-based experiments that emulate search queries and basic interactions with the search results, e.g., interacting with those search results whose price is above or below the price average on the total of the results, or those results belonging to a list of previously visited websites. Github hosts our extended version12.
AdFisher interacts with Selenium, a web browser automation tool. Selenium allows to run a unique instance of Firefox creating a fresh profile, with new associated cookies, the so called Firefox profile. The Firefox profile that is used is stripped down from what is installed on the machine, to only include the Selenium WebDriver.xpi plugin. Further, we take advantage of a plugin to automatically obtain the Python code for recording actions on web pages (e.g., clicking, 11 12
https://github.com/tadatitam/ info-flow-experiments
https://github.com/tienhv/Adfisher_for_ GoogleShopping typing, etc.), provided by Selenium IDE for Firefox13. All the experiments are done on the Firefox web browser version 43.0.4, controlled by Selenium in Python, under XUbuntu 14.04.
To simulate different real users, we browse from different IP
addresses, implementing a solution based on SSH tunneling to remote
computers. To simulate many computers from one geographic area,
we use the VPS services of Digital Ocean14. It is well known that,
when a user enters a query to Google, the query is unpredictably sent
to many distributed servers, to retrieve the results. This could produce
noise due to inconsistent data among different servers. To avoid the
issue, we query only towards specific Google servers IP addresses,
as in [
Finally, we use the following metrics to evaluate the search results
of the test:
• Jaccard Index: given two sets P and Q, Jaccard Index is 1 when
the sets are identical and 0 when their intersection is empty.
• NDCG (Normalized Discounted Cumulative Gain), measuring the
similarity between a given list of results and the ideal list of
results. Originally introduced as the non-normalised version DCG
in [
We have extended the AdFisher functionalities to handle Google Shopping pages and to mimic the behaviours described in Section 3. We automatically implement the whole experiments for the affluent and control users. The extended AdFisher also stores the query results for further analysis, as the calculation of the NCDG metric.
For the training phase, we emulate two user profiles logged into
Google. We consider 80 websites for each type of user profile and we
let the profiles visit all of them. Such choice has been driven by [
For the test phase, we still consider the two trained profiles, plus two control ones. All feature the same behaviour, which consists of querying Google Shopping with the same test keywords. Then, we collect the results. Figure 1 and Figure 2 show the first results showed to affluent and control users, for a specific test query. We extract links and prices of all the results to calculate the metrics listed in Section 3.
The training and the test phases are repeated eight times per user. Each session lasts around 90 minutes.
Table 2 shows NCDG values for each test session (results are for query “luxury shoes”), for each user.
Figures 3 and 4 plot, resp., the Jaccard index and the Kendall index, for the eight test sessions about “luxury shoes”. The blue lines represent are calculated over the results pages of Affluent 1 and Control 2 users, while red lines are calculated over the result pages of Affluent 2 and Control 1 users. This is to consider two users trained in same ways and connected from two different machines. Jaccard index shows evidence of results customization, while Kendall index says that, most of the times, the results for affluent and control profiles have a level of agreement (featuring a positive values for that index).
Figure 5 plots the values obtained calculating the average NCDG of the two affluent users and the two control users, over the eight test sessions. The test query is “luxury shoes”. Average NCDG indicates that even if affluent and control users follow the same pattern of pricing ordering, the former is closer to the ideal list results (where the most expensive products are in the first positions). shows, at a glance, how, in some cases, NCDG values are higher for affluent users than for control. The control users almost always obtain NCDG values lower than the affluent (or comparable in those cases with very similar values for the two kind of profiles). However, we argue that results are also affected by the specific search query over Google Shopping. As an example, Figure 7 shows the NCDG values for the test query “luxury jeans” for affluent 1 and the average of results of the two controls, over the eight sessions. The picture clearly indicates NCDG values that are greater for the affluent user. Instead, “mens dress casual shoes” (not shown in a Figure over the eight sessions) provides a higher value for the control which is an opposite result with respect to our expectations. While these initial results are promising, there is the need of further evaluation, where more, and more generic keywords, should be tested, at a larger scale. It is worth noting that, as introduced and motivated at the end of Section 3, the two affluent users (as well as the two control ones) are identical in terms of behaviour. Further, browser and OS settings are the same, while the IP address from which they browse is different training websites outfitideashq.com seriousrunning.com lululemonmen. com storelocate.us, skateboardingmagazine.com haircutinspiration.com training keywords mens fashion shoes, trendy boots, formal shoes for men, jogging shoes, cheap designer shoes, athletic shoes test keywords mens dress casual shoes, luxury shoes, dance boots, comfortable shoes, women trendy boots, luxury jeans, casual jeans luxury shoes Affluent 1 Control 1 Affluent 2 Control 2 (different devices, from the same geographical area). We have not compared directly the two affluents, since their results pages could be different for uncontrollable effects, such as timeouts or network delays (we are indeed emulating different IP addresses). The existence of sources of noise is also the reason why we have chosen to show, for some experiments, the average of the results of the two users. 5
We have designed and implemented a methodology to train and test user behaviours on Google Shopping, for evaluating a potential price steering, based on the willingness to pay attitude of the users. We have analysed the results list of affluent and control users. Affluent users were trained over eight training sessions. The results lists were obtained over eight test sessions, one at the end of each training session. The outcome of the experimentation is that, for most of the test queries, the result list of the affluent user is biased towards more expensive products than the one of the control user. However, the experiments results pave the way for further investigation. Indeed, we can imagine to 1) mimic queries from different geographical areas (not considered here, but recognised by past work as an impact factor for price manipulation); 2) use location via IP address as the major measurement, instead of artificial user profiles, because in some countries (like USA), the location/postcode is a strong indicator for economic situation, religion and race; 3) augment the number of training and test queries; 4) expand the duration of each training and test experiment; 5) mimic queries by different kind of users, e.g., mimic budget users, which always search for cheap products and services. Our experimental approach is general enough to be applicable to other e-commerce websites, like, e.g., Amazon.com and eBay.com. 6
This research has been partially funded by the Registro.it project MIB (My Information Bubble).